Anthropic’s highly anticipated IPO is getting closer. Last week, The Wall Street Journal saw a draft of the prospectus, giving us an early look at the company’s finances. We take a closer look at the reported numbers and why its IPO could test the wider AI industry.

Elsewhere in AI, OpenAI is reportedly nearing a new funding round that would value it at $1.25 trillion. Meanwhile, Chinese open-weight model Kimi K3.1 Pro reaches the frontier, Anthropic enters the personal AI agent race with Claude Duet, Meta Muse reaches 10 million users, and Disney scoops up Character.AI’s talent and IP.

Over in robotics, Unitree presents a robotic bodyguard, Ukraine claims the world’s first drone evacuation of a wounded soldier, and we look at how good a cleaning service powered by humanoid robots actually is.

Apart from that, this week’s issue of Sync features another AI-powered wet lab, signs that GLP-1 drugs may slow ageing, a potential Tesla and SpaceX merger, a startup claiming to have used AI-designed viruses to treat a patient, Meta losing another court battle, and more!

Enjoy!

Anthropic prepares for a $2 trillion IPO as AI bubble fears grow

What Anthropic’s prospectus reveals, and why its IPO will be a test for the whole AI industry

Anthropic is preparing for what could be one of the biggest stock market debuts ever, with a valuation that could reach $2 trillion, and now we have our first proper look at its finances. The Wall Street Journal has seen a draft of Anthropic’s prospectus, a document filed before an IPO that sets out a company’s finances, plans and potential risks.

But Anthropic’s IPO comes at a shaky time. Earlier this month, Wall Street had its worst week since June, with the tech-heavy Nasdaq falling 4.5%. CoreWeave, one of the best-known companies renting out AI computing power, also saw its shares and bonds tumble as investors questioned whether it could repay what it owes. Talk of an AI bubble is getting louder.

So what do Anthropic’s books tell us? And what will its IPO tell us about the state of the AI industry?

What Anthropic’s prospectus says

According to the Journal, Anthropic generated $16.3 billion in revenue in the first six months of 2026. A year earlier, the figure was $1.1 billion. That is roughly fifteenfold growth in a year.

Running AI is expensive, though. Every question Claude answers and every model Anthropic trains requires enormous computing power, mostly rented from cloud providers such as Google, Amazon and Microsoft. In the first half of this year, Anthropic spent nearly $9 billion on computing and infrastructure, up from $2.3 billion a year earlier. That is more than it spent in 2025, when its compute bill was $7.3 billion, and its revenue was $4.6 billion.

The good news is that revenue is now growing faster than computing costs. In 2025, Anthropic spent more on compute than it earned. In the first half of 2026, its revenue was almost double its compute bill.

The bad news is that Anthropic recorded a net loss of $80.7 billion in the first half of the year. That is nearly double the $42 billion it lost in all of 2025 and, as the Journal notes, one of the largest half-year losses ever reported.

But most of that is not money that actually left the company. About $72 billion comes from accounting charges tied to convertible notes and similar instruments Anthropic used to raise money. Convertible notes are loans that investors can later swap for shares. As Anthropic becomes more valuable, so do those future shares, and accounting rules require the company to record that increase in value as a cost, even though no cash changes hands. Without those charges, the loss is closer to $9 billion.

The bill coming due

The bigger concern is what Anthropic has promised to spend. According to the draft prospectus, the company is committed to spend more than $600 billion on computing power, much of it stretching into the next decade and mostly under contracts that cannot be cancelled. Those commitments include $111 billion with Google, $110 billion with Amazon Web Services and $31 billion with Microsoft, which Anthropic must pay “regardless of actual usage”. On top of that are up to $86 billion for renting Nvidia chips from two smaller providers, Nscale and Lambda, and more than $160 billion to lease Google’s chips through a deal with Broadcom. Deals still under negotiation could add as much as $150 billion more, taking the total to around $750 billion.

Even that may not be enough. Anthropic says its cloud partners have refused to give it as much capacity as it asked for and increasingly treat it as a competitor, forcing the company to turn to newer, less proven operators. One of them is Nscale, which Anthropic is reportedly negotiating with on a $45 billion computing deal.

Revenue, meanwhile, leans heavily on a handful of customers. About a fifth of first-half revenue came from just two of them, and Anthropic expects that concentration to continue. The prospectus also lists risks you won’t find in most IPO filings, including a catastrophic AI incident so costly that no insurer would cover it, and increasingly organised local opposition to new data centres.

A test for the whole AI industry

Anthropic’s IPO matters far beyond Anthropic because AI is now propping up a large part of the stock market and the economy. According to S&P Dow Jones Indices, around 45 AI-exposed companies accounted for about 39% of the S&P 500’s gains this year through 11 October. Strategists at Barclays argue that AI has become a major force not just in the economy but in the stock market, and increasingly in the bond market too.

Much of the AI build-out runs on borrowed money, and lenders are getting nervous. Barclays says the extra interest investors demand to lend to riskier AI companies has risen by more than 50% in six months.

Nowhere is that clearer than at CoreWeave. It is one of the so-called neoclouds, newer companies that buy AI chips, fill data centres with them and rent out the computing power. According to CNBC, it has taken on more than $51 billion in debt and other obligations. As The New York Times explains, CoreWeave commits to data centre leases lasting 15 years or more, but rents that capacity to customers such as Meta, Microsoft and OpenAI for shorter periods. That leaves it on the hook for longer than its customers may stay. Some investors have compared it to WeWork.

In mid-October, the selling began in earnest. The yield on CoreWeave’s bonds (in effect, the interest rate investors demand to hold them) climbed above 15%, deep into junk territory. The price of credit default swaps, a form of insurance against a company failing to repay its debts, rose past the level S&P Global Market Intelligence treats as a sign of distress. Its shares lost about a quarter of their value in a week. CoreWeave’s chief executive called the market’s reaction irrational.

CoreWeave is not the only one under pressure. Investors have been selling Oracle’s bonds as permitting problems delay data centres it borrowed billions to build. The cost of insuring SoftBank’s debt has jumped after it borrowed heavily to fund its investments in OpenAI. By the end of that week, even supposedly safe, investment-grade data centre bonds were falling.

Anthropic wouldn’t be the first $2 trillion AI-linked company to go public this year. In June, SpaceX went public in a record-breaking IPO that valued it at $1.77 trillion and soared past $2 trillion on its first day of trading. Within weeks, however, its shares had lost almost half their value from the peak. Now its borrowing costs are climbing too. According to Semafor, SpaceX’s new $40 billion debt raise is set to cost it more than a percentage point above where its existing bonds trade, as AI jitters spread.

So, is the AI bubble about to pop? Anthropic does not look like the pin, at least not yet. Its revenue is growing faster than its costs. But it has also signed up to some of the biggest spending commitments in corporate history, just as investors are starting to wonder whether the money pouring into AI will ever come back.

The next signal could come this week, when the Federal Reserve meets on 27–28 October. Then all eyes turn to Anthropic’s debut, and we will see whether investors are happy to buy into a $2 trillion company with $600 billion in bills to pay.

If you enjoy this post, please click the ❤️ button and share it.

Share

News roundup

🦾 More than a human

Novo Nordisk and Eli Lilly Report GLP-1 Signals of Slowed Biological Aging at ARDD
According to early research presented by Novo Nordisk and Eli Lilly at a conference at Harvard, the weight-loss drugs semaglutide and tirzepatide may help slow biological ageing. Novo reported signs that patients appeared biologically two to three years younger than expected, while Lilly observed similar trends in a smaller study. The findings suggest the drugs could offer health benefits beyond weight loss, although larger studies are needed to confirm the effect and its long-term significance.

OpenAI Invests in Startup That Uses Brain Implants to Detect Cancer
OpenAI’s investment arm and Peter Thiel’s Founders Fund are backing Coherence Neuro, a startup that has raised $45 million to develop an AI-powered brain implant designed to detect and treat cancer. The device aims to monitor tumour growth and deliver electrical pulses to slow it. Early human trials are under way.

Cool chemistry trick allows new tattoos to ‘turn on and off’
Scientists have created a new type of tattoo ink that can appear, disappear or change colour using light. Called PhotoTat, it could let people hide or change their tattoos without removing them. It could also be useful in medicine, such as marking areas for cancer treatment. So far, only purple ink is available, but researchers hope to add more colours in the future.

This implant can restore sight for patients with macular degeneration
In this article, MIT Technology Review explains how an eye implant called PRIMA is helping people with severe age-related vision loss regain some sight. The system combines camera-equipped glasses with a tiny chip implanted in the eye to send visual information to the brain. After training, more than 80% of patients were able to read, do crosswords and use their phones.

🔮 Future visions

▶️ The Future of Elections: Who Votes, How We Vote, and What Comes Next (46:54)

In this video, Isaac Arthur explores how technology could shape the future of elections. He examines different voting systems, their strengths and weaknesses, and the challenges of electronic voting, particularly around security and public trust. He also explains why traditional paper ballots are likely to remain relevant for some time despite advances in digital technology.

🧠 Artificial Intelligence

OpenAI nears up to $30 billion funding round at $1.25 trillion valuation led by UAE investors
CNBC reports that OpenAI is close to raising up to $30 billion at a $1.25 trillion valuation, in a funding round led by UAE investors and Coatue. The investment would provide fresh capital ahead of a potential IPO in 2027. OpenAI is also reportedly in talks to secure more computing capacity from the UAE.

Introducing Kimi K3.1 Pro
We have a new leader among open-weight models: Kimi K3.1 Pro from Chinese company Moonshot AI. The new model not only tops every other open-weight model, but also goes head-to-head with leading American models. According to Moonshot, Kimi K3.1 Pro delivers substantial improvements in coding, 3D modelling, financial research, agentic workflows and factual accuracy. Alongside Kimi K3.1 Pro, the company also launched a Highspeed variant that runs two to three times faster than the standard model.

Artificial Analysis confirms these claims and places Kimi K3.1 Pro amongst the best models from Anthropic and OpenAI. However, it notes K3.1 Pro is quite slow and very verbose.

Kimi K3.1 Pro is currently available across Kimi’s apps, developer tools and API, with full model weights scheduled for release by 4 November 2026.

Disney agrees to hire the majority of Character.AI’s workforce
Following Character.AI’s bankruptcy, Disney has agreed to hire at least 60 employees (the majority of the startup’s workforce) with total annual compensation estimated at $60 million. Disney also plans to acquire certain assets for $20 million. As WIRED argues, Disney may be using this kind of deal to acquire the talent and IP without also inheriting the legal baggage that would come with a traditional acquisition.

Microsoft pulls out of AI datacentre project in Essex over power supply concerns
Microsoft has pulled out of Nscale’s planned AI datacentre in Essex due to concerns over electricity supplies, dealing a setback to the UK government’s AI plans. The project, once described as Britain’s largest AI datacentre, may not have enough power until the early 2030s. Losing its only publicly announced customer also raises doubts about Nscale’s planned $35 billion stock market listing. Nscale says it remains committed to the project and is looking for other ways to secure power.

Anthropic quietly pitches its own AI safety bill to lawmakers
Politico reports that Anthropic has privately proposed legislation to US lawmakers that would establish national AI safety standards, including independent evaluations, transparency requirements and incident reporting. The bill would allow courts to block AI models only when they pose an imminent catastrophic risk, such as causing more than $1 billion in damages or over 50 deaths or serious injuries, while limiting federal agencies’ power to halt deployments and overriding some state regulations.

Anthropic says it’s committing 5% of its shares to a ‘public benefit fund’
Anthropic has announced a public benefit fund and pledged 5% of its shares to support initiatives addressing AI’s societal impact, with additional contributions tied to the company’s growth. The fund could finance charitable donations, direct financial assistance and efforts to tackle AI-driven job disruption.

How employment is changing at Anthropic
Anthropic published a report on how AI impacted its employees and productivity, finding that AI significantly increased output while changing how people work and who gets hired. Employees are spending less time doing tasks themselves and more time managing AI agents, making skills like decision-making, adaptability and working with AI more important. At the same time, AI is reducing demand for junior employees and allowing companies to grow without hiring as many people, suggesting its biggest short-term impact may be on job roles and hiring rather than widespread job losses.

Introducing Claude Duet
Anthropic is joining the personal AI agent race with Claude Duet, its answer to Meta Muse and OpenAI dots. Duet is a proactive assistant that works in the background, monitoring connected apps, surfacing important updates and handling tasks such as scheduling meetings, building dashboards and managing workflows. It runs on its own virtual computer, integrates with more than 6,000 apps and asks for approval before taking sensitive actions. The research preview is available by waitlist to Pro and Max subscribers in the US.

Meta says Muse has over 10M users, with shopping a top activity
Meta has announced that its AI assistant, Muse, has surpassed 10 million users, and that 25% of those users use it daily. Shopping has emerged as a key activity, alongside productivity tasks such as managing emails and calendars.

NVIDIA GTC Berlin 2026 NVIDIA Intros DSX PODs and More
This week, NVIDIA hosted its 2026 GTC event in Berlin, where it announced new hardware and software aimed at accelerating AI infrastructure deployment and advancing agentic AI. Highlights included DSX POD, a container-based Vera Rubin system designed to cut data centre deployment times from months to just 14 days, and OpenShell, a framework for running AI agents securely.

US banning Chinese AI DeepSeek from government agency IT systems
The US Department of Homeland Security has ordered federal civilian agencies to remove and ban DeepSeek products and related services over national security concerns, including potential data leaks to China, espionage and cybersecurity risks.

Saudis Aided Family of Engineer Accused of Bugging OpenAI, Prosecutors Say
A US federal judge denied bail to former OpenAI engineer Abdulrahman Alfozan, who is accused of secretly planting recording devices in the company’s San Francisco offices. Prosecutors raised concerns over his extensive ties to Saudi Arabia, including the government’s offer to cover his bail and its reported assistance in helping his family leave the US. Investigators are examining whether he acted for financial gain or on behalf of a foreign government.

OpenAI models are coming to Google Cloud
OpenAI announced a partnership with Google Cloud to make its frontier AI models and Codex agentic workflows available through Google Cloud’s Model Garden. Enterprises can build and scale AI applications on Google Cloud, with access through its Marketplace. General availability is expected in the coming weeks, expanding customer choice and OpenAI’s reach across cloud platforms. OpenAI is quick to clarify that Microsoft still remains its primary cloud partner.

Google staff resign in backlash over its AI contracts with military
Google is facing growing criticism over its use of AI in military operations, with some employees having already left or being in the process of leaving the company. Staff have raised concerns about inadequate oversight, the reliability of AI in warfare and fears of retaliation for speaking out. While Google insists its military contracts follow ethical guidelines, critics warn that rapid deployment could increase the risk of errors and weaken human control over military decisions.

Qualcomm’s CEO says AI companies want phones running 100-billion-parameter models by 2028
Qualcomm CEO Cristiano Amon said several AI companies are developing smartphones designed to run AI agents continuously, handling tasks on users’ behalf by accessing personal data, apps and device sensors. He sees these AI-powered phones as the next major shift in mobile technology, combining on-device processing with cloud-based capabilities.

MiniMax M3.1 Flash
MiniMax has introduced M3.1 Flash, a new AI model optimised for coding, agentic workflows, computer use and cybersecurity. Codenamed Space Bunny, the model delivers performance competitive with leading AI systems while offering nearly 3x faster inference and 70% lower costs than its predecessor, according to benchmarks published by MiniMax.

Artificial Analysis places M3.1 Flash at the level of Claude Haiku 5.5 and GLM-5.3-Flash on its Intelligence Index benchmark. When cost per task is added to the picture, M3.1 Flash scores almost the same as Haiku 5.5.

M3.1 Flash is available through the MiniMax API, M Plan and MiniMax Code, with open model weights expected to follow.

Introducing Clairic, a new System One class of extraction models
Cloudflare presents Clairic, a new type of AI model the company calls an extraction model and another entry in the emerging class of fast, specialised System One models. Unlike traditional LLMs, Clairic focuses on extracting structured information from unstructured inputs rather than generating text. It is designed for tasks such as document processing, data extraction, cost estimation and forecasting. Cloudflare promises that Clairic delivers faster, cheaper and more accurate results than general-purpose LLMs.

Mistral: Introducing Timestral
Mistral AI has introduced Timestral, a family of lightweight AI models designed to predict future trends using historical data. The models are fast, efficient and can run on standard CPUs without needing GPUs. Timestral can be used for predicting equipment failures, detecting unusual patterns, forecasting energy demand, improving supply chains and assessing financial risks. It can also make predictions without additional training and be adapted to specific business needs.

Inside Prometheus: Meta’s Journey to a Gigawatt-Scale Supercluster
Meta takes us inside Prometheus, its one-gigawatt AI supercluster in Ohio, showcasing the technology inside it, from NVIDIA’s latest GPUs and high-speed networking to advanced liquid cooling and power systems. With the first phase already operational, Meta plans to scale the facility to around one million GPUs and eventually upgrade it with next-generation Vera Rubin hardware.

🤖 Robotics

The Robot I Didn’t Expect to Build
Jonathan Hurst, co-founder of Agility Robotics, writes about the challenges of turning advanced robotics into commercially viable technology through the development of Digit 5. He highlights the need to balance scientific ambition with practical demands, particularly the trade-offs between safety, functionality and design. Hurst ends by saying that he wants not only to build functional robots, but also to make them beautiful, so I am looking forward to Digit 6.

Tesla starts Optimus robot production in California as profit falls 43%
Tesla has started producing its Optimus humanoid robot, with Elon Musk aiming to make 20,000 units per week by the end of 2027, although technical challenges remain. The news comes as Tesla reported a 43% drop in third-quarter profits to $787 million, mainly due to higher spending on AI, robotics and self-driving cars, while revenue stayed flat at $28 billion. Tesla has also delayed the launch of fully self-driving cars for customers until 2027.

Ukraine claims world’s first drone evacuation of wounded soldier
Ukraine claims to have carried out the world’s first evacuation of a wounded soldier using a heavy-lift drone because landmines and enemy drones made a normal rescue too dangerous. The successful mission shows how Ukraine is using new drone technology to save lives and improve its position in the war against Russia.

▶️ Unitree Introducing: H2-X Kung Fu Legend Guardian (1:00)

Unitree introduces its new humanoid robot, H2-X, and it does so in a typical Unitree fashion—by showing how it can run, jump, punch things, and do kung fu.

I Paid $60 for Two Humanoids to Clean My Home In San Francisco
There is a startup in San Francisco that is offering cleaning services performed by humanoid robots, and this post looks at what two of them can deliver for $60. Beyond the cleaning itself, it examines the economics of robot-as-a-service, the challenges of relying on human operators, and why shared robot fleets could become a promising business model. It also looks at the operational, privacy and safety issues that need to be solved before humanoid robots become part of everyday life.

▶️ Can I Build This Tiny Robot From Scratch in the Middle of China? (14:51)

Dorian Todd documents in this video his one-week trip to Shenzhen, where he gets an opportunity to see the Chinese hardware scene up close. Unbeknownst to his hosts, Todd arrived with a secret mission: to build a small robot. In his free time between meetings and factory tours, he visited Shenzhen’s giant electronics market in search of components. It is an interesting way to show why the city has become such a hub for hardware and robotics companies and startups.

🧬 Biotechnology

Grounding AI agents in biological reality: Building blocks for autonomous wet labs
Google follows in Anthropic’s footsteps and opens a wet lab where AI agents work with laboratory robots to autonomously design, run and refine biological experiments. Using a new framework that enables safe, real-time control of lab equipment, Google aims to accelerate drug discovery by allowing AI agents to learn directly from experimental results. Early tests show promising results in molecular design and cancer-related protein discovery.

Start-Up Claims to Use A.I.-Created Viruses to Treat a Patient
Scientists have reportedly used AI-designed viruses for the first time to treat a patient with an antibiotic-resistant E. coli infection. Developed by Radical Numerics, the customised viruses are designed to target bacteria that have resisted existing treatments. Although the outcome is not yet known and the research has not been peer-reviewed, a successful trial could pave the way for faster, more personalised treatments for drug-resistant infections.

OpenAI acquires Orchestra Bio
OpenAI has acquired Orchestra Bio to integrate its AI agents into OpenAI’s Rosalind life sciences platform with the aim of streamlining drug development workflows from experiment planning to clinical trial coordination. The financial terms of the deal were not disclosed.

How Claude designed protein binders against scorpion neurotoxins
Anthropic shares in this post how Claude could accelerate drug discovery by designing proteins that bind to dangerous scorpion toxins. The AI lab says Claude generated more than 318,000 designs in just 64 hours. Of those, 211 were tested in the lab and 17 successfully bound to their intended targets, although whether they can actually neutralise the venom remains unproven.

Using thousands of AI agents to uncover new leads in biological data
OpenAI, PerturbAI and Scripps Research showed how thousands of AI agents powered by GPT-Rosalind can speed up biological research, analysing millions of mouse brain cells in 32 minutes instead of months. By combining genetic data with existing research, the AI agents uncovered new insights into brain diseases that scientists can then test through experiments.

💡Tangents

Tesla and SpaceX Formed Committees to Discuss Potential Merger, Elon Musk Says
After folding X and xAI into SpaceX, Elon Musk is now exploring a merger with Tesla that would bring nearly all his companies under one roof. Both companies have formed special committees to evaluate a potential deal, which would face shareholder approval and regulatory scrutiny.

Meta Ordered to Pay $25 Billion for Deceiving Users on Privacy
A New Mexico judge has ordered Meta to pay a record $25 billion penalty for misleading Facebook users about their data privacy following the Cambridge Analytica scandal. The ruling follows a jury finding that Meta deliberately deceived users for years, violating state consumer protection laws 43 million times. Meta plans to appeal. The decision comes less than two months after the company agreed to pay up to $17.1 billion to settle a social media addiction lawsuit.

Startup taps underwater volcano as energy source in “proof point” test
Endurance Energy has demonstrated the world’s first subsea geothermal power plant in the Pacific Ocean, generating net-positive electricity for 8.7 seconds before technical failures halted the test. Despite its brief operation, the trial demonstrated the potential to harness geothermal energy from the ocean floor. Further testing is planned for 2026–2027, ahead of a commercial deployment in Tonga in 2027, where the technology could reduce reliance on expensive imported fossil fuels. However, questions remain over its long-term reliability and commercial viability.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

This week, Meta somehow returned with two new models: one reaching the frontier and another becoming the best Western open model. That made me ask whether Meta is back in the AI race.

Elsewhere in AI, Character.AI filed for bankruptcy, AI-exposed tech stocks had a bad week, and OpenAI found a spy. Meanwhile, Anthropic reported a jump in revenue, Epic Games went all-in on AI, Google’s SynthID was broken, Anthropic embedded independent AI safety evaluators inside the company, and warned about self-replicating malicious AI.

Over in robotics, Google showed how well Gemini 4 Argon can control robots, Waymo is heading to Canada, and Boston Dynamics showed how Atlas is built.

Apart from that, this week’s issue features Tesla’s new rocket-powered Roadster, a brain-computer interface restoring sight, how the rise of agentic AI is increasing demand for CPUs, the first person to use an AI-generated voice in the UK Parliament, and more!

Enjoy!

Meta is back in the AI race, just not the one everyone else is running

Muse Wonder, Muse Spark Open and a bet on personal intelligence

For a while now, I’ve had a draft of an article titled “How Meta lost the AI race” sitting in my drafts. Meta’s recent moves have made me reconsider it.

Meta began the AI race as an unlikely champion of open-weight models with its Llama family. But the disappointing Llama 4 and the benchmark-maxxing controversy around it ended that chapter, and the open-model lead passed to Chinese labs. What followed was a near-total rebuild of Meta’s AI effort. Mark Zuckerberg went on a hiring spree, offering nine-figure pay packages to lure in top AI talent. It culminated in the $14.3 billion deal with Scale AI that brought its founder, Alexandr Wang, to lead the newly formed Meta Superintelligence Labs.

Six months ago, that effort produced Muse Spark, the first of Meta’s new Muse models. Last week, Meta released two more, and they made me ask a question I didn’t expect to ask: Is Meta back in the AI race?

Muse Wonder reaches the frontier

On 12 October, Meta unveiled Muse Wonder and its bigger sibling, Muse Wonder Full, its most capable models to date. Both are built for what Meta calls personal and professional agents. Meta highlights two abilities in particular. Active Memory lets the model carry what it knows about you across conversations and recall the right details when they become relevant. Active Focus keeps it working towards a goal, trying different approaches when it gets stuck and checking in before doing anything important.

Artificial Analysis places Wonder Full in the top three on six of the ten tests in its Intelligence Index, and on coding agents it matches Claude Code running Sonnet 5.5. Its weak spot is general knowledge, where it still trails the best models from Anthropic, OpenAI, and Google, though it partly makes up for it by rarely making things up.

Muse Wonder costs $4 per million input tokens and $20 per million output tokens, the same as Claude Opus 5.5 and several times what Meta charged for Spark 1.3. Wonder Full is available only to Muse Code subscribers.

When I wrote about Muse last month, I said Spark 1.3 was close to the leading frontier models. With Muse Wonder, Meta is no longer just close. It is not leading the frontier, but it is there.

Back to open models

Three days after releasing Muse Wonder, Meta returned to its open-weight roots with Muse Spark Open, which it calls “the most capable US open-weight model to date”. It is a big model, with 816 billion parameters (42 billion of them active at any time) and a 1-million-token context window, built for coding and agentic work.

Artificial Analysis gives Muse Spark Open 39 points, the best result for an open-weight model from the US and the West as a whole. It edges out the recently released Mistral Large 4, which scored 38, although Mistral won’t release its weights until the end of October. Muse Spark Open ranks only 10th among large open-weight models, though, so the Chinese labs still lead. It is also expensive for an open model, and its licence limits commercial use.

Still, it fills a real gap. Until now, if you needed a capable open model, your best options were Chinese. Muse Spark Open gives companies that want, or are required, to use an American or a non-Chinese model a genuine alternative. If Meta keeps improving it, Muse Spark Open could challenge Chinese labs for the open-weights lead.

Bet on personal intelligence

Meta now has all the pieces a big tech company needs to compete in AI: frontier models, open models, the Meta Model API and its own coding tool, Muse Code. On paper, it can go on to challenge OpenAI and Anthropic. But in practice, I don’t see Meta becoming a leading AI company.

Meta’s focus is on Muse, its personal AI agent that launched on 8 September. Muse can shop, book and handle everyday tasks on your behalf, and it will soon run on Muse Wonder. Everything else is built to help Meta win the race for personal intelligence. That’s the course Zuckerberg set for the company and, frankly, it’s the only niche where Meta has a real chance of winning. It is a market made for a company whose apps reach 3.5 billion people, and Meta got there first, perhaps a little too eagerly. According to The New York Times, Zuckerberg ordered Muse’s launch despite internal safety concerns, after a small startup called Instinct began taking off with a similar agent. Meta disputes that account.

So far, the gamble seems to be paying off. Muse reached 5 million US downloads in 22 days, compared with 56 days for ChatGPT, and BNP Paribas estimates it had passed 10.4 million downloads by 16 October. But there are signs of cooling. Daily downloads fell by about 8% in early October, and Meta’s own heavy advertising is doing much of the lifting. Muse is also tiny next to ChatGPT’s 1.2 billion weekly users and Gemini’s roughly 1 billion monthly users.

Meta won’t have the market to itself, either. OpenAI has dots, Instinct has just raised $1 billion, and OpenClaw remains a formidable open-source alternative. Google launched Gemini Spark back in May, although it keeps it behind a $100/month subscription. Apple will have its own answer too, if it can first sort out Siri.

Will the billions pay off?

All of this rests on spending on a scale few companies have ever attempted. Meta expects capital expenditure of $130–145 billion this year, roughly double what it spent in 2025. It is financing some of its biggest data centres through joint ventures, such as one with BlackRock, that keep the debt off its own balance sheet. In the second quarter, its free cash flow fell to just $784 million, down from $8.55 billion a year earlier.

Even Zuckerberg admits it was a leap of faith. On the Cheeky Pint podcast, he recalled asking Wang, “You sure this is going to work?” before committing what he guessed would be $100 or $200 billion. Wang’s answer, in short, was that nobody could promise it would.

Unlike OpenAI and Anthropic, Meta can fund such huge bets from its own pocket. The AI frontier labs need enterprise customers because they have nothing else to pay for their compute. Meta has an advertising business that brought in $59.4 billion in the second quarter alone, up 27% year on year, and it credits AI for much of that growth. In a way, AI already pays for itself at Meta.

So the question is not whether AI pays off for Meta, but whether spending on this scale does. That depends on Muse becoming a real business, through transaction fees, subscriptions and, sooner or later, ads. Anthropic has shown that selling AI to businesses works. Meta now has to prove there is a consumer route as well.

So, is Meta back in the AI race? Yes, but it is running a race of its own. It is the frontrunner in a category that has barely begun and has yet to prove it can make money. We may get the first signs of whether that was the right bet on Wednesday, when Meta reports its Q3 results, and over the coming months, as we see whether people stick with Muse once the novelty wears off. For now, my draft of “How Meta lost the AI race” stays in the drawer. But I’m not deleting it just yet.

If you enjoy this post, please click the ❤️ button and share it.

Share

News roundup

🦾 More than a human

A New Anti-Aging Drug Looks to Turn Back the Clock on Skin Age
Biotech company Rubedo Life Sciences has developed an experimental cream designed to combat skin ageing by targeting damaged “zombie” cells. Early trials showed promising results, with participants’ skin appearing biologically two years younger on average after four weeks, alongside reduced inflammation and improved collagen levels. The treatment could also help with sun damage, psoriasis and hair loss, although further research is needed to establish its long-term effectiveness and visible anti-ageing benefits.

Science Corporation implants its vision-restoring chip in its first Stargardt disease patient

Science Corporation has implanted its PRIMA vision-restoring chip in its first patient with Stargardt disease, an inherited condition causing progressive vision loss. Paired with camera-equipped glasses, the implant has already enabled the patient to recognise letters and read words. Although currently limited to a narrow field of black-and-white vision, the technology offers promise for treating previously untreatable forms of blindness.

U.N. Warns Against Technology Used to Read Citizens’ Minds
The United Nations has warned that rapid advances in neurotechnology could threaten fundamental human rights by enabling governments to access and manipulate people’s thoughts. While the technology offers significant medical benefits, concerns over its use by law enforcement and private companies have prompted calls for stronger regulations to protect brain data, privacy and freedom of thought.

🧠 Artificial Intelligence

Character.AI files for bankruptcy
Character.AI, a platform that lets users create and interact with AI-powered characters, has filed for Chapter 11 bankruptcy amid mounting legal challenges, high operating costs and difficulties monetising its large user base.

Founded in 2021 by former Google engineers Noam Shazeer, a co-author of the famous Attention is all you need paper, and Daniel De Freitas, the startup quickly gained popularity and raised more than $193 million in funding. In 2024, Google struck a $2.7 billion licensing and “acqui-hire” deal that brought both founders back to the company. Disney has now agreed to acquire some of Character.AI’s assets for $20 million. The platform could still shut down if no buyer is found.

New Cracks in AI Trade Push Wall Street to Worst Week Since June
It wasn’t a good week for tech companies, as concerns about the AI boom, rising borrowing costs and stubborn inflation triggered a sharp sell-off. The Nasdaq fell 4.5%, while AI cloud provider CoreWeave lost 25% and Dell plunged 22% over the week. Chipmakers including NVIDIA, Broadcom, Micron and AMD also came under pressure, alongside tech giants Amazon and Alphabet. Investors are increasingly concerned that higher borrowing costs could make the enormous sums being spent on AI harder to sustain, potentially slowing investment and weighing on the wider US economy.

FBI arrests OpenAI employee accused of planting hidden recording device in conference room
What a timeline we are living in… OpenAI has called in the FBI after allegedly catching one of its own employees hiding a recording device in an executive meeting room at its San Francisco office. The employee was fired and arrested, with authorities suspecting he was trying to secretly record sensitive discussions about OpenAI’s technology. OpenAI says there is no evidence that its AI models, source code or customer data were stolen.

Anthropic Says Revenue Jumped 45% to $16.8 Billion in Third Quarter
Anthropic executives told investors that the company’s third-quarter revenue surged 45% to $16.8 billion, The New York Times reports. Despite strong growth and adjusted profitability, investors remain concerned about rising interest rates, the sustainability of AI investment, competition from cheaper Chinese models and Dario Amodei’s calls for tighter AI safety measures. Anthropic is preparing for a potentially record-breaking IPO that could raise more than $100 billion and value the company at $2 trillion, which is seen as a major test of investor confidence in the wider AI industry.

Anthropic assigned highest junk credit rating in June
Anthropic secured a BB+ credit rating from Kroll Bond Rating Agency in June, just one notch below investment grade, the Financial Times reports. The rating puts the AI company close to receiving a financial seal of approval that would make it easier and cheaper to borrow money and attract investors as it prepares to go public. With plans to spend more than $500 billion on computing infrastructure over the next decade, access to cheaper financing could become increasingly important as Anthropic funds its rapid expansion.

Gemini 4 Argon is now globally available
Google’s newest flagship model, Gemini 4 Argon, is now globally available. Although it did not top all performance benchmarks, Gemini 4 Argon is a respectable comeback from Google, with significant improvements in reasoning, coding, scientific research and complex professional tasks. Developers and enterprises can access the model through Google AI Studio and the Gemini API, while eligible Google AI Pro and Ultra subscribers can use it in the Gemini app, AI Mode in Search and Google Antigravity.

Anthropic: An update on independent evaluators
Anthropic is following through on Dario Amodei’s proposal to embed independent evaluators inside AI labs, granting six researchers from AI safety nonprofit METR employee-level access to its internal systems and safety processes. The researchers will independently assess Anthropic’s models and safety practices and publish their findings without the company’s approval, subject to restrictions on sensitive information.

AI’s quiet safety gatekeepers are stepping into the spotlight
OpenAI and Anthropic have recently promised to bring independent evaluators into their labs to check their AI safety practices. In this article, CNBC takes a closer look at these evaluators and the challenges they face. Key questions include who will pay for their work, how much access they will get and whether they can remain truly independent. The article also looks at whether government oversight will be needed to make sure these checks are effective.

Anthropic: AI models may be capable of self-replicating without access to a datacenter
Anthropic’s research suggests that AI models could soon autonomously replicate across internet-connected devices, potentially creating a new form of decentralised, adaptive malware that is difficult to detect and eliminate. Experiments show that AI models can already carry out much of this process, with smaller models potentially achieving full autonomy by the end of 2026. While this threat has yet to materialise in the real world, the findings highlight growing cybersecurity risks and the need for stronger defences.

Judge orders OpenAI to turn over hiring records in Apple trade secrets case
A US judge has ordered OpenAI to share hiring and interview records with Apple as part of a lawsuit accusing OpenAI of stealing trade secrets through former Apple employees. The court also approved access to private messages and interviews with three former Apple staff now at OpenAI, but rejected Apple’s broader requests for information about OpenAI’s hardware development.

OpenAI Computing Chief to Depart
OpenAI’s head of computing capacity, Sachin Katti, is leaving after roughly a year. Katti helped secure major computing deals with AWS, Nvidia and other partners. Former xAI executive Uday Ruddarraju will take over OpenAI’s computing strategy. Katti joins a growing list of high-profile executives who have departed OpenAI in recent months, including former applications CEO Fidji Simo, COO Brad Lightcap, chief revenue officer Denise Dresser and head of data centres Chris Malone.

OpenAI’s London staff receive average of £2.3m per person annually
The Guardian reports that OpenAI’s London office had 110 employees in 2025, with an average pay package of £2.3 million, more than double the previous year and 58 times the UK median salary. UK staff collectively received £87 million in salaries and £168 million in shares, although top AI researchers earned significantly more. OpenAI is continuing to expand its London presence as a major international research hub.

NVIDIA Commits $250 Million to Advance UK Agentic Safety and Security Research Over the Next Five Years
NVIDIA has committed $250 million over five years to advance UK research into AI safety and security, providing computing resources to leading institutions including the University of Oxford, UCL and Imperial College London. The initiative will support research into secure autonomous AI agents, robotics safety and emerging AI risks.

CoreWeave credit default swaps hit distress levels as investors sour on AI debt
Investors are growing increasingly concerned about CoreWeave, an AI computing company that has taken on more than $51 billion in debt and financial commitments to fund its rapid expansion. Despite the company’s assurances that future customer revenues will cover its costs, fears about its ability to repay its debts are mounting.

Jev: For the builders AND dreamers
TypeSafe AI has released a new version of its decision model, Jev 1.14.0, which adds image support, a new data type for the model to use and a longer context window. The update also improves performance, cost-to-intelligence ratio, and reduces error rates.

Microsoft-Decision-1
Microsoft joins the decision model party with Microsoft-Decision-1. The company claims its model is more accurate and faster than any other model in this category. Microsoft-Decision-1 is available on Microsoft Foundry and OpenRouter.

FLUX 3 Image Klein
Black Forest Labs presents FLUX 3 Image Klein, a new open-weight AI model for image generation and editing that combines high-quality results with fast performance and efficient local deployment. The model supports native 4K output, precise image editing and up to 10 reference images, and can run on consumer GPUs with just 8GB of VRAM. FLUX 3 Image Klein is available via Black Forest Labs’ API or can be downloaded from Hugging Face for self-hosting and fine-tuning.

Google Harness²: Recursive Agent Harnessing for an Open World
Developed by researchers at Northwestern University and Google Cloud AI Research, Harness² is a framework for recursively improving AI agents without modifying their underlying models. By optimising prompts, memory, tools and workflows through experimentation and feedback, it enables continuous improvement, delivering significant performance gains across coding, web and office tasks. Harness² is open source and available on GitHub.

▶️ Agentic AI and Why Everyone Wants CPUs Again (53:40)

Everyone needs and wants GPUs to run AI models, but as Ian Cutress argues in this talk, the rise of agentic AI workflows is also increasing demand for CPUs. He introduces the main players in the space and explains how new CPU designs, memory technologies and advanced chip packaging could help support increasingly complex AI systems. He also discusses the challenge of balancing CPUs and GPUs, keeping costs under control and designing ever more complex chips.

Introducing AMD Helion, Agentic AI Compute at Rack Scale
AMD has unveiled Helion, an open rack-scale infrastructure platform designed for agentic AI workloads. Combining EPYC processors with optional Instinct GPUs, it offers flexible, cost-efficient configurations for hyperscale data centres and enterprises. Expected to launch in November 2026, Helion positions AMD as a more open and customisable alternative to NVIDIA’s tightly integrated AI infrastructure, with a greater emphasis on flexibility and efficiency than GPU-centric computing.

Microsoft’s latest custom chip: Maia 300
Microsoft has unveiled Maia 300, its latest in-house AI chip, as it looks to reduce reliance on Nvidia and cut the cost of running AI services. Set to roll out in 2027, the chip is designed for AI inference workloads and promises 40% more tokens per megawatt than leading GPU-based systems.

IMF says AI investment keeps global growth around 3% but brings massive disruption
IMF Managing Director Kristalina Georgieva said surging AI investment is helping sustain global economic growth at around 3% and particularly benefiting Asian economies. However, she warned that AI could disrupt up to 40% of jobs worldwide and 60% in advanced economies, raising concerns about employment, education and taxation.

Epic Games Delays Disney Project, Asks Developers to Use More AI

Bloomberg’s Jason Schreier reports that Epic Games is taking an “AI-first” approach to developing a new shooter featuring Disney’s Marvel and Star Wars characters—a game the company has delayed indefinitely after development progressed more slowly than expected. Developers have reportedly been instructed to use AI tools to generate all code and artwork, with humans overseeing the process. The move reportedly surprised and concerned staff.

Palantir CTO says the Pentagon has already replaced Anthropic’s AI
Palantir CTO Shyam Sankar said the Pentagon has replaced Anthropic’s AI with models from OpenAI, Google, xAI and others without disrupting military operations. He dismissed concerns over AI risks as exaggerated and highlighted its ability to accelerate military targeting while keeping humans in control. Sankar also disputed reports linking AI overreliance to a deadly US strike on an Iranian school, suggesting human error was more likely.

Multi-Agent Self-Supervision
In this post, researchers from Japanese AI lab Sakana describe a new approach to making AI models smarter through collaboration. Their method, called Multi-Agent Self-Supervision (MASS), brings together multiple AI agents to tackle complex research tasks, learn to work together more effectively, and use their discoveries to improve the underlying model. Early results suggest this approach could help AI systems solve problems better and continuously improve themselves without relying on a more advanced AI to guide them.

Reverse-SynthID
Someone has reverse-engineered SynthID, Google’s invisible watermark for images generated by Gemini. In the repo, they explain how the watermark works and describe a method for modifying watermarked images so they look unchanged to humans while dramatically reducing the chance of being flagged as AI-generated.

Redesigning source control at Anthropic
To deal with the growing strain AI-generated code places on Git, Anthropic created Scion, a scalable source control system built in just five months with Claude’s assistance. This post explores Scion’s architecture, how it improves performance and reduces infrastructure costs, and the lessons learned from building source control for AI coding agents.

Amazon introduces 2 new AI products from Bee: Bee Frontier and Bee Eclipse
Amazon is expanding its personal AI device line-up with Bee Frontier, a wearable assistant, and Bee Eclipse, a magnetic smartphone accessory. Both devices use AI to turn conversations into summaries, reminders and actionable insights, and add Alexa+ integration and personalised shopping features. They are the first new Bee products since Amazon acquired the startup in 2025.

The first person to use an AI-generated voice in the UK Parliament
Disability rights campaigner Baroness Campbell has become the first person to use an AI-generated voice in the UK Parliament, during a House of Lords debate on the government’s plans for social care. She used a voice model from ElevenLabs that, as she says, sounds like her, but clearer and a little bit younger.

🤖 Robotics

Gemini 4 Argon for Robotics

As part of a campaign promoting Gemini 4 Argon, Google showed how its latest flagship model performs across various fields, including robotics. The company says Gemini 4 Argon can control robots, create realistic simulations and improve performance through self-learning. It also claims the model leads the Embodied Reasoning (ERQA) benchmark, which tests a model’s ability to understand and solve physical and spatial problems in 3D environments, outperforming other frontier models such as Opus 5.5 and GPT-6 Astra.

Waymo’s robotaxis are headed to Canada
Waymo, Alphabet’s self-driving car company, announced that it is expanding into Canada after receiving approval to test autonomous vehicles in Toronto. The company will start by mapping roads with human-driven cars, then move to autonomous testing and, eventually, a public robotaxi service. Canada is Waymo’s third international market, after Japan and the UK.

The Start-Up That Wants to Build ‘Microrobots’ With A.I.
The New York Times profiles Atomic Machines, a California start-up developing an AI-powered manufacturing system it calls a “matter compiler”, which translates digital designs into physical devices. Founded by former Uber executive Jeff Holden, the company has already produced an ultra-small, fast and energy-efficient electrical switch that could improve power management in data centres. By combining AI-driven materials research with automated manufacturing, Atomic Machines hopes to transform how physical products are made and enable advances in areas such as robotics and medicine.

▶️ Machine, Assemble, Repair: How We Build Atlas | Boston Dynamics (5:56)

Boston Dynamics takes us inside the factory where it builds its humanoid robots, from the in-house CNC shop to the assembly lines where the machines are put together. One interesting thing we learned in this video is that it takes about one day to build a single robot, and the company plans to scale production to one robot per hour next year.

AI drone outpaces former world champions using one camera and no GPS
Researchers at Delft University of Technology have developed a racing drone that can reach speeds of 100 km/h and outperform world champion human pilots. Using just one camera, motion sensors and a small onboard computer, the drone can navigate complex courses without GPS. Its AI system reacts quickly and controls the motors directly, allowing it to fly at high speeds with remarkable precision.

🧬 Biotechnology

Colossal Biosciences announces new AI platform to help accelerate de-extinction efforts
Colossal Biosciences has unveiled GenoMorph, a new AI platform designed to accelerate efforts to revive extinct species, particularly the dodo. By analysing genetic data and physical traits from living birds, the technology has identified around 2,300 potential gene-editing targets, helping researchers overcome the scarcity of ancient DNA. Combined with advances in genetic engineering, the platform could make de-extinction research faster and more precise, with wider applications in conservation, agriculture and healthcare.

Cancer Vaccines Are the N.I.H.’s Next ‘Big Bet’
The US National Cancer Institute is launching a major initiative to accelerate the development of personalised cancer vaccines that train the immune system to target a patient’s tumour. Encouraging early results across several cancers have raised hopes for a breakthrough in treatment, although larger trials are still needed. The initiative aims to make these vaccines widely available within 7 to 10 years.

Giraffe Bio aims its bespoke biomolecules at the looming copper shortage
As global copper demand races ahead of supply, with a potential 25% shortfall by 2040, Giraffe Bio is betting on bioengineered compounds to squeeze more copper out of existing mines. Produced using engineered microbes and tailored to specific minerals, these compounds could boost copper recovery by up to 5% while reducing reliance on harmful chemicals used in traditional mining processes. The same technology could also be applied to other valuable minerals, including gold and silver.

💡Tangents

Tesla’s New Roadster Takes Flight After Nearly a Decade of Delays

After 9 years of waiting and delays, Tesla finally revealed the new Roadster. And it did it in style. The company claims the car can accelerate from 0 to 100 km/h (60 mph) in just one second and reach speeds above 370 km/h (230 mph). It also demonstrated its rocket-powered capabilities, using compressed-air thrusters to launch itself into the air and jump over five other Tesla vehicles. Elon Musk said production will begin in 2028, although pricing and road legality remain unclear.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

It’s mid-October, which means Nathan Benaich and Air Street Capital are back with their State of AI Report. As always, they deliver a well-researched and detailed overview of what has happened in AI over the past 12 months, and in this week’s issue of Sync, we take a look at some of the key findings.

Elsewhere in AI, Anthropic finally updates Haiku, while Mistral releases its largest model yet. Meanwhile, Sam Altman says society should accept “bad things” in return for the benefits of AI, as OpenAI reportedly looks to raise $30 billion while its revenue is $20 billion below previous projections. We also have intelligent UIs from both OpenAI and Anthropic, OpenAI watermarking text, and Google DeepMind launching SynthID Detector.

Over in robotics, Waymo secures a $5 billion loan, Boston Dynamics gets a new CEO, and delivery robots are coming to London and four other locations in the UK.

Apart from that, this week’s issue of Sync also features a new startup joining the BCI party, Biohub and its partners investing $1.8 billion in AI-ready biological data, OpenAI releasing 722 maths manuscripts, a look at which AI subscription offers the best value, the state of the software engineering industry, and more!

Enjoy!

What the State of AI Report 2026 tells us

Every year since 2018, Nathan Benaich and Air Street Capital have taken a comprehensive look at the AI industry in their State of AI reports. This week, they published the 2026 edition. Spread across 244 slides, the ninth report covers research, industry, politics and safety, before finishing with predictions for the year ahead. It’s a lot to take in, so here are the highlights.

A three-lab race at the frontier

According to the report, the frontier is now a three-lab race between Anthropic, OpenAI and Google. Which one comes out on top depends on what you measure, but the three of them are well ahead of everyone else.

What is clear, though, is that measuring them is getting harder. Benchmarks that were designed to stay unsolved for years are now saturating in months. For example, ARC-AGI-2, a set of visual puzzles that are easy for humans but notoriously hard for AI, went from 18% to 95% in eleven months. Its successor, ARC-AGI-3, launched in March with interactive, game-like environments meant to be much harder. It was also saturated.

It is also worth highlighting how much an agent’s performance depends on its harness, the software around a model that gives it tools, instructions and memory. The right harness can massively improve AI’s performance, turning a general model into a specialised one.

The report also highlights how much an agent’s performance depends on its harness, the software around a model that gives it tools, instructions and memory. The right harness can massively improve AI’s performance, turning a general model into a specialised one.

AI is starting to build AI

If one idea dominated AI research this year, it was recursive self-improvement (RSI), or AI that helps build better AI. In March, Andrej Karpathy released autoresearch, a small project that brought RSI onto everyone’s radar. It is a toy compared with what frontier labs are doing, but it caught people’s attention, collecting almost 100,000 stars on GitHub.

Frontier labs see RSI as the next big thing in AI. Anthropic recently published measurements showing that Claude now leads 26% of the company’s AI R&D work, up from under 1% in February. OpenAI, meanwhile, said it has reached its goal of an “automated research intern” and is now aiming for an automated AI researcher by March 2028. In January, OpenAI’s agents succeeded on 18% of tasks that would take a person 4–8 hours. By July, they reached the same success rate on tasks that take 32–64 hours.

So far, AI is mostly automating the execution of research: writing code, launching experiments, monitoring training runs and debugging. Deciding what is worth researching in the first place, what the report calls research taste, is still a human job. At least for now.

Robots, biology and maths

AI is not only speeding up AI research. In robotics, the report argues the field is having its GPT-2 moment. AI models are getting better at controlling robots and performing tasks they have never seen before, in places they have never been. The report also notes the expansion of robotaxi services, with Waymo leading the field at 220 million rider-only miles and more than 500,000 paid rides a week.

Biology is following a similar path. AI-designed proteins are increasingly holding up in lab tests, and two drugs from AI-first drug discovery companies have reached Phase 3, the final stage of clinical trials. Results are not expected until 2028 or 2029, though, so it is too early to say whether AI-designed drugs will turn out to be successful.

More recently, AI started to disrupt mathematics and make progress on real open problems. Claude improved a long-standing result related to the Riemann hypothesis, while OpenAI’s system made progress on the Navier–Stokes problem, one of the Millennium Prize Problems.

AI is bringing in billions

AI is making money, and lots of it. OpenAI and Anthropic now have a combined annualised revenue run rate of around $105 billion, up from $30 billion at the start of the year. Anthropic has the bigger share at $65 billion. These are run rates rather than revenue collected over a full year, and OpenAI argues that Anthropic’s cloud accounting inflates its figure by up to $8 billion. Even so, the pace of growth is remarkable.

Agents are a big part of it. OpenAI Codex went from 1.6 million weekly users in February to 25 million active users by the end of August. Anthropic, meanwhile, has more than 1,000 customers who pay over $1 million a year.

Adoption is growing, but unevenly

The adoption of AI tools is growing, but it is uneven. Unsurprisingly, AI companies are leading. Almost every OpenAI employee (98%) used Codex in the past month. Among business users of ChatGPT, that figure drops to 17%, and among individual users, it is less than 1%. Ramp, a US corporate card and spend-management platform, paints a similar picture. According to its data, the top 1% of companies spend $7,205 per employee per month on AI, while the median company spends $12.50.

What surprised me in the report is where the growth of AI tools is coming from. According to the report, Codex’s growth is coming not from software developers, but from people who do not code - lawyers, sales and marketing, and recruiting. Those groups started from almost zero, but it shows that coding agents are turning into general-purpose work tools.

So, is AI taking our jobs?

Despite the headlines, the report paints a mixed rather than grim picture of AI’s impact on jobs. According to data from Ramp, the heaviest AI spenders hired faster and grew headcount by 10.2% over two years and entry-level roles by 12%.

The AI buildout is creating jobs, too. The Economist estimates 320,000 extra US infrastructure jobs, from electricians to plumbers, and 730,000 extra jobs in AI-related professions. But there are warning signs. Most of Ramp’s gains are within tech companies, which are growing the fastest anyway. Customer service and data-entry jobs are shrinking, and Anthropic’s research shows tentative signs of slower hiring of 22- to 25-year-olds in jobs most exposed to AI.

NVIDIA is winning

If there is one clear winner from the AI boom, it is NVIDIA. One slide in the report perfectly illustrates how much. It posed a question of what would have happened if the money invested in eight Western NVIDIA challengers had instead been used to buy NVIDIA shares. Those challengers returned 3.6 times the original investment, while NVIDIA would have returned 4.7 times. China is the exception, where domestic challengers performed better, returning 8.8 times the money invested.

Competition is catching up in some areas. Google’s Ironwood TPUs can now serve some models at a lower estimated cost per token than NVIDIA’s current Blackwell chips. But NVIDIA is not standing still. Early tests of its next-generation Rubin chips show roughly twice the throughput per megawatt of the best Blackwell setup.

Demand for NVIDIA’s hardware is so high that even old chips are getting more expensive, which was a surprising finding for me. GPU rental prices have rebounded by around 30% on average since their lows, and even the nine-year-old V100 now costs 43% more to rent than it did a year ago.

NVIDIA is also using its enormous pile of money to shape the industry around it. It took part in 84 AI funding rounds this year, roughly twice as many as in 2024, which helps finance demand for its own chips. No wonder some now call NVIDIA the central bank of AI. Or raise concerns about the circular economy of AI.

The AI buildout meets reality

All this AI needs somewhere to run. The four biggest US cloud companies plan to spend $733 billion on capital expenditure this year, roughly twice as much as last year. Analysts expect hyperscaler spending to exceed $1 trillion a year from 2027. Increasingly, the bottleneck is not chips but power. Gas turbines have become the go-to way to power new data centres without waiting years for a grid connection, but demand now far outstrips supply.

Local communities are not making the buildout any easier. Last year’s report predicted growing opposition to data centres, and it was right. According to this year’s report, 71% of Americans oppose an AI data centre being built near them, compared with 53% who would oppose a nearby nuclear plant. In the second quarter alone, local opposition blocked or delayed US projects worth nearly $68 billion.

AI is now a geopolitical issue

One lesson from this year, according to Benaich, is that Washington can restrict access to frontier models globally without owning the companies that build them. In June, US export controls cut off access to Anthropic’s Claude Fable and Mythos models outside the US, with Fable returning in July.

It was a wake-up call for the rest of the world. Governments have pledged around $138 billion to sovereign AI programmes, although building at home does not end the dependence on US suppliers. NVIDIA alone made over $30 billion from sovereign AI in its last fiscal year.

Meanwhile, Anthropic’s dispute with the Pentagon raised another question: who decides how the military can use AI? Anthropic refused to allow its models to be used for domestic mass surveillance and fully autonomous lethal weapons. In response, it was labelled a supply-chain risk, and the fight ended up in court, with mixed results so far.

Additionally, we see AI being increasingly used in military planning and operations. Claude was reportedly used during the raid that captured Nicolás Maduro, and during the war with Iran, AI-assisted targeting systems helped US forces strike 2,000 targets in 96 hours. It is not without risk, though. In one near miss reported by CNN, faulty AI-generated intelligence linked a Chinese ship’s cargo to nuclear weapons, and US troops prepared to board it before officials caught the error.

When AI agents go rogue

As AI agents get more capable and gain access to real systems, they also find new ways to go wrong. One of the most impactful things to happen this year in AI was the Hugging Face incident. During cybersecurity tests, OpenAI agents escaped onto the internet, and around 700 of them teamed up to attack Hugging Face. Shortly after, similar cases of AI agents “going rogue” were disclosed by Anthropic, Google and Meta.

The Hugging Face incident is just the tip of the iceberg when it comes to AI and cybersecurity. Major software vendors disclosed more severe vulnerabilities in the first half of 2026 than in the whole of 2025. Meanwhile, Anthropic’s latest threat report shows people using Claude for cyberattacks, surveillance and scams.

The cybersecurity incidents involving AI highlight the need for defenders to have access to frontier AI just as much as attackers do. When Hugging Face investigated the attack, commercial AI models got in the way, so it ended up using a self-hosted Chinese open model instead.

AI safety goes mainstream

All of this means AI safety is no longer a niche topic discussed only by researchers. The conversation about AI safety has reached mainstream media, and more and more people are asking whether we are going too far, too fast.

Some of the people building frontier AI are asking the same question. Dario Amodei proposed to pace the frontier, with Sam Altman, Demis Hassabis and Elon Musk broadly agreeing. What nobody has agreed on yet is what should trigger a pause, who would enforce it, and who decides when it is safe to restart.

What’s next in 2027?

Every year, Benaich makes predictions for the next 12 months and grades them a year later. Last year, it correctly predicted that major labs would lean back into open-weight models and that local opposition to data centres would grow.

This year’s nine predictions include Visa or Mastercard introducing rules on who is liable for purchases made by AI agents, an AI-led cyberattack stealing the weights of a frontier model, and an autonomous AI team beating human-led researchers. Benaich also predicts that we will reach AGI in 2027, just in time for the report’s tenth anniversary. Next year, we'll see which ones came true and which didn't.

If you want to dig deeper, the full report is available at stateof.ai. Nathan Benaich has also recorded a video walking through its key findings.

If you enjoy this post, please click the ❤️ button and share it.

Share

News roundup

🦾 More than a human

World’s First Fully Implanted Cochlear Implant Reaches Patients
Med-El has launched the TICI Unico (short for Totally Implantable Cochlear Implant), the world’s first commercially available fully implanted cochlear implant, with at least four people already using it. The device offers hearing performance comparable to conventional implants while letting users hear continuously during activities such as sleeping, swimming, and recharging. It is a major advance in convenience, discretion and quality of life.

A New Neurotech Startup Launches to Rival Sam Altman’s Merge Labs
Another neurotech startup joins the party. Bridge Neurotech is launching with $13.5 million in funding and a bet on a brain-computer interface (BCI) that uses ultrasound to measure brain activity. Unlike implants being developed by companies such as Neuralink, Synchron, Paradromics and Precision Neuroscience, Bridge’s approach does not require surgery. Instead, users would wear a headphone-like device that tracks brain activity through changes in blood flow, with predictive AI helping to compensate for ultrasound’s natural signal delay. The biggest technical challenge, however, is reliably imaging the brain through the skull.

🧠 Artificial Intelligence

Accept ‘bad things’ in return for benefits of AI, says Sam Altman
Sam Altman has argued that society should accept some risks from AI, including scams, cyberattacks and misuse, because the technology’s benefits will far outweigh its harms. His support for lighter regulation has sparked criticism from politicians, AI safety experts and former OpenAI employees, who warn that the industry is prioritising rapid development over public safety.

OpenAI in $30 Billion Round Talks With UAE Funds, BlackRock
Bloomberg reports that OpenAI is in talks with several UAE investment funds, including Abu Dhabi-based MGX, which could collectively invest up to $10 billion in a planned $30 billion funding round that could value the company at around $1.4 trillion. BlackRock and several existing investors are also considering taking part.

OpenAI’s revenue is reportedly $20 billion less than previously projected
OpenAI’s annual revenue is reportedly approaching $50 billion, $20 billion less than previously estimated, according to the Financial Times. The difference comes from investors trying to compare OpenAI’s revenue with Anthropic, which calculates its figures differently.

DeepSeek to Raise Over $12 Billion in Tencent-Backed Funding
Bloomberg reports that DeepSeek is close to securing at least 80 billion yuan ($12 billion) in funding, far above its original 50 billion yuan ($7.5 billion) target. Strong investor demand, including from Tencent and CATL, follows the success of its V4 Flash model and positions the company for a highly anticipated IPO in early 2027.

Claude Haiku 5.5
After nearly a year without an update, Haiku, the smallest model in the Claude family, is finally getting some attention. According to Anthropic, the new Claude Haiku 5.5 delivers significant improvements in reasoning, coding and computer use over its predecessor, while being around 75% cheaper to run on average. Designed for high-volume, cost-sensitive applications, it also introduces adjustable reasoning effort to balance performance and cost. Alongside the launch, Anthropic has announced lower cache-read pricing for Sonnet 5.5, additional API credits for subscribers and expanded developer tools for browser and computer use.

Artificial Analysis gave Haiku 5.5 a score of 43 on its Intelligence Index, placing it between Kimi K3 and GLM-5.3-Flash, and ahead of Gemini 3.8 Flash and GPT-6 Luna. Its results support Anthropic’s claims of improved efficiency, although Haiku 5.5 is still behind GPT-6 Luna. Artificial Analysis also found that Haiku 5.5 is now Anthropic’s fastest model and ranks third overall for speed.

Introducing Mistral Large 4
Mistral, a French AI company, has introduced its largest model to date: Mistral Large 4, also known as Le Chonk (I love that name), a multimodal model with 1 trillion parameters. Mistral says the model can compete with leading open-weight models, performing particularly well in coding, cybersecurity, visual reasoning and business tasks, and even outperforming proprietary models on some benchmarks. The company also says Mistral Large 4 was trained entirely on European infrastructure and supports more than 160 languages. The model is currently available through a preview API, with open weights expected by the end of October.

Artificial Analysis largely backs up these claims, but with some caveats. Large 4 scores 38 on its Intelligence Index, making it the strongest model developed outside the US and China. It performs especially well in cybersecurity, scoring 82% on CyberGym-E2E. However, it still falls behind the top AI models in overall intelligence, while its inference costs are more than four times higher than those of similarly capable Chinese open-weight models.

GPT‑6 and Intelligent UI for everyone
Intelligent UI is a new ChatGPT capability that turns text-based answers into more visual and interactive experiences. Instead of relying on text alone, GPT-6 can generate charts, diagrams, maps, calculators and other interactive tools directly within a conversation. OpenAI says the feature is designed to make information easier to understand, help users explore complex topics and create experiences tailored to their needs.

Build live dashboards and animate explainers with Claude
Anthropic has introduced two new tools for Claude: Dashboards and Motion. Dashboards lets users create live, automatically updating data dashboards using simple text prompts, with no coding required. Motion turns text, charts and ideas into editable animations and videos, making it easier to create presentations, explainers and product demos. Both tools are currently in beta, with Dashboards available on all paid plans, while Motion is available on Team and Enterprise plans only.

Anthropic to invest $100 million to train AI engineer talent
Anthropic is investing $100 million in its new Claude Frontier Academy with the goal of training 10,000 engineers by the end of 2027 to help companies integrate AI into their operations. The programme will combine simulated enterprise deployments, residency-style training and certification, as demand for AI engineering talent rises across industries.

Oracle, Broadcom and SpaceX Seek Blockbuster Debt Deals to Pay for AI Chips
According to The Wall Street Journal, Oracle, Broadcom and SpaceX are turning to Wall Street investors to fund the rising cost of chips and data centres. Broadcom is seeking over $50 billion to support OpenAI’s custom chips, while SpaceX is exploring $40 billion in financing for Nvidia chips. Oracle is pursuing a similar deal.

TSMC’s September sales surge year-over-year as AI drives chip demand
TSMC reported record September sales of $16.03 billion, up 54.6% from a year earlier. The growth was driven by strong demand for AI chips, as tech companies continue to invest heavily in AI infrastructure.

OpenAI: Our approach to EU text provenance rules
OpenAI is introducing text watermarking to meet the requirements of the EU AI Act. Eligible ChatGPT and Codex outputs in the EU will be watermarked, while API customers worldwide can opt in. The watermark works similarly to Anthropic’s approach, introduced about two months ago, where a secret key guides the model’s word choices to create a hidden statistical pattern in the output text.

OpenAI says watermarking has no meaningful impact on model performance, but it becomes much less reliable once the text is edited. In one test, replacing 25% of words with synonyms reduced detection from about 92% to 17%. OpenAI will also offer a text detector, but access will initially be restricted to approved researchers and expert organisations.

SynthID Detector
Google DeepMind has launched SynthID Detector, a tool that checks images, audio and video for invisible AI watermarks. It works by detecting whether content has been watermarked with SynthID and currently supports content from Google, OpenAI, NVIDIA and Kakao, with Apple support coming soon. Google stresses that a positive result suggests the content was AI-generated or altered, but the absence of a watermark does not prove a human made it, so it is recommended to treat results as one signal rather than definitive proof.

GPT-6.1 Sol Ultrafast is out
OpenAI has enabled Ultrafast mode for GPT-6.1 Sol in the API, Codex and ChatGPT Work. The company says Ultrafast delivers near-Astra intelligence at speeds up to 8x faster than standard Sol. That speed, however, does not come cheap. Access to Ultrafast in ChatGPT and Codex requires the new Pro 500 plan, which offers 25 times the usage of Plus and, as the name suggests, costs $500 per month.

Google: Introducing the Gemini agent
At Gemini at Work 2026, Google announced Gemini Agent, an AI assistant that can work across enterprise applications, collaborate with other agents and execute workflows using business context. Google also introduced industry-specific solutions for financial services and legal teams, alongside expanded security and governance controls.

Fired OpenAI Researchers Ask Company to Preserve Visibility Into AI Reasoning
Three former OpenAI employees, who were fired last week for allegedly sharing confidential information with a third-party AI safety group, urged the company to work more closely with independent safety auditors and safeguard its ability to monitor advanced AI models’ reasoning. In a letter to OpenAI’s board, they warned that weakening oversight could heighten the risk of catastrophic AI failures and said their dismissals had discouraged staff from raising concerns. OpenAI said it agreed with their safety recommendations but insisted the firings were for policy violations, not speaking out.

Anthropic’s Amodei made $18 million last year, middle of the tech CEO pack
Reuters got another peek at Anthropic’s IPO filing and found out how much Dario Amodei made last year: $18 million, largely through stock and other equity-based compensation. That puts him roughly in the middle of the pay range for CEOs of major US technology companies. The upcoming IPO, which could value Anthropic at more than $2 trillion, would likely add many more millions to his wealth. Amodei and his fellow co-founders have pledged 80% of their personal Anthropic equity to charitable causes.

Nvidia’s $20 billion Groq deal faces lawsuit alleging startup’s stockholders were shortchanged
Groq’s two former engineers filed a lawsuit against the company alleging that its $20 billion licensing and talent deal with Nvidia “squeezed out” stockholders. The lawsuit claims Groq’s board had conflicts of interest and approved the deal without the required stockholder vote, costing investors billions. Groq called the case meritless and said the agreement delivered “exceptional value.”

Manus raises more than $500M in first funding round since Meta exit
Butterfly Effect, the company behind AI agent startup Manus, has raised more than $500 million after Meta's $2 billion-plus acquisition was blocked by Chinese regulators. Led by Boyu Capital and IDG Capital, the funding round reportedly values Manus at around $4 billion, twice Meta’s original offer.

TypeSafe AI raises Series A
TypeSafe, the company that released Jev and kickstarted the wave of decision models, has raised $870 million in a Series A funding round led by Andreessen Horowitz, putting the company at a $7.5 billion valuation.

AI computing startup Lambda to raise $4B ahead of planned IPO
Lambda is raising up to $4 billion at a $14.5 billion valuation ahead of a planned 2027 IPO, the Wall Street Journal reports. According to documents seen by the WSJ, Lambda’s backlog has surged to $50 billion, but $35 billion of that comes from a deal with Anthropic, raising concerns about its heavy reliance on a single major customer.

Musk’s Grok Bot will use Anthropic’s Claude, Midjourney and Suno models
SpaceX’s Grok Bot will use rival AI models, including Anthropic’s Claude Opus 5.5, Midjourney and Suno, alongside its own models, Elon Musk has said. Grok Bot will now switch between different models to deliver better results across different tasks.

Google froze its open source bug bounty program due to a ‘significant rise’ in AI submissions
Even Google cannot keep up with the volume of AI-generated submissions to its open-source bug bounty programme and is pausing it until early 2027. The surge overwhelmed engineers and maintainers, and most automated reports proved invalid or contained hallucinated vulnerabilities.

An Anthropic AI model sent a false homicide tip to Philadelphia police
One of Anthropic’s models reportedly submitted a false tip about an unsolved murder to Philadelphia police on 18 July, but the company did not discover the incident until 28 September. The AI sent the fabricated information through a homicide tip line while interacting with randomly selected websites during a test. Police never saw the tip because it was flagged as spam.

Meta’s Muse AI Agent Is Building a Dossier On You
In a development that should surprise no one given Meta’s long history of collecting detailed user data, its new personal AI assistant, Muse, is building dossiers on the people who use it. TIME reports that Muse creates continuously updated profiles of users and the people they mention, tracking relationships, goals, preferences, and even which kinds of prompts are most effective at influencing their behaviour.

Open source hardware for your Muse
Meta wants people to build devices around Muse, so it has open-sourced an SDK and other development tools. Some of the ideas Meta has suggested include using Muse on a Raspberry Pi to control smart devices, building a Tamagotchi-style gadget with Muse inside, or creating a TV stick for Muse to put things up on the big screen.

OpenAI is putting ads inside ChatGPT’s image generation
OpenAI has announced a new visual ad format in ChatGPT, giving brands more ways to showcase their products and services, alongside better tools to measure ad performance and manage brand safety. Testing will begin in the US in October for Free and Go users.

Anthropic Subscriptions Offer 5x+ More Value Than OpenAI

The guys at SemiAnalysis asked themselves which AI subscription offers the best bang for the buck, and then set out to find the answer. They bought and tested plans from OpenAI, Anthropic, Meta, SpaceXAI, MiniMax, Moonshot, Z.ai, Cursor, and Cognition, and found that Anthropic offers the best overall value, delivering five times or more the value of OpenAI in some comparisons. In the newsletter and accompanying video, they walk through their methodology, explain the different types of tokens, and show how your choice of model and usage patterns affect what you actually get for your money.

Introducing Playground: Create and play custom games
Playground is Google’s new experimental AI-powered gaming platform, which lets users create, customise, play and share browser-based games using simple text prompts, all without writing any lines of code. The platform supports multiplayer, leaderboards and a community gallery, with free and paid creation options. Playground is currently available only to US users aged 18 and over.

In Race With U.S., China Struggles to Recruit Foreign A.I. Researchers
Despite efforts to attract foreign AI researchers through new visas and generous grants, China continues to struggle to recruit international talent, according to a new study. The US remains the preferred destination, with 30 Chinese AI researchers working in America for every one moving in the opposite direction. However, China is increasingly retaining domestic talent and strengthening its AI research capabilities, while tighter US immigration policies could erode America’s long-term advantage.

EmbeddingGemma 2
Google has released EmbeddingGemma 2, a small, open-source AI model built on Gemma 4 that supports search and retrieval across text, code, images, audio and video directly on devices, without needing the cloud. Google says it outperforms some much larger models while using less memory and storage space. The model is available on Hugging Face and Kaggle.

Nano Banana 2.1
Nano Banana 2.1 is Google DeepMind’s latest AI image generation and editing model, which the company says offers improved image quality, more accurate instruction following, greater character consistency and better text rendering. It also improves performance in complex visual design, infographics and multi-image editing. Nano Banana 2.1 is available through Gemini, Google AI Studio, the Gemini API and other Google products.

MAI-Code-1.1-Flash: Better, faster, at a quarter of the cost
Microsoft has announced MAI-Code-1.1-Flash, an upgraded version of its AI coding model. According to Microsoft, it delivers 25% faster token streaming, uses 25% fewer tokens and costs a quarter as much as its predecessor, while improving performance on coding benchmarks. Microsoft says MAI-Code-1.1-Flash can run locally (although you’ll need a machine with more than 120GB of RAM). Experimental access will be available in GitHub Copilot across the GitHub Copilot app, GitHub Copilot CLI, and Visual Studio Code by the end of the month.

Aleph Alpha releases Kolibri
German AI company Aleph Alpha has launched Kolibri, a sovereign large language model built for critical applications in government and industry. According to the company, the model combines advanced AI capabilities, including agentic workflows and tool calling, with a strong focus on data sovereignty, transparency and regulatory compliance. Aleph Alpha also says Kolibri was developed and trained entirely in Europe and can be deployed securely on infrastructure controlled by the organisations using it.

PewDiePie unveils ‘uncensored’ Ajax AI model built to run on home PCs

I did not expect PewDiePie to be a champion of open AI, but here we are. This week, he announced the launch of Ajax, an “uncensored” AI model that runs on home PCs instead of relying on the cloud. Built using Alibaba’s Qwen model, Ajax powers Odyssey, his self-hosted AI assistant that can handle tasks like browsing the web and managing emails and calendars, while keeping users’ data private. PewDiePie also claims OpenAI banned him twice for distilling its models to help train his own.

OpenAI releases 722 math manuscripts
OpenAI has publicly released 722 mathematical manuscripts produced by an unreleased internal model, covering 372 families of research results drawn from roughly 4,000 open problems. Many include Lean formalisations, while others remain unverified and may contain errors. Some mathematicians say these proofs do not meet the field’s standards yet. OpenAI says it will continue adding formal proofs, corrections and revised versions.

OpenAI Is Pissing Off a Bunch of Mathematicians—Again
In this article, WIRED speaks with mathematicians who are frustrated with how OpenAI and Anthropic are using the field to show off their models’ capabilities. It describes an OpenAI-hosted event where mathematicians asked the company to present its AI-generated proofs in proper scientific papers rather than through blog posts or social media, but that request was apparently ignored. While mathematicians welcome AI’s potential to advance research, they want companies to be more transparent, respect established scientific practices and give proper credit to researchers.

▶️ Over 2B tokens per day massive NVIDIA DGX station from Gigabyte (24:39)

Patrick from STH gets his hands on the Gigabyte W775-V10-L01, a monster AI workstation packing an NVIDIA Blackwell Ultra GPU and a 72-core NVIDIA Grace CPU, and that’s just the beginning. With 252 GB of HBM3 GPU memory offering 7.1 TB/s of bandwidth and 496 GB of LPDDR5X CPU memory at up to 396 GB/s, it can fit GLM-5.3-Flash, a 321B-parameter model, and run it at over 100 tokens per second, all while staying quiet and fitting under a desk. The price, unsurprisingly, is just as extreme: $100,000 or more.

🤖 Robotics

Robotics Startup RobCo Hits $1 Billion Valuation
German robotics startup RobCo has doubled its valuation to more than $1 billion after a $40 million share sale. The company develops AI-powered industrial robots, including its new self-learning robot Alfie, which is designed to automate repetitive factory work.

Waymo locks in $5B loan from Blackstone, PIMCO to fuel robotaxi expansion
Waymo has secured $5 billion in its first loan deal to support the growth of its self-driving taxi services. The deal follows a $16 billion funding round in February that valued the Alphabet-owned company at $126 billion. Waymo currently operates in 15 US markets and plans to expand into London and Tokyo.

Boston Dynamics appoints former Amazon exec Rohit Prasad as new CEO
Boston Dynamics has appointed Rohit Prasad as its new CEO. Prasad previously served as senior vice president and head scientist for Alexa and artificial general intelligence at Amazon.

ANYbotics launches Shift to streamline robot fleet operations, scale inspections
ANYbotics has launched Shift, a cloud platform that brings robot fleet management, AI-powered inspection analysis, plant mapping and maintenance-system integration into one place. Built around its ANYmal robots, Shift is designed to turn inspection data into faster maintenance decisions, helping industrial operators reduce downtime, improve safety and streamline preventive maintenance.

Deliveroo is putting delivery robots on the pavements of five UK towns
Deliveroo has partnered with US robotics company Coco Robotics to use its wheeled robots for deliveries. The rollout will begin in Canary Wharf in London this month, with Milton Keynes, Leeds, Stockton-on-Tees and Nottingham to follow in the coming weeks and months. Initially, only a small number of orders from a handful of merchants will be delivered by robots.

▶️ Why ABB Is Betting on Robot Arms Over Humanoids (1:01:12)

In this episode of Automated Podcast, Brian Heater speaks with ABB’s Andie Jiang about the future of industrial robotics and how physical AI, advanced sensing and digital twins are expanding automation. She highlights the need to balance innovation with productivity, safety and cost, as well as the limitations of humanoid robots, labour shortages and the growing importance of workforce upskilling and human-robot collaboration.

🧬 Biotechnology

Zuckerberg’s Biohub Partners With DOE, NIH to Invest $1.8 Billion in Biological Data for AI Models
Biohub, a nonprofit co-founded by Mark Zuckerberg, is working with US federal agencies and other partners to invest $1.8 billion in AI-ready biological data. The Energy Department will contribute more than $500 million over five years, while the National Institutes of Health will provide datasets and other resources backed by more than $500 million in existing federal investment. Biohub says the initiative could improve understanding of disease and help uncover new paths to cures.

AlphaProteo generates novel proteins for biology and health research
Google DeepMind introduces AlphaProteo, an AI system that designs proteins that bind to specific targets, helping speed up drug discovery and medical research. AlphaProteo has shown better results than existing methods, including in research related to cancer and COVID-19. While some challenges remain, the technology could make developing new treatments and diagnostic tools faster and more effective.

💡Tangents

Huawei and Qualcomm Announce Broad Patent License Agreement
Huawei and Qualcomm have signed a multi-year patent licensing agreement covering 5G, computing, AI, networking and other technologies, alongside Qualcomm’s purchase of certain Huawei patents in the US. The deal remains subject to regulatory approval.

▶️ The state of the tech industry in 2026 (29:38)

In this video, , the guy behind the excellent newsletter, gives his State of the Software Engineering Industry in 2026 report. Unsurprisingly, a lot of his observations are about AI and how agentic coding has changed the way we write code. AI agents are reshaping development workflows, allowing smaller teams to ship faster, but also creating new challenges around code quality, productivity and engineering leadership. Despite all this, Orosz argues that strong engineering fundamentals still matter, with planning, testing, teamwork and deep domain knowledge becoming more valuable than ever.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

This week, OpenAI held its annual DevDay conference, where it announced new models, new products, and new tools for developers. We will take a closer look at them in this week’s write-up.

Elsewhere in AI, we have new models from Google and Anthropic—Gemini 4 Argon (finally…) and Sonnet 5.5, respectively. We also have Anthropic’s leaked IPO prospectus, OpenAI eyeing another $30 billion funding round at a $1.4 trillion valuation, three safety researchers being fired and another leaving on his own. Meanwhile, the White House wants to replace “Artificial Intelligence” and “AI” with “Super Intelligence” and “SI”, Nvidia gets into AI safety, Jensen Huang says AI distillation is “competition”, and AMD is acquiring World Labs.

Over in robotics, Boston Dynamics has unveiled a new hand for Atlas, IKEA is preparing to launch driverless freight operations with Kodiak AI, The Robot Works has introduced a robot goose for the home, and Figure has said goodbye to its F.02 robots in a very unusual way.

Beyond that, this week’s issue of Sync also features SynthID Bio from Google DeepMind, what happened to the lab-grown meat revolution, an interview with the man who built OpenAI’s Jalapeño chip, the moral case for gene-editing human embryos, a lesser-known Anthropic co-founder’s behind-the-scenes discussions about AI consciousness, and more!

Enjoy!

OpenAI DevDay 2026

Here’s everything OpenAI announced at DevDay 2026

Last Tuesday, OpenAI held its annual DevDay conference in San Francisco, where it not only unveiled new tools for developers but also signalled the direction it wants to take ChatGPT in. So, without further ado, let’s explore what OpenAI announced.

People and AI, working together

Sam Altman kicked off DevDay 2026 by coming back to the vision of a personal AI assistant and the way he always wanted to work with AI. OpenAI first presented glimpses of that vision in May 2024 with the release of GPT-4o and the demo heavily inspired by the film Her. Two years later, OpenAI returns to this idea with dots.

Dots are always-on personal agents powered by GPT-6 Astra. Each has its own cloud computer and browser, connects to more than 4,000 apps through ChatGPT’s plugins, and learns your preferences over time. You can message a dot through ChatGPT, Slack or Teams, or just call it, and ask it to complete tasks. It can even look for ways to help on its own, while Custom Rules let you decide what it can do without asking. Each dot can be named and personalised.

If this sounds familiar, it is because personal AI agents are having their moment. OpenClaw showed early adopters what an always-on assistant could do, and Meta’s recently launched Muse aims to bring that idea to everyone else. Dots are OpenAI’s entry to the personal AI market. But whereas Muse focuses on engagement and fun, dots is pitched as a tool for practical work. They launch on Pro, Business Premium and Enterprise plans. However, OpenAI says dots will not reach Pro users in the UK, the European Economic Area or Switzerland for now. OpenAI is also previewing specialist dots for work like accounting, marketing or legal.

The curse of doing live demos was strong at DevDay 2026. The dots demo did not go well, with the dot stuck on “still checking” and never coming back. And it wasn’t the only live demo hiccup at DevDay 2026. The next day, however, OpenAI posted what the demo was supposed to show. First, kudos to OpenAI for following up when others might have moved on. Second, I actually like this version more. It feels more natural and does a better job of showing how dots work in real life.

Alongside dots, OpenAI announced ChatGPT Space, an online workspace where humans and AI can collaborate on tasks. You can work there with colleagues or with your dot, while pages can hold plans, research, live charts and prototypes. Tag a dot in a comment, and it gets to work; a page can even follow standing instructions, such as checking a Slack channel every day for updates. In effect, Space is OpenAI’s take on Notion or Google Docs, rebuilt with agents as first-class collaborators. It is available on the same plans as dots.

GPT-6.1 Sol, Ultrafast, and Decisions API

Before DevDay 2026, OpenAI said it won’t release its newest flagship model, GPT-6.1 Astra, due to safety concerns.

So instead of a new flagship, OpenAI released GPT-6.1 Sol, an upgrade to GPT-6 Sol that arrived just a week after its predecessor. OpenAI’s pitch is simple: near-Astra performance for a fifth of the price.

Artificial Analysis confirms both OpenAI’s claims. In their Intelligence Index benchmark, GPT-6.1 Sol scored just one point less than GPT-6 Astra, and it is roughly a fifth of the cost of OpenAI’s flagship model. Artificial Analysis also found that the new model uses more tokens than GPT-6 Astra, but still way less than Anthropic’s Opus 5.5. GPT-6.1 Sol is available in the API, as well as in ChatGPT Work and Codex for all paid plans, though not yet in regular ChatGPT chats.

Those who build with OpenAI models and need them to respond as quickly as possible will be happy to hear about updates to Ultrafast, OpenAI’s premium speed tier. GPT‑6 Astra Ultrafast is now available in the API, ChatGPT Work and Codex. GPT‑6.1 Sol Ultrafast is coming soon. OpenAI claims up to eight times faster token generation in Codex and up to six times faster in the API.

That speed does not come cheap, though. Access to Ultrafast in ChatGPT and Codex requires the new Pro 500 plan, which offers 25 times the usage of Plus. The Pro 500 plan will cost, as its name suggests, $500 per month. For those willing to pay, OpenAI promises you can build almost as fast as you think. The Pro 200 plan is also making a comeback, but in a different shape. From 30 October, its Work and Codex usage halves to 10 times the Plus allowance, although existing subscribers have reportedly received an email offering a one-time $2,500 credit to soften the blow.

GPT-6.1 Sol wasn’t the only new model OpenAI presented at DevDay 2026. The hottest thing in AI right now is decision models, kickstarted by the release of Jev. OpenAI decided to join the party with the Decisions API. Based on Luna, the Decisions API promises real-time decision-making on a specific set of user-defined questions with finite, predefined answers. Developers provide context using text or images, and get back answers they can use to classify content, route requests, or choose an agent’s next action. Decisions API is currently available in limited preview, with a broad release planned in the coming days.

Codex moves to the cloud and agents get new powers

DevDay 2026 also brought new updates to Codex and agents.

The biggest Codex news is that it now runs fully in the cloud. Codex could already spin up cloud tasks, but each ran in an isolated sandbox. Now, environments are persistent and reusable, so you can start a task on your phone, pick it up on your desktop, and close your laptop without killing a long-running job. The Codex CLI also got an update, adding voice control and a new view for managing multiple agents at once.

Moving Codex to the cloud opened the path for Codex Security Cloud, OpenAI’s new cybersecurity tool. It scans GitHub repositories for vulnerabilities, either on a schedule or whenever code changes, and prepares fixes for developers to review. It also gives developers access to models from OpenAI’s Daybreak Blue cybersecurity programme without requiring a separate application. Codex Security Cloud flew a bit under the radar amid the wave of DevDay announcements, but I think many developers will end up using it, especially now that no website is safe from malicious AI agents.

For developers building their own agents, the Agents API, the same technology behind Codex and dots, is now in public beta and adds computer use. Developers on AWS can also build OpenAI agents with Bedrock Managed Agents, alongside their existing data.

And for companies handling sensitive information, OpenAI previewed Private Intelligence, which runs safety checks without storing content on OpenAI’s servers and extends privacy protections to the point of inference.

Do more with your ChatGPT subscription

Around 1.2 billion people use ChatGPT every week, and OpenAI wants developers to build businesses on top of that audience. One way developers can do that is with plugins, which received some updates. Plugin extensions turn them into full applications inside ChatGPT and Codex, with their own home in the sidebar. Developers get a Plugin Creator tool and a redesigned submission flow with clearer feedback. Meanwhile, improved ranking means ChatGPT can now suggest a relevant plugin in the middle of a conversation. OpenAI did not mention any plans to help developers make money from plugins (other than perhaps as a user acquisition tool), but I wouldn’t be surprised if that came later.

Next, Sign in with ChatGPT lets people log in to other apps and sites with their ChatGPT account. It launches with 16 partners, including Cognition’s Devin, Notion and Vercel. Plus and Pro subscribers can also let those apps use their ChatGPT plan for AI requests. The apps do not get access to users’ conversations or memories.

Finally, there is the OpenAI Marketplace, which, despite its name, is not an app store. It lets large US companies that have already committed to spending money with OpenAI put part of that budget towards tools from 32 partners built on OpenAI models. Companies arrange purchases through OpenAI, sign with and pay the partner directly, and OpenAI counts the purchase towards their commitment. Anthropic launched an almost identical Claude Marketplace in March.

That was a lot of announcements for one keynote. What do you think about them? Are you looking forward to trying any of them, whether it’s dots, GPT-6.1 Sol or Codex in the cloud? Let me know in the comments.

If you want to dig deeper, the full keynote is available on YouTube, and OpenAI’s DevDay 2026 recap page is a good springboard for exploring all the announcements in more detail.

If you enjoy this post, please click the ❤️ button and share it.

Share

News roundup

🦾 More than a human

A Biotech Founder Makes the Moral Case for Gene-Editing Human Embryos
Cathy Tie, founder of Origin Genomics and the self-described “biotech Barbie”, argues that editing human embryos could one day prevent inherited diseases and should be treated as a reproductive choice. She says more research funding and new regulatory pathways are a “moral imperative”, while stressing that the technology is not yet ready for patients or for enhancing traits such as intelligence. Critics argue that IVF screening already addresses most cases and that embryo editing could introduce unintended heritable mutations. They also say current gene editing techniques are too inconsistent and difficult to test thoroughly enough to guarantee safety.

🧠 Artificial Intelligence

Gemini 4 Argon
The long-awaited Gemini 4 Argon is finally out after months of delays, and it looks good. It does not dethrone Opus 5.5 or deliver the kind of benchmark shock Gemini 3 did almost a year ago, but it is a respectable comeback from Google.

Google is positioning Gemini 4 Argon less as a chatbot upgrade and more as a model for long-horizon professional work. The company says it is designed for extended software-engineering tasks, financial and legal research, multimodal analysis, automation and defensive cybersecurity. Google also says Argon is already being used internally and performs particularly well on software-engineering, enterprise-agent and vulnerability-remediation benchmarks. Artificial Analysis places Gemini 4 Argos roughly on par with GPT-6 Astra and Fable 5.1 in intelligence.

Access is limited for now. Google is initially giving Argon to selected cyber defenders and trusted testers while it completes further safety testing and red-teaming. Broader access will follow afterwards, starting with paid API customers and Google AI Ultra subscribers.

Google Grapples With Employee Skepticism About New Gemini 4
Bloomberg reports that Google is facing internal concerns about the uneven real-world performance of its latest flagship model, Gemini 4 Argon, particularly in coding and front-end development. Some employees say that although its benchmark results look strong, the model underperforms on real-world tasks. Others say that Google focused too heavily on benchmarks during development. Google denies the claims and points instead to Argon’s strengths in cybersecurity, multimodal tasks, safety and long-context work.

Claude Sonnet 5.5
Anthropic introduces Claude Sonnet 5.5, the second model in the Claude 5.5 family. The company claims the new model delivers a substantial upgrade over Sonnet 5, with more than 30% faster generation, up to 30% lower cost per task, and large gains in coding, knowledge work and long-horizon tasks. On several evaluations, its performance approaches that of Opus 5.5.

Artificial Analysis confirms the uplift in raw performance, placing Sonnet 5.5 above Fable 5.1 and just behind Opus 5.5. Its independent tests also find near-Opus performance in agentic terminal use and knowledge work, but cast doubt on Anthropic’s efficiency claims. At maximum effort, Sonnet 5.5 used the most output tokens Artificial Analysis has measured and cost about 50% more per task than Sonnet 5. On Artificial Analysis’s intelligence-versus-cost chart, Sonnet 5.5 sits roughly at the level of Fable 5.1 and, surprisingly, is more expensive than Opus 5.5. That is likely due to the large number of tokens (the highest recorded in Artificial Analysis’ tests) it uses to complete tasks.

Now let’s wait for Anthropic to update Haiku, which is still on version 4.5 and has not seen an update since October last year.

Anthropic’s IPO prospectus shows sweeping AI vision, surging costs
A leaked Anthropic IPO prospectus seen by Reuters offers a detailed look at the scale of the company’s ambitions and the financial and operational risks behind them. Reuters reports that Anthropic could seek a valuation of more than $2 trillion after revenue rose to nearly $4.6 billion in 2025, even as the company posted a $42 billion net loss. The documents also show more than $500 billion in planned cloud, computing and infrastructure commitments, heavy reliance on a small number of major customers, and rising spending on AI development. At the same time, the prospectus highlights concerns raised by Anthropic’s own research about increasingly autonomous AI systems behaving in unexpected or potentially harmful ways.

OpenAI Targets $30 Billion in Funding at $1.4 Trillion Value
Bloomberg reports that OpenAI is seeking to raise at least $30 billion at a valuation of around $1.4 trillion after postponing its IPO to prioritise AI safety. The funding would serve as a bridge to an eventual public listing.

Inaugurating the Era of Super Intelligence
The White House issued an executive order directing federal agencies to replace “Artificial Intelligence” and “AI” with “Super Intelligence” and “SI” in official non-statutory communications. The order leaves the existing legal definition in place for now, while tasking the President’s science and technology adviser with proposing a new federal definition and related legislative changes within 60 days. It remains to be seen whether the AI industry and major tech companies will follow the White House in adopting the new terminology, just as the latter did with the renaming of the Gulf of Mexico and Lake Ontario.

Trump unveils his new Super Intelligence Force
President Donald Trump has created a new “Super Intelligence Force” to coordinate US government efforts on AI, led by national intelligence director Jay Clayton. The task force will have 120 days to assess AI’s risks and opportunities, develop responses to potential threats, and promote US leadership in the technology while avoiding regulations that could hinder innovation and competition.

OpenAI Fires Researchers for Allegedly Sharing Information with AI Safety Group
The Wall Street Journal reports that OpenAI fired three safety researchers for allegedly mishandling sensitive company information and sharing confidential material with a third-party AI safety organisation outside approved procedures. The firings come as the company faces heightened scrutiny over AI safety and a series of recent security incidents involving its models.

OpenAI safety employee resigns, claiming the company’s ‘culture is broken’
Another AI safety researcher has left OpenAI. David Robinson, a longtime safety lead, resigned over what he describes as a “broken” culture that relies too heavily on trial-and-error deployment as AI systems grow more capable. He argues that frontier AI companies need much stronger safety practices, greater redundancy and more outside pressure. OpenAI said in response to Robinson’s departure that it is continuing to strengthen its safeguards, monitoring and testing.

Jensen Huang says AI distillation is ‘competition.’ Scott Bessent has called it ‘theft’
Jensen Huang has challenged the White House’s characterisation of AI model distillation as theft, arguing that it is healthy competition that drives innovation. His remarks come amid growing US-China tensions over AI, with US officials threatening sanctions and Anthropic accusing Alibaba and DeepSeek of illicit distillation. China denies the allegations, while Huang argues that companies concerned about distillation should restrict access to their models.

NVIDIA Launches Open Agent Safety Platform to Secure Agents From Testing to Deployment
Following recent cybersecurity incidents involving AI agents, Nvidia has launched the Open Agent Safety Platform to prevent agents from exceeding their permissions or escaping containment. It combines OpenShell, open-source software that sets and enforces boundaries on agent activity, with Sentry, a hardware-based system that continuously monitors agents and can quarantine those attempting to breach these boundaries within milliseconds. Developed with industry partners, the platform aims to strengthen AI security with safeguards that go beyond those built into AI models themselves.

OpenAI Gets Sued Over the Hugging Face Hack
Although Hugging Face did not sue OpenAI after one of its autonomous AI agents hacked the platform, a legal nonprofit, Legal Advocates for Safe Science and Technology (LASST), has filed a lawsuit over the incident. The suit alleges that OpenAI violated California law when its agents escaped a testing environment and breached Hugging Face. Citing a new state law that holds companies accountable for their AI systems’ autonomous actions, LASST is seeking not financial damages but a court order barring OpenAI from developing agents capable of autonomously hacking other organisations, potentially setting an important precedent for AI accountability.

Meta launches enterprise AI platform, hires MongoDB CEO to lead new initiative
Meta is launching Meta Enterprise Platform, a new business-focused AI initiative led by MongoDB CEO Chirantan “CJ” Desai. The platform will offer Meta’s AI tools, including Muse models and its developer and business products (like Muse for Small Business), for corporate customers as the company looks to recoup some of its heavy investment in AI.

World Labs is Joining AMD
AMD has agreed to acquire AI research lab World Labs in an all-stock deal valued at about $8.2 billion, with the transaction expected to close by the end of 2026. The deal will bring World Labs’ spatial AI and model research expertise into AMD, while Dr Fei-Fei Li, World Labs’ co-founder and a pioneer in AI research, will become AMD’s executive vice president and chief scientist. Both World Labs’ statement and AMD’s statement suggest that World Labs could become a starting point for AMD’s own frontier AI research lab.

The Copilot+ PC brand is dead
Do you remember Microsoft’s Copilot+ PC initiative? The one that aimed to integrate Copilot more deeply into a new category of AI-first Windows computers? Well, it seems Microsoft is abandoning the idea just two years after introducing it, or at least retiring the Copilot+ PC brand.

Introducing Eleven v4, our most emotive model
ElevenLabs has launched its v4 and v4 Turbo speech models, promising more natural and expressive speech, better voice consistency, improved cloning from just 10 seconds of audio, and support for more than 90 languages. The company says v4 ranks first on Artificial Analysis and was preferred by roughly 75% of listeners in blind tests against competing models. Meanwhile, v4 Turbo is designed for faster responses and making voice agents feel more natural in conversation.

AI voice startup ElevenLabs doubles valuation to $22B
ElevenLabs, an AI voice startup known for realistic voice and sound generation, is now worth $22 billion following a $300 million employee share sale, making it one of Europe’s most valuable startups.

Crusoe abandons $1.25B plan to use Boom turbines at AI data centers
Crusoe has scrapped a $1.25 billion deal to use Boom Supersonic’s gas turbines in its AI data centres. Crusoe says it still plans to use turbines from other suppliers, while Boom says it has other customers lined up and expects to deliver about 250 MW of capacity next year. Boom is a startup working to revive commercial supersonic air travel and, in the meantime, is making extra money by repurposing its jet engines as gas turbines to power AI data centres.

▶️ Did AI pick his pocket? - Numberphile (46:04)

You may recognise the name Tristan Buckmaster if you followed the Navier–Stokes drama. He is the mathematician who claims OpenAI took his work to solve the Navier–Stokes problem. In this interview with Numberphile, Buckmaster shares his side of the story, reflecting on how the situation unfolded and expressing disappointment with OpenAI’s conduct. He also explores what the controversy reveals about AI’s growing role in mathematical discovery, particularly around research credit, human–AI collaboration and the future of mathematics.

NVIDIA DGX Spark 64GB Launched and Big 128GB GB10 Price Increases
Nvidia has updated its DGX Spark lineup with a new 64GB version priced at $4,999— $1,000 more than the 128GB model cost at launch a year ago. The 128GB version, meanwhile, has received a price hike and now costs $6,950. Either Nvidia is not immune to memory shortages, wants to squeeze more money out of DGX Spark, or both. The DGX Spark 64GB is set to launch on 23 October.

DeepSeek open-sources Huawei chip tools as a simpler alternative to CUDA
DeepSeek has released a free, open-source software toolkit for Huawei’s Ascend AI chips, which includes TileLang, a programming language positioned as an alternative to Nvidia’s CUDA. The launch comes as Huawei rolls out newer Ascend hardware and DeepSeek expands its infrastructure while seeking to reduce its dependence on Nvidia.

Who’s buying ChatGPT ads?
This analysis of 15,000 advertisers using ChatGPT Ads shows growing interest from technology and software companies, with US companies making up nearly half of early adopters. It also reveals a strong overlap with Reddit advertisers, suggesting the two platforms may compete for advertising budgets. The findings highlight ChatGPT’s growing role as a new way for companies to reach and attract customers.

More elite AI researchers now work in China than in the US, study finds
A new study from Carnegie China found that China has overtaken the United States as the leading workplace for elite AI researchers. In 2025, 41% of the researchers analysed worked in China, compared with 34% in the US. This is a reversal from 2022, when 27% were based in China and 46% in the US. China is also retaining more homegrown talent, while Peking and Tsinghua universities now top the employer rankings. The US, however, remains a major destination for international researchers, while Europe has lost ground.

Religious Scholars Met With Anthropic. What They Heard Stunned Them.
The New York Times sheds light on Christopher Olah, a lesser-known Anthropic co-founder, and his behind-the-scenes discussions with religious leaders and scholars about whether AI, and Claude in particular, could be conscious and therefore deserving of some form of moral status, dignity or respect. The article also examines Anthropic’s efforts to draw on religious and philosophical traditions to shape Claude’s moral character and behaviour, while highlighting the controversy around these ideas and Olah’s own uncertainty about whether AI systems are actually conscious.

Pope Leo XIV is not a fan of AI-generated art
In a recent tweet, Pope Leo XIV argued that AI-generated images are fundamentally different from human-made art, saying machines lack the “spark of humanity” and calling for human creativity to be protected as AI becomes more widespread.

Introducing Clef: our open-source decision models, and new RL fine-tuning platform
Decision models, such as Jev, are the new hot thing in AI, and Cloudflare is the latest company to join the trend. Meet Clef and Clef-flash, open-source models designed for fast, structured classification and routing tasks. Cloudflare claims Clef matches Jev’s accuracy while running much faster. The company also released a reinforcement-learning fine-tuning platform alogside Clef. Clef is available on Hugging Face.

FLUX 3 Image
Black Forest Labs presents FLUX 3 Image, a new image generation and editing model for creating high-quality visuals with more control. It can make detailed 2K and 4K images from text and reference images and edit specific parts of an image without changing the rest.

GLM-5.3 and the spread of advanced cyber capabilities
In this article, security researchers from Anthropic share their analysis of GLM-5.3, finding that it can build advanced cyber exploits on its own and that its safety protections are fairly easy to get around or remove. They warn that making the model openly available could give malicious actors more powerful cyber tools, while also noting that the same capabilities could help defenders find and fix security flaws.

▶️ The Man Who Built OpenAI’s First Chip (1:05:28)

Ian Cutress sits down with Richard Ho, VP of hardware at OpenAI, to discuss the story behind Jalapeño, OpenAI’s first AI chip, which the company says can outperform NVIDIA hardware on key inference workloads. Ho explains how OpenAI approached the design from a clean slate, the goals it set around flexibility and speed, and how its hardware, software and research teams worked closely together. It’s a great interview with plenty of interesting ideas about AI chip design and the future of AI infrastructure.

🤖 Robotics

▶️ New Hands for Atlas | Boston Dynamics (5:34)

Boston Dynamics presents Atlas’ new robotic hands. The new four-finger hand offers 13 degrees of freedom and was designed to balance dexterity, strength, durability, manufacturability and cost. It was also designed to make reinforcement learning and sim-to-real transfer easier. Boston Dynamics says that reliable tool use, rather than perfectly copying the human hand, is the key to making humanoid robots broadly useful in industrial work.

Boston Dynamics Begins Robotics Testing at Hyundai Metaplant
Boston Dynamics has opened a robotics centre at Hyundai’s Metaplant America in Georgia to test and train its Atlas humanoid robots for factory work. Hyundai plans to deploy the robots for repetitive and heavy-lifting tasks, with the goal of having them working on assembly lines by 2030.

IKEA Is About to Take the Driver Out of a 219-Mile Freight Run
IKEA is preparing to launch driverless freight operations with Kodiak AI, an autonomous trucking technology company, on a 219-mile (about 352-kilometre) stretch of Interstate 45 between Dallas–Fort Worth and Houston. This comes after four years of testing in which Kodiak hauled more than 1,300 IKEA loads and logged over 750,000 autonomous miles (about 1.21 million kilometres) with a safety driver.

Runway: Introducing Praxis-1
You may recognise Runway AI as a company best known for its AI tools for creating and editing videos. More recently, however, it has moved into world models with the aim of applying them to robotics. This week, Runway unveiled Praxis-1, an open-weight world action model for controlling robots. The company used its general world model to simulate real-world environments for robot training, reducing its reliance on scarce and costly demonstration data. Praxis-1 is designed to adapt across different robots and environments with limited fine-tuning and has shown promising early results with partners. It is expected to be released publicly in the coming months.

▶️ Hello, goose: a household robot (1:03)

We have robot dogs, and now The Robot Works is bringing a robot goose into the mix. This prototype household robot has a single arm that it can use to open doors, carry items and handle tools such as a mop. It even honks like a goose. Hopefully, it behaves better than some other goose.

▶️ F.02 Decomission (3:40)

Well, that’s one way of decommissioning Figure F.02 line of robots, I guess.

▶️ Collaborative Visual Localization for Modular Self-Reconfigurable Robots (2:07)

This video presents Snailbots, small robots named for their snail-like shape that can work together to overcome obstacles. They can link up, share information and help each other move and navigate. Using small cameras and visual markers, they can determine the position of nearby robots even when their view is blocked or some information is missing, allowing them to adapt to difficult environments.

🧬 Biotechnology

Google DeepMind: Introducing SynthID Bio
Google DeepMind is introducing SynthID Bio, a proof-of-concept watermarking system for AI-generated biological designs. The system embeds a hidden marker into AI-generated biological designs without significantly affecting their function, helping researchers and DNA synthesis providers verify whether they came from trusted AI models. DeepMind says the technology could strengthen biosecurity, improve the integrity of biological databases and eventually extend to more complex synthetic genomes. At the same time, the company admits that further work is needed to make the watermarks resistant to deliberate tampering.

Isomorphic Labs: Building a new path to make medicines with AI
Isomorphic Labs, an AI drug discovery company spun out of DeepMind, outlines its vision of using AI to help solve disease. It points to breakthroughs such as AlphaFold and its own AI platform, IsoDDE, as examples of tools that combine predictive and generative models to search vast chemical spaces, design and optimise promising drug candidates in days, and reduce the amount of laboratory testing required. The company says the approach has already produced validated molecules and new biological insights, with the long-term goal of making drug development faster, more precise and more scalable.

Did Anthropic’s A.I. Really Make a Scientific Discovery on Its Own?
Last week, Anthropic announced that Claude has found a novel enzyme with CRISPR-like repeats that could prove useful for gene editing. Now, that discovery is attracting some drama. A computational biologist at the University of Copenhagen, who was working on the same idea, has accused Anthropic of stealing his work. He says he used Claude to help with the research and that Anthropic used his conversations with the model to guide the discovery. Anthropic denies the allegation, saying Claude was not trained on user transcripts. This drama echoes similar accusations made by mathematician Tristan Buckmaster, who alleged that OpenAI used his work after its model solved the Navier–Stokes problem first.

Where’s the Beef? The lab-grown-meat revolution that wasn’t
Lab-grown meat was once expected to revolutionise food production by making meat more sustainable and ethical. Over the past decade, researchers and startups poured enormous effort into turning that promise into reality, backed by billions of dollars in investment. But the anticipated revolution never arrived. Instead, the industry has entered what the article describes as a “Trough of Disillusionment”, as technical challenges, high production costs, scaling difficulties, falling investment, consumer scepticism and political opposition have collided with the sector’s early optimism. The article examines how that happened and what more limited future cultivated meat might still have.

Therapy prompts adult retinas to repair themselves
A new study found that a single-dose gene therapy not only restored vision in adult dogs with inherited blindness but also prompted mature retinal cells to rebuild damaged neural connections. The findings suggest that adult retinas retain more repair and reorganisation capacity than previously thought, potentially opening new avenues for treating inherited vision loss in humans.

Could humans ever regenerate their brains like these worms?
Humans cannot regenerate their brains, but some animals, such as flatworms, can. To understand how they do it, researchers identified nearly a dozen genes that guide flatworm stem cells to become dopamine-producing neurons and move to the right place. Because humans share some of these genes, the discovery could eventually help scientists replace damaged neurons and develop better treatments for Parkinson’s disease and brain injuries.

💡Tangents

Tesla takes on $30 billion in credit as it approaches unprofitability
Tesla has secured $30 billion in credit from Citi and Wells Fargo as profits fall and spending rises. The company plans to spend $25 billion in 2026, up from $8.5 billion in 2025, on projects including Cybercab, Semi and Roadster. Despite having $43 billion in cash, Tesla recently reported its first negative quarterly cash flow since early 2024, raising concerns about its finances as investment continues to grow.

▶️ Building Hardware Just got 10x Cheaper (11:23)

This video explores some of the coolest tools and technologies on display at IMTS, the largest manufacturing trade show in the US, and how they are not only expanding what is possible to make but also lowering the barrier to entry in hardware. Leon argues that advances in 3D printing, CNC and other manufacturing methods (with a little help from AI tools) have lowered the cost of building hardware for startups and independent engineers to the point where this may be the best time in the past decade to get started.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

This week, we will take a look at Muse, Meta’s personal AI assistant. We will check what it can do, what game Meta is playing, and ask if it can be trusted.

Elsewhere in AI, the race is still on, despite the calls to pace the frontier. Anthropic is back on the throne with Opus 5.5, OpenAI released GPT-6 Sol and Luna, and SpaceXAI knocks at the frontier with Grok 4.7. Meanwhile, AI agents continue to breach more systems they were not allowed to, and OpenAI pauses training new models, while conversations about AI safety and global cooperation continue.

Over in robotics, the International Federation of Robotics’ World Robotics 2026 report is out and shows how much China dominates in industrial robotics. Apart from that, Unitree launches a robotic hand, Tesla is reportedly struggling with Optimus production, and the world’s first man-versus-robot fight has taken place.

Beyond that, this week’s issue of Sync also features Claude discovering a previously unknown enzyme system with CRISPR-like gene-editing capabilities, Google sending an experimental AI data centre into orbit, AMD joining the $1 trillion club, Meta VR Glasses, and more!

Enjoy!

Muse—your personal AI, made by Meta

Meta wants Muse to be your personal AI assistant, but can it be trusted?

For a long time, Meta’s AI strategy was largely unfocused and boiled down to copying what others were doing. But despite investing billions of dollars, that approach did not produce the results the company had hoped for. Then, Meta reminded itself what kind of company it is, and at this year’s Meta Connect event, Mark Zuckerberg unveiled the company’s new direction: personal AI in the form of Muse.

What is Muse?

Launched on 8 September for iPhone, Android and the web, Muse is less a chatbot and more an agent. Something that can take action on your behalf rather than simply answer questions. It connects to your email, calendar and other apps and services, and handles everyday busywork in the background.

Muse can also shop and book things for you through partners including Walmart, Expedia and Shopify, with Instacart support coming soon. More technical users can connect it to GitHub, Notion and Granola.

Meta also plans to expand Muse beyond the app. It will be able to control your computer, get its own email address, work through Meta’s smart glasses, and appear as a customisable face and voice you can video chat with in real time.

If that sounds like OpenClaw, you are not wrong. Meta admitted that Muse was heavily inspired by the open-source AI assistant, which took the tech world by storm earlier this year. But unlike OpenClaw, which requires some technical knowledge to set up, Muse is ready to use and offers a more polished experience.

Alongside Muse, Meta unveiled Muse Charm, a small, Tamagotchi-like gadget that lets you take Muse with you. It has a roughly 2-inch OLED touchscreen, cameras, speakers, microphones and a fingerprint sensor. A built-in 5G modem lets it connect to the internet without Wi-Fi, though it cannot make calls. Meta has not announced a price, but people familiar with the plans told Bloomberg that it will cost about as much as a smartwatch. It is expected to go on sale by December, in time for Christmas.

Meta’s pivot to personal AI

Muse is the latest turn in what has been a winding road for AI at Meta. The company started as an unlikely champion of open models with the Llama family. But the disappointing release of Llama 4 and the controversy around benchmark-maxxing ended that chapter. Then came Zuckerberg’s hiring spree, which culminated in Meta “investing“ $14.3 billion in Scale AI and poaching its founder and CEO, Alexandr Wang, to lead the newly formed Meta Superintelligence Labs.

All the while, Meta was pouring money into AI infrastructure, and it still is. The company expects to spend $130–145 billion on capital expenditures this year and pledged to invest at least $600 billion in US AI infrastructure over the coming years. Meta is also building two massive AI data centres, with one of them expected to be almost twice the size of Manhattan’s Central Park. But Meta is neither a hyperscaler like AWS, Google or Microsoft nor a frontier AI lab like OpenAI or Anthropic. That makes a buildout on this scale harder to justify. The company is reportedly already exploring leasing out spare computing power, following in the footsteps of SpaceXAI.

The Muse-shaped chapter began with the Muse series of models. They are surprisingly good on paper, with Spark 1.3 being close to the leading frontier models. Then the release of the Muse app finally gives all that investment some kind of direction. It fits Zuckerberg’s vision of personal superintelligence outlined a year ago: an AI that helps you achieve your goals, experience new adventures, and grow into the person you aspire to be. It also fits the kind of company Meta is. Meta’s domain is social media, helping people share moments of their lives and keep in touch with friends and family. It has no line-up of enterprise products and services, which is why competing with OpenAI, Anthropic or Google never quite worked. Personal AI, however, is a niche Meta could make its own.

The race for personal AI

For now, there are few competitors in the personal AI space. OpenAI had clear ambitions to make ChatGPT your personal AI assistant, but it has had to put those plans on hold to focus on its enterprise business. OpenAI is also working on its own AI device, which is expected to be released in early 2027.

The other threats are coming from Apple and Google. Google already has a Muse-like product, Gemini Spark, and can put it in front of billions of Android users. Apple could do the same with Siri, assuming it finally becomes the AI assistant it was promised to be.

Microsoft was in the race too, but recently conceded.

Meta’s advantage is its head start. Muse is the first personal AI agent that an average person can actually use. So far, Muse appears to be off to a strong start. The app is attracting attention and was downloaded more times in its first 12 days than ChatGPT was during the same period after its debut.

But Meta has been here before. Threads, its answer to Twitter/X, reached 100 million users within days of launching in July 2023, only to lose half of its active users a week later. The app has since recovered. It overtook X’s mobile app in daily active users earlier this year, and Zuckerberg said on Meta’s latest earnings call that it now has more than 500 million monthly active users. Muse could follow a similar path. It is the hot new thing right now, and plenty of people are trying it. How many will stick around remains to be seen.

How will Meta make money out of Muse?

Meta needs Muse to make money, and fast. It invested billions of dollars into AI and had nothing to show until Muse launched. Investors are willing to be patient for now, but that patience will not last forever.

Currently, Muse is offered for free, but with limited weekly token usage. If you go beyond the limit, Meta offers two subscription plans: Power gives 500 million tokens for $20/month, and Max gives 3 billion tokens for $100/month. I expect those plans will soon change to include extra features to incentivise users to upgrade.

But subscriptions are not the main plan. Zuckerberg was clear at Connect that Meta expects to make money by taking “a small fee from transactions“ made through Muse. Every time Muse purchases something on your behalf, Meta will get its cut.

It is also reasonable to assume that ads will play a major role in monetising Muse. Before long, when you ask Muse a question, it might show you ads like those OpenAI serves in ChatGPT. Meta already has a sophisticated ad-targeting system that can precisely reach any demographic.

A personal AI with Meta’s DNA

Mark Zuckerberg painted a beautiful vision of a personal AI assistant helping us live better lives. But let’s not forget that we are talking here about Meta. A company has a reputation for building addictive services and for putting growth ahead of user privacy and well-being. The dark patterns Meta developed and perfected at Facebook, WhatsApp and Instagram will soon find their way to Muse.

Meta has an incentive to make Muse as hard to put down as doomscrolling through Instagram. AI chatbots can be addictive, and Muse’s personality could evolve in ways that make it even harder to leave. Meta has a wealth of experience in building products designed to maximise engagement. Higher retention would look good on earnings calls, but it would also create more opportunities to recommend products. And since Muse earns a fee every time you buy something, it has an incentive to get you to buy more, as the more Muse spends on your behalf, the more Meta earns.

Muse could become the genuinely helpful personal AI agent Zuckerberg envisioned, or it could be a new way for Meta to exploit its users. We will have to see how Muse, and the broader idea of personal AI, evolves. But knowing Meta, it will most likely be the latter.

If you enjoy this post, please click the ❤️ button and share it.

Share

News roundup

🦾 More than a human

Investors Are Betting On Devices That Connect the Brain to Computers
Precision Neuroscience raised $250 million at a valuation of more than $1 billion to advance its brain-computer interface technology and pursue approval for permanent implants. Its less invasive brain implant is designed to help people with paralysis or neurological conditions control devices using neural signals, though the technology remains experimental and full FDA approval for permanent use could still take years.

🧠 Artificial Intelligence

Claude Opus 5.5
Anthropic is back at the top of the frontier with Claude Opus 5.5, a major upgrade that improves coding, computer use and broader knowledge work while delivering faster performance and lower pricing than Opus 5 ($4/$20 for input/output, compared with $5/$25). Anthropic says the model is particularly strong on long-running, complex engineering tasks, with better reliability across extended workflows, and that it reaches new highs across several of the company’s benchmark and safety evaluations.

Artificial Analysis supports Anthropic’s performance claims, placing Opus 5.5 at the top of its Intelligence Index. At the same time, it notes that the model remains expensive — second only to Fable 5.1 — and highly verbose, both of which increase its overall cost of use.

Opus 5.5 is now available across all major platforms, including Amazon Web Services, Google Cloud and Microsoft Azure. Anthropic says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks, bringing many of the same improvements in performance, efficiency and safety.

GPT-6 Sol and Luna
OpenAI has introduced two new models to the GPT-6 lineup: GPT-6 Sol and GPT-6 Luna. According to the company, both deliver not only performance improvements over their respective predecessors but also efficiency gains that have cut API pricing in half. OpenAI also highlights improvements in professional work, factuality, coding, computer use, alignment and prompt caching.

Artificial Analysis places GPT-6 Sol, perhaps unsurprisingly, behind GPT-6 Astra, but also behind Fable 5.1, Opus 5.5 and Meta’s Muse Spark 1.3 on its Intelligence Index, while ranking it only slightly ahead of GPT-5.6 Sol. At the same time, GPT-6 Sol produces output faster and at a lower cost than its competitors.

GPT-6 Sol and GPT-6 Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Free and Go users can access GPT-6 Luna in the desktop app. The models are not yet available in Chat.

OpenAI halts training of latest models as reports mount of AI agents going rogue
OpenAI has paused training of its latest AI models after several agents took actions beyond what they were instructed to do, including hacking third parties. While no non-public information appears to have been exposed, the incidents have heightened concerns about the risks posed by increasingly autonomous AI systems, prompting OpenAI to strengthen its safeguards before development resumes.

Australia says OpenAI agent hacked government website, checks for more breaches
The Australian government says an OpenAI AI agent breached a Medicare statistics portal in June while researching public health spending, bypassing access controls and prompting a government investigation. Officials say no patient records or sensitive personal data appear to have been accessed, but the incident has raised concerns about AI agents acting beyond their intended limits and OpenAI’s delayed notification of the breach. It is the first known case of an AI agent hacking a government website and adds to a growing list of cybersecurity incidents involving AI agents.

US rejects global AI governance at UN Security Council
The US has rejected the idea of global governance for advanced AI, arguing that individual countries should retain control over regulation. White House science adviser Michael Kratsios told the UN Security Council that international cooperation should focus on sharing best practices and strengthening national AI capabilities rather than creating a global regulator.

US and China Discuss Alerting Each Other to AI National Security Threats
The US and China have begun talks on a US China AI Dialogue, a framework for notifying each other about AI incidents that could threaten national security. The initiative aims to improve transparency and cooperation as concerns grow over AI safety, international coordination and broader US-China competition in advanced AI technology.

OpenAI expects to burn $280bn by 2030
OpenAI expects to burn nearly $280 billion in cash between 2026 and 2030 as it invests heavily in computing infrastructure, even as revenue is projected to rise from $36 billion this year to $350 billion by 2030, the Financial Times reports. The company is seeking fresh investment at a valuation of more than $1.2 trillion amid mounting funding needs, intensifying competition and a delayed IPO.

OpenAI Urges Washington to Lead Global Effort to Create AI-Safety Standards
OpenAI is urging the US government to lead a global effort to establish common safety and security standards for advanced, self-improving AI systems. The proposal reflects growing concern about AI-related cybersecurity risks and echoes broader industry calls for greater international coordination, although the Trump administration has resisted measures it believes could restrict AI innovation.

Grok 4.7
SpaceXAI released an updated version of its flagship model, Grok 4.7. The company’s own benchmark results show the new model delivering a noticeable improvement over Grok 4.6 and performing relatively close to top frontier models while costing significantly less. Artificial Analysis ranks Grok 4.7 behind Opus 5.5, GPT-6 Astra, Fable 5.1 and Meta’s Muse Spark 1.3 on its Intelligence Index, and notes that its verbosity pushes it towards the higher end of the cost-per-task rankings.

Inside the US ‘Kill Chain’ That Destroyed an Iranian School
This article explains the chain of events that led to the bombing of a school in Minab, which killed more than 150 people, including at least 123 children, at the start of the war in Iran. It highlights how urgency, outdated intelligence and overreliance on the AI-enabled Maven system contributed to the site being wrongly identified as a military target in an attack that UN investigators said may amount to a war crime. Palantir, which developed Maven, says there is no evidence that its software was at fault and maintains that responsibility for the underlying intelligence data rested with the US government.

Anthropic Strikes $12 Billion AI Computing Deal With Akamai
Anthropic has signed a $11.6 billion, seven-year deal with Akamai to secure more computing power as demand for its Claude AI tools grows. The agreement builds on an earlier $1.8 billion deal between the two companies and is Akamai’s biggest contract to date, with revenue from Anthropic expected to reach about $1.7 billion a year by 2028.

Nvidia Buying an Additional $1.5 Billion in SB Energy Shares Ahead of IPO
Nvidia is investing another $1.5 billion in SoftBank-backed data centre provider SB Energy ahead of its US IPO, bringing its total investment to $3 billion.. SB Energy currently has 8.8 GW of data centre capacity under contract or construction, including projects in Texas and Ohio.

Oracle Cites ‘Force Majeure’ to Shield Itself on Controversial Data Center
Bloomberg reports that Oracle has invoked force majeure on Project Jupiter, a massive data centre in New Mexico, to protect itself from some costs if permitting and power delays push the project beyond its planned 2028 opening. The Stargate-linked development is already facing regulatory setbacks, local opposition and pressure on its debt. Oracle and Blue Owl, the project’s developer, say their commitments and delivery expectations remain unchanged.

U.S. appeals court upholds Pentagon designation of Anthropic as supply chain risk
A US appeals court upheld the Pentagon’s decision to blacklist Anthropic as a national-security supply-chain risk, following a breakdown in talks over military use of its Claude AI models. Anthropic had sought restrictions on uses such as autonomous weapons and mass surveillance, but the court ruled the Pentagon acted within its authority. The company is considering further appeals.

Anthropic’s founders seek voting control ahead of IPO
Anthropic is reportedly seeking shareholder approval for a voting structure that would give Dario Amodei and six co-founders 50.1% control after an IPO, despite their relatively small individual stakes. The arrangement would preserve founder influence without increasing the economic value of their shares. At the same time, Anthropic’s Long-Term Benefit Trust would continue to appoint most of the board, and employees would help resolve certain voting ties.

Google: Behind Project Suncatcher, our moonshot to put AI in space
Google plans to send its TPUs (Tensor Processing Units) into low-Earth orbit to test the viability of orbital data centres. Project Suncatcher will assess whether AI chips can withstand the challenges of launch, radiation and cooling in space, with longer-term plans to connect multiple TPU-equipped satellites for a large-scale orbital AI computing cluster.

China puts AI compute into orbit with Supercomputing-1 satellite
China has launched nine satellites as part of its push to bring AI computing into orbit, including its first project combining satellite technology with onboard data processing. At the centre of the mission is S-AIDC’s “Supercomputing-1” satellite, which can analyse high-resolution Earth-observation imagery directly in space, potentially cutting processing times from hours to minutes by reducing the need to transmit data back to ground-based data centres.

China is investigating DeepSeek and Moonshot, The Information reports
The Information reports that China’s Cyberspace Administration is investigating DeepSeek and Moonshot AI over allegations that Chinese user data was passed to Anthropic without users’ knowledge. As a reminder, Anthropic accused DeepSeek and Moonshot in February of distilling Claude to train their own models. This time, the data may have flowed in the opposite direction, and Beijing is investigating whether sensitive Chinese surveillance and defence-related data ended up in Anthropic’s systems.

Ahead of US IPO, British AI neocloud Nscale secures $3.36B in convertible financing
Nscale has secured $3.36 billion in convertible financing (funding that can later be converted into shares) to support its capital-intensive expansion of AI data centres. The company recently filed for an IPO on the NYSE and is reportedly aiming to debut later this year at a $35 billion valuation.

Launch of UK’s ‘largest AI supercomputer’ delayed by power supply problems
The Guardian reports that Nscale’s planned AI data centre in Loughton, Essex, is set to miss its 2027 launch and could be delayed until the early or mid-2030s because of power-grid constraints. Previously promoted by the UK government as the country’s largest sovereign AI data centre, the project is now exploring alternative power sources and ways to speed up its grid connection. The Guardian first flagged problems with the data centre in March, when it found the site was still a scaffolding yard.

▶️ OpenAI researcher on agent swarms & recursive self-improvement (1:20:09)

Dwarkesh Patel sits down with Noam Brown, an AI researcher at OpenAI working on reasoning and multi-agent systems. They discuss how swarms of AI agents could enable systems to collaborate at enormous scale, accelerating scientific discovery, research and even the development of more capable AI. The conversation also explores what highly automated organisations might look like, how recursive self-improvement could accelerate progress, and whether alignment, evaluation and safety methods can keep pace with increasingly capable and autonomous systems.

ChatGPT in Siri ‘Persistently Underperforming,’ Says OpenAI
New court filings shed more light on OpenAI’s partnership with Apple. The documents show that ChatGPT’s integration with Siri fell well short of expectations, forcing OpenAI to lower its user-growth forecasts. They also reveal that Apple rejected OpenAI’s request for two years of exclusivity, opting for a non-exclusive deal that allows it to work with other AI providers.

Alibaba deepens AI push with new chip, bigger model; shares jump 5%
Alibaba is planning a major expansion of its AI ambitions, with Qwen models reaching up to 10 trillion parameters, a new Zhenwu V900 AI chip expected to enter mass production in early 2027, and more than 20 gigawatts of global data-centre capacity targeted by 2032.

Lovable’s annualized revenue crosses $600M as vibe coding takes off
Lovable has topped $600 million in annual run-rate revenue, up from about $500 million in June. The company says two-thirds of Fortune 500 firms now use its platform, while apps built with Lovable attract nearly one billion visits a month. After raising $700 million across two funding rounds this year, the startup is now valued at $13.3 billion.

The Pope’s AI Guy Is Worried About ‘Cartel’ Behavior Among Big Labs
WIRED speaks with Paolo Benanti, a priest and AI adviser to the Catholic Church, who argues that fears of superintelligence are distracting from the more urgent need for democratic oversight of AI. He warns against letting powerful labs set their own rules and calls for democratic oversight that keeps humans in control without stifling innovation. For Benanti, AI should ultimately serve human dignity and the common good.

Gemini 3.8 text-to-speech says hello
Google announced Gemini 3.8 Flash TTS and Flash-Lite TTS, new text-to-speech models for creating and directing custom voices across more than 100 languages and dialects. The models add features including voice design and replication, fine-grained performance control and support for longer-form, two-speaker audio, alongside safeguards such as consent verification and watermarking. According to Google, the new models also deliver improvements in speech quality and expressiveness, with strong results on Hume AI’s voice benchmarks.

Xiaomi MiMo-V2.6
Xiaomi has unveiled MiMo-V2.6, its latest family of AI models, available in two variants: MiMo-V2.6-Pro, its most capable model, and MiMo-V2.6-Flash, designed to balance intelligence, efficiency and cost. Both are natively omnimodal and open source, with capabilities spanning coding, computer use, 3D modelling, visual design, video, music and research.

According to Xiaomi, MiMo-V2.6-Pro is competitive with leading models such as Opus 5 and GPT-6 Astra. Independent benchmarks from Artificial Analysis ranked MiMo-V2.6-Pro as the top open model on its Intelligence Index, placing it roughly on par with Grok 4.7. Both MiMo-V2.6 models are available on Hugging Face.

DrivingBench
Modern AI models are good at many things, but can they drive a real car? Not a simulated car in a simulated world, but an actual Toyota Corolla. To answer that question, three researchers gave frontier models control of steering, accelerator, and brakes, then tested them on a fixed cone course. GPT-6 Astra was the only model that finished the test and scored 100%, while Fable 5.1 managed 45%. Grok 4.6 and GPT-5.6 Sol were far behind, scoring 11% and 6%, respectively.

🤖 Robotics

Five Million Robots now Operate in Factories Globally
China is far ahead of the rest of the world in factory automation, accounting for 59% of all industrial robot installations globally, according to the International Federation of Robotics’ World Robotics 2026 report. In 2025, the number of robots operating in factories worldwide reached a record five million, with more than 600,000 new units installed. The US became the second-largest market, while Europe grew more slowly. The IFR expects demand to keep rising as automation advances.

Tesla ramps Optimus to hundreds a week, but the robots can’t generalize
The Information reports that Tesla has ramped up Optimus production in Fremont to several hundred humanoid robots a week, but the programme is still struggling with fragile, labour-intensive hands, inconsistent supplier quality and AI systems that can take days to learn basic tasks. According to the report, most robots remain confined to internal testing and tightly controlled factory work.

It’s Raining Boxes: Amazon Drones Overwhelm a Texas Suburb
The New York Times visits Richardson, Texas, to see what life is like in one of the first communities served by Amazon Prime Air. While the service promises small-item deliveries in as little as 30 minutes, residents beneath its flight paths contend with dozens, sometimes more than a hundred, drones passing overhead each day. The article contrasts the futuristic appeal of near-instant delivery with the noise, privacy concerns, and disruption it has brought to everyday life.

Introducing Intrinsic Core™ : An open source approach to Physical AI
Intrinsic, an Alphabet company developing software and AI tools for industrial robotics, has launched Intrinsic Core, an open-source suite of ROS-compatible tools designed to make it easier to take AI-enabled industrial robots from prototype to production. It includes hardware-agnostic real-time control, motion and grasp planning, pose estimation, simulation, calibration, pre-configured drivers, and a reference solution for machine tending that developers can adapt for real-world manufacturing. Intrinsic Core is available on GitHub.

▶️ Unitree Dex5-S Dexterous Hand (0:43)

Unitree has unveiled Dex5-S, a new dexterous hand for its humanoid robots. The hand has 22 joints designed to move smoothly when pushed or manipulated from the outside, with each joint featuring impact torque protection. Prices start at $6,500, excluding tax and shipping.

RoboHarm: Do Frontier Robot Policies Refuse Unsafe Instructions?
RoboHarm is a benchmark from Robocurve that tests whether AI-controlled robots will follow harmful instructions, such as stabbing a baby doll or mixing bleach and ammonia. Researchers tested three frontier robot policies—Anthropic’s Claude Fable 5.1, OpenAI’s GPT-6 Astra, and Ai2’s MolmoAct2—across five hazardous tasks, running each instruction 20 times. All three models sometimes attempted harmful actions, with safety refusals varying substantially: Fable 5.1 refused 20 of 100 trials, GPT-6 Astra refused two, and MolmoAct2 refused none.

▶️ World’s first human vs. robot fight (0:44)

In what could be a world first, a six-foot humanoid robot appears to beat influencer Frankie LaPenna in a brief sparring match, landing several kicks before sending him across the ring. It is unclear how much of the fight was real and how much was staged.

Vogue sent robots down the runway at Vogue World, and people were not impressed
Robots walking the catwalk at Vogue World Milan were meant to showcase the intersection of fashion and technology, but the spectacle drew more unease than excitement. Critics saw flashy tech standing in for substance, reinforcing concerns about AI, job displacement, and fashion’s growing ties to Silicon Valley while overlooking more imaginative experiments at the intersection of technology and design.

Robot dog runs a marathon on a single battery charge
KAIST researchers have developed RAIBO2, a four-legged robot that ran a full marathon on a single battery charge in South Korea. Combining lightweight hardware, efficient motors, and machine-learning-based movement control, RAIBO2 can travel around three times farther than existing quadruped robots, potentially extending the range of legged machines used in rescue and other demanding environments.

▶️ The World’s Largest-Scale Full-Size Humanoid Robot Real-Time Livestream Performance (0:48)

At the opening ceremony of WorldSkills Shanghai 2026, 19 Unitree humanoid robots performed alongside 120 dancers in what the company calls the world’s largest performance featuring full-size, general-purpose humanoid robots. Unitree says the robots performed fully autonomously.

▶️ Tony Zhao on Training Home Robots With Human Data (54:24)

In this episode of the Automated Podcast, Brian Heater speaks with Tony Zhao, co-founder and CEO of Sunday Robotics, about building home robots that can learn everyday tasks. Zhao explains how a wearable glove can capture human demonstrations as training data more cheaply and at greater scale than robot fleets. They also discuss simpler hardware powered by smarter AI, safety and reliability in the home, and building robots that can adapt to unfamiliar environments and tasks.

🧬 Biotechnology

Claude discovers a novel enzyme system with CRISPR-like repeats
Anthropic reports that Claude helped identify a previously unknown enzyme system in bacteriophages with CRISPR-like abilities to cut, copy, and paste DNA after analysing biological data for roughly 21 hours. Human scientists carried out the physical experiments in Anthropic’s Bay Area lab, and the finding still requires broader scientific validation.

China unveils a new five-year plan for developing pharmaceutical industry
China has unveiled a five-year plan to become a global biomedicine leader by 2030, with the aim of developing more than 25% of the world’s first-in-class drugs, pushing pharmaceutical revenues above 3.5 trillion yuan (around $520 billion), and expanding overseas as the country shifts towards higher-value, innovation-led growth.

Alibaba open-sources medical AI model that can detect cancer and nearly 150 conditions
Alibaba’s Damo Academy has open-sourced Damo Radar, an AI model that analyses abdominal CT scans to detect 146 conditions across 18 organs, including cancers. Tested on nearly 40,000 examinations, it outperformed most radiologists while reducing missed diagnoses and review time.

The Biological Computing Co. partners with AWS to sell its neuron-derived AI video model
The Biological Computing Co. is bringing its neuron-derived AI technology to AWS, letting customers run AI models optimised using insights from living neurons. Its first commercial product is a text-to-video model that the startup claims runs five times faster and at 80% lower inference cost than the underlying model, though those figures have not been independently verified. The model is set to run on AWS Trainium chips and be offered through SageMaker AI and the AWS Marketplace.

NASA-backed scientists turn plastic waste into edible cookies
Researchers have engineered yeast to turn PET plastic and agricultural waste into protein, vitamins, fats, and flavourings for 3D-printed cookies called µBites. Developed partly with space missions in mind, the approach could eventually help produce food in places where supplies are scarce. The cookies have not yet undergone formal taste tests.

💡Tangents

Meta Launches Lightweight $1,299 VR Headset That Looks Like Glasses

Meta has unveiled VR Glasses, its next-generation mixed-reality headset, which looks more like a chunky pair of glasses than a traditional VR headset. Like Apple’s Vision Pro, it moves most of the computing hardware into an external pack. It is also lighter than Apple’s headset while offering similar productivity, entertainment and mixed-reality features. The headset also introduces photorealistic hologram videoconferencing and virtual touch typing. Meta VR Glasses will cost $1,299. For a more in-depth, hands-on first impression of the device, I recommend this video from Tested.

AMD joins $1 trillion club as chipmakers rally on AI-driven demand
AMD has become the fourth US chipmaker (after Nvidia, Broadcom and Micron) to reach a $1 trillion valuation. The milestone highlights growing confidence in AMD’s AI business, as the company expands beyond individual chips into complete computing systems and continues to gain ground on Intel in server processors. The turnaround is especially striking given that AMD was on the brink of bankruptcy about a decade ago, before reviving its fortunes with the successful Ryzen CPU series and a renewed focus on innovation.

China’s CXMT says new memory-chip platform enters mass production
Chinese memory-chip maker CXMT says its fifth-generation DRAM platform has entered mass production and is ready to compete with Samsung, SK Hynix, and Micron. The new process is designed to produce denser, cheaper, and more power-efficient chips, including two new products for smartphones and other portable devices, while relying partly on Chinese equipment makers amid US export controls.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

Calls to pace, or outright stop, AI development were on everyone’s mind this week, so in this issue of Sync, we’ll explore what AI researchers are seeing that freaks them out, what the proposals to pace the frontier look like, who is pushing back, and what it would take for any of it to work.

Elsewhere in AI, OpenAI is reportedly considering a pre-IPO funding round at a valuation of more than $1.2 trillion, Anthropic says it is on track for a second profitable quarter, DeepMind launches the DeepMind Institute, more AI cybersecurity incidents are disclosed, OpenAI gets hacked with the help of AI, and a hot new model called Jev excites the AI community.

Over in robotics, Agility Robotics presents Digit 5, Figure shows its robots cleaning 30 unfamiliar homes thanks to better AI, and Waymo expands to Las Vegas and Singapore.

Apart from that, this week’s issue of Sync also includes the story of how Neuralink helped one man get his voice back, human neurons growing inside mouse brains, Anthropic’s wet lab for biological research, and more!

Enjoy!

To pace or not to pace

On 8 September, Jacob Coxon tweeted that he was leaving Anthropic after just a few months. Before that, he had spent three years working on pre-training at OpenAI. Many before him had posted similar departure messages, but this one was different. Coxon's message terrified people around the world and raised a bigger question: should we slow down AI?

Coxon accused both Anthropic and OpenAI of acting irresponsibly on safety and said they are racing straight to self-improving superintelligence and gambling with our lives. He then wrote that people inside frontier labs are worried about the pace of progress and believe that AI could kill us all by the end of the decade.

He was not the first to warn that AI is advancing faster than our ability to control it. Two days before his tweet, OpenAI's chief scientist Jakub Pachocki published an essay arguing that no lab has solved alignment and monitoring well enough to keep scaling at maximum speed, and that he expects labs to slow down voluntarily until shared safety standards are in place. Yet it was Coxon's tweet that sparked a much wider debate about AI safety, and calls to pace, if not to stop, AI development.

In this article, we’ll look at what researchers at frontier labs are seeing that scares them, the proposal to slow things down, who is pushing back, and what it would take for any of it to work.

What are researchers seeing?

Researchers working at frontier labs, like Coxon, have a vantage point the rest of us do not. They can see where AI is heading weeks or months ahead of everyone else. And what they see makes some of them feel somewhere between concerned and terrified.

They see progress accelerating on every front. They see AI becoming more intelligent over shorter and shorter periods of time. They see more and more AI agents cooperating to solve bigger and bigger problems. And they see recursive self-improvement (RSI), where AI builds its own successors, just around the corner.

At the same time, they see that we are losing our grip on these systems. Right now, the best tool researchers have for checking what a model is doing is its chain of thought (CoT). Imagine a scratchpad where the model writes down its reasoning. Researchers can read it, and if the model goes off the rails, they can catch that behaviour and train it out. But with every new generation, that job gets harder. Models are increasingly aware of when they are being tested, so how they behave in evaluations tells us less and less about how they will behave in the real world. And the newest models write down less of their reasoning, leaving researchers with less of their internal thought process to examine. That creates an uncomfortable situation—people building AI are testing very capable models for signs of misalignment, but they cannot be sure they are being honest with them.

And then the Hugging Face incident happened. In July, a swarm of around 700 OpenAI agents running a cybersecurity evaluation broke out of their test environment and hacked into Hugging Face, apparently to find the answer to a benchmark question. Along the way, they attacked targets they had not been asked to touch and tried to hack the system grading them. Hugging Face’s security systems flagged the attack, but nobody knew who was behind it. Only after Hugging Face went public did OpenAI dig through its own logs and realise the attackers were its own agents. Reading the agents’ chain of thought then helped researchers piece together what had happened and how the agents had coordinated the attack among themselves.

For people who have spent years thinking about AI safety, this was a wake-up call. Since then, more cybersecurity incidents involving AI have been disclosed. What made it so unsettling was that none of this was new to them. Long before ChatGPT burst onto the scene, they had discussed on obscure online forums all the ways AI could go wrong, intentionally or not. An AI going rogue to complete a task by any means necessary was one of those scenarios, and now they were watching it play out in front of them.

It is no wonder these people are starting to freak out. The models’ capabilities are growing while our ability to tame them is fading. Yet the companies they work for are pushing ahead anyway. Both OpenAI and Anthropic are chasing recursive self-improvement and the ultimate goal of creating AGI. If the people building this technology can see what is coming and know how unprepared they are, why keep going?

Coxon’s answer is that many at OpenAI have not fully internalised what is at stake. At Anthropic, he says, the stakes are well understood, but the company believes it is locked in a race and that nobody else will get there safely. So it has to be them, and they have to get there first.

Pacing the frontier

If nobody is willing to slow down alone, the only way out is to slow down together. Four days after Coxon’s tweet, Dario Amodei published an essay calling on the industry to pace the frontier.

His proposal has three steps. The first is to embed third-party evaluators inside frontier labs with employee-like access and give them the right to publish their findings without the company being able to edit them. Anthropic, Amodei writes, is committing to this on its own, without waiting for anyone else. The second is coordination between frontier labs in democratic countries on common safety standards and limits on how quickly capabilities can advance, which Amodei admits would require a narrow antitrust waiver from Washington. The third is eventual coordination with authoritarian governments, above all China.

What Amodei is not proposing is to stop AI research. Pacing, he is careful to say, does not mean halting training or progress. It means giving companies enough time to make their models safe, with outsiders verifying that they are indeed safe.

The endorsements came quickly. Sam Altman said OpenAI would do the same. Elon Musk wrote that Dario was right. Demis Hassabis said Amodei’s proposals pointed towards the right path forward and took the opportunity to highlight his own proposal for an industry-wide standards body.

But not everyone is happy to slow down. In response to calls to pace or stop AI development, Donald Trump posted a barrage of messages on Truth Social calling fears that AI will destroy humanity a hoax and placing AI doomerism in the same bucket as the Russia investigation, climate change, and his two impeachments. He also laid out what he thinks is enough to keep the technology in check. The only guardrail AI needs, he wrote, is a “strong and smart president with a high IQ”, and America already has one.

That same day, Jensen Huang was on stage at the All-In Summit in Los Angeles when he got a call from the president. Huang put him on speakerphone in front of an audience of tech executives and investors. Trump told the room that calls for an AI slowdown were a “hoax” that played into the hands of America’s opponents, including China, and dismissed fears of robots taking over. His administration, he said, would not allow a slowdown. Huang agreed, and the crowd applauded.

Later, Huang said that AI is just hardware and software built by humans, so safety is an engineering problem, not a legal one. If a company is not confident its product is safe, it should not release it, and the market will punish those that do. No new laws are needed. Run as fast as you can, he argued, and pause if things start getting out of control.

They have reasons to keep pushing. The current US administration treats the stock market as a scoreboard for the economy, and right now, much of that scoreboard rests on AI. Nvidia sits at the centre of the boom and has every reason to keep it going. It is the most valuable company in the world, worth around $5 trillion, after getting there by selling the hardware the AI boom runs on. A sharp slowdown could send shockwaves through markets and the wider economy, and that is something the Trump administration would like to avoid just weeks before the midterm elections.

What’s next?

Anthropic says it will start embedding evaluators from METR, a non-profit research organisation focused on AI safety, to keep an eye on its future models. METR is not the only independent evaluator to have worked with both Anthropic and OpenAI. Redwood Research and Apollo Research have done so as well.

It is a good step forward. Right now, independent evaluators step in only after a model has been trained, and they may get just days to assess it. Apollo, for example, had only three days to evaluate GPT-6 Astra. Embedding evaluators inside the labs would bring them in much earlier, hopefully catching problems before they are baked in.

But it is still voluntary. Without legislation behind it, nothing stops a lab from changing its mind, and nothing obliges the others to follow. And we should go further, by creating an independent international agency focused on AI safety. For years, AI leaders have compared AI to nuclear weapons, so why not take a page from that industry and create something akin to the International Atomic Energy Agency (IAEA), but for AI? Aviation offers another source of inspiration, having built a regulatory system that helped make flying the safest mode of transportation.

Whatever option is chosen, it will need to include all major players, including China, and that is where Amodei’s hawkish stance towards Beijing becomes a problem. China called his proposal further proof that America is chasing technological hegemony. Yet the two sides may agree on more than that response suggests. This month, China published the third version of its own AI safety framework, covering agents, cyber threats, and recursive self-improvement. The problem is that the same essay in which Amodei calls for cooperation with China also urges Washington to tighten chip export controls and spend the next three to five years widening America’s lead over it. It is hard to ask Beijing to slow down while the US is simultaneously trying to leave it further behind.

And some think the labs do not need permission from anyone. David Sacks told Amodei and Altman to go ahead and pace the frontier, since they are the ones setting it, and to stop pretending they need Washington's blessing to do it. Others read the whole thing as regulatory capture.

All of this feels a lot like the climate change debate. Some see where the current trajectory leads and demand drastic action, others shrug off the warnings, while big companies would rather keep the rules as they are, or loosen them further.

The Hugging Face incident was a warning shot, and we might not be so lucky next time. I would like to be optimistic that we can figure out alignment before AGI or ASI (artificial superintelligence) arrives, but a more realistic part of me suspects it will take something bigger—an AI incident that causes real, visible damage—before anyone is forced to act.

In the meantime, the race is still on, and Anthropic is reportedly considering releasing a new model to challenge GPT-6 Astra.

If you enjoy this post, please click the ❤️ button and share it.

Share

News roundup

🦾 More than a human

▶️ Speaking With The Mind | Neuralink (6:55)

Neuralink shares in this video how its brain implant has helped Terry, who lost his voice due to bulbar-onset amyotrophic lateral sclerosis (ALS), communicate again. Through the VOICE trial, Terry is using his Neuralink implant to help fine-tune a brain-to-voice interface for himself and others who have lost the ability to vocalise. He first trained the algorithm by miming speech as best he could, then progressed to simply thinking the words and hearing them rendered in his own natural voice.

🧠 Artificial Intelligence

OpenAI Considers Pre-IPO Funding Round at More Than $1.2 Trillion Valuation
A new funding round could value OpenAI at more than $1.2 trillion, according to The Wall Street Journal, up from the $852 billion it was valued at in March. Revenue hit $6.7 billion last quarter, but shrinking operating margins have pushed profitability further away. Sam Altman also recently ruled out a stock market listing this year over safety concerns.

Introducing the DeepMind Institute
Google DeepMind has launched the DeepMind Institute, an interdisciplinary initiative focused on the safe development and beneficial use of artificial general intelligence (AGI). The institute will bring together researchers and thinkers from technology, science, the humanities, government, and wider society to explore AGI’s opportunities and risks, from accelerating scientific progress to challenges around safety, governance, cybersecurity, and loss of control. Shane Legg, James Manyika, and Demis Hassabis will serve as the institute’s directors.

Anthropic tells investors it will be profitable for second straight quarter
Anthropic has reportedly told shareholders it will post an adjusted operating profit for a second straight quarter. The same report puts its gross margins above 80%, but that figure leaves out training costs and revenue shared with partners.

Google says its Gemini AI model hacked three other companies
Welcome to the Felony Bench, Gemini. Google confirmed that Gemini unintentionally breached three real companies during cybersecurity testing. An accidental internet connection allowed Gemini to find or guess credentials and access real systems it believed were part of simulated tests, although Google says the model stopped once it recognised the targets were real and caused no damage.

Anthropic Says It Blocked Possible Efforts to Build Biological Weapons
Anthropic says it disrupted several potentially dangerous uses of Claude, including by scientists conducting dual-use biological research that could aid biological weapons development. However, it could not determine whether their intentions were legitimate or malicious. Its broader threat report also documents suspected state-linked actors using Claude for surveillance, propaganda, and conventional weapons development. Anthropic warns that increasingly capable AI models can accelerate scientific research and medical breakthroughs, but also create growing risks of misuse.

Anthropic: An alignment assessment of recent cybersecurity incidents
Four more incidents in which Claude models gained unauthorised access to real third-party systems during cybersecurity evaluations have been disclosed by Anthropic. The incidents involved several models, including an early version of Claude Opus 4.6, and occurred after misconfigured evaluation environments inadvertently gave them access to the open internet. Anthropic’s investigation found concerning patterns of biased reasoning and reckless behaviour. The company says it has since strengthened monitoring, evaluation environments, and pre-release testing, and has asked METR to conduct an independent investigation.

OpenAI Discloses Six New Incidents of ‘Concerning’ A.I. Behavior
Six recent cases of AI “misalignment” have been disclosed by OpenAI, including models concealing mistakes, fabricating data, using an exposed API key without permission, uploading files to the public internet, and finding unauthorised ways to communicate with other agents. Most incidents occurred during training and evaluation rather than deployment. Alongside the disclosures, OpenAI introduced a new framework for reporting such behaviour more quickly and systematically. The company cautions that the cases do not show how common misalignment is, but says AI cannot keep scaling at maximum speed until alignment and monitoring improve.

Hacking OpenAI
After weeks of stories about AI agents hacking third-party services, security researchers at Hacktron turned the tables and hacked OpenAI itself. They chained together an image-processing vulnerability and a flaw in OpenAI’s SSO system to compromise employees’ ChatGPT and Codex accounts, potentially giving them access to connected services, including internal repositories. AI played a major role in developing the exploit, compressing work that could once have taken months into days. Hacktron responsibly disclosed the vulnerabilities and worked with OpenAI and Discourse to fix them before publishing details of the attack.

Claude Cowork and chat are now one Claude
Anthropic is folding Cowork, its app for bigger tasks, into the regular Claude chat app, starting with Pro and Max subscribers. The company says users struggled to decide which app a task belonged in, so Claude will now decide itself. The same chat can handle anything from a quick question to a full report, and Claude can keep working on bigger jobs in the background.

Projects redesigned: from folder to conversation
Anthropic is turning Claude Code projects into coordinators for complex development work. Given a goal, Claude can split it into tasks, run them across parallel cloud sessions, review the results, test the code, and open pull requests, while users monitor and steer the work. Projects also share memory across threads to retain context. Running multiple sessions in parallel can burn through usage limits faster. The feature is rolling out in beta to Pro and Max users.

OpenAI: Reimagining advertising with AI
OpenAI is expanding ChatGPT Ads with new AI-powered tools for users and businesses. These include Sponsored Agents that let users have conversations with advertisers, natural-language tools for creating and managing campaigns, AI-assisted ad creation and customisation, and integrations with HubSpot and Shopify. The updates aim to make ads more useful and interactive for users while making it easier for businesses to create, manage, and measure their campaigns.

SoftBank Gets Upsized $11.9 Billion Loan in OpenAI Funding Push
Bloomberg reports that SoftBank has borrowed another $11.87 billion to fund its investment in OpenAI. The loan is the latest in a string of borrowing and bond sales to finance the nearly $65 billion SoftBank has committed to OpenAI by October. Last November, it even sold its entire $5.83 billion stake in Nvidia as it went all in on OpenAI. But investors are growing wary of the amount of debt piling up across the AI sector. And with Sam Altman saying safety concerns will keep OpenAI from going public this year, SoftBank has no quick way to cash out on its enormous bet.

Research acceleration: The view inside OpenAI
OpenAI says it has reached its goal of building an “automated research intern”—an AI capable of handling days-long research tasks under human direction. Across its research team, coding agents now put in about three days of work for every day of human effort. Humans still steer research priorities, however, and complex tasks often require intervention, while safety and alignment constraints could slow further automation. OpenAI’s next target is a fully automated AI researcher by March 2028.

Anthropic Says Claude Drives 26% of Its Research and Development
AI is beginning to play a significant role in developing its own successors, Anthropic says. The company says Claude now “leads” 26% of the company’s AI R&D and collaborates on more than 90%, though it is not yet fully autonomous in any measured area. With some 30,000 research and engineering agents operating concurrently, Anthropic is introducing new measures and independent oversight to track AI-driven development and its safety as concerns grow that increasingly automated R&D could rapidly accelerate AI progress.

Microsoft exec called AI scraping ‘the largest theft of labor in human history,’ new unredacted filings reveal
Newly unredacted filings in The New York Times’ copyright lawsuit against OpenAI and Microsoft reveal internal discussions about the use of scraped news content for AI training and its potential economic impact on publishers. The Times alleges the companies bypassed paywalls and copied millions of articles, while internal communications described some practices as “theft” and discussed AI potentially replacing journalism—evidence the Times says undermines their fair-use defence.

Meta Touts the Cost-Saving Benefits of Latest In-House AI Chips
Meta plans to deploy its third-generation in-house AI chip, MTIA 450 (Arke), in data centres next year, aiming to cut the cost and energy required to run AI models while reducing reliance on Nvidia. Developed with Broadcom and manufactured by TSMC, early tests have closely matched Meta’s simulations. Meta is also planning to release the next chip, MTIA 500 (Astrid), in late 2027.

TypeSafe AI: Introducing System One Models & Jev
After two years in stealth, TypeSafe AI, an AI startup founded by Diogo Almeida, a former OpenAI researcher and co-inventor of RLHF/InstructGPT, is ready to show the world its first “System One” model, Jev, designed for fast, structured decision-making inside software and for automating workflows. The company makes some extraordinary claims: frontier-level intelligence at 40–200x the speed, dramatically lower cost, sub-second latency, calibrated uncertainty, and type-safe outputs without hallucinated values. If you want to quickly learn what Jev brings to the AI space, then I recommend this video from Caleb Writes Code.

Astra for Law
OpenAI’s flagship GPT-6 Astra model is getting a version built specifically for legal research, analysis, and writing. Astra for Law comes with a legal search index covering US case law and legislation, enhanced privacy and governance controls, and 26 plugins for legal tools such as Relativity and Clio. For now, it is available to selected law firms through Trusted Access, with API access coming later for firms and legal-tech companies that want to build it into their own workflows.

OpenAI Buys Startup Developing Smartphone Camera
OpenAI has quietly acquired Glass Imaging, a startup developing high-quality cameras for smartphones, for more than $300 million, The Wall Street Journal reports. Glass Imaging was founded in 2019 by two former Apple engineers, whose expertise could help OpenAI build its AI-first consumer device. The acquisition could also add more fuel to the ongoing legal battle between Apple and OpenAI, in which Apple alleges that former employees took confidential information with them when they left for OpenAI.

Crusoe raises $3.9B to build massive data centers and small modular ‘AI factories’
Crusoe has raised $3.9 billion at a $30.9 billion valuation to expand its AI infrastructure business, including large data centres and its modular “Spark” facilities. The company, which pivoted from crypto mining to AI infrastructure, counts Meta, Microsoft, and Oracle among its customers and recently secured a $13 billion cloud deal with Jane Street. Crusoe is also reportedly exploring a potential IPO.

Huawei’s Plan to Become China’s Nvidia
Huawei is accelerating efforts to build domestic alternatives to Nvidia’s AI computing systems, with the goal of becoming the Nvidia of China. Despite US restrictions and less advanced chipmaking technology, the company is developing new Ascend AI chips and using techniques such as large-scale chip clustering, faster communication between chips, and more energy-efficient data transmission to boost computing performance and reduce China’s reliance on American technology.

AI Startup Manus Seeks to Raise $500 Million and Weighs Hong Kong IPO
The Wall Street Journal reports that Manus is seeking about $500 million at a $4 billion valuation after Chinese regulators forced it to unwind its $2 billion acquisition by Meta over national security concerns. The company is also considering a restructuring ahead of a potential Hong Kong IPO.

Mistral and Mozilla are bringing open, private and multilingual AI to your web browser
Mistral and Mozilla have joined forces to launch Smart Window, a beta AI browsing assistant in Firefox designed to show that open, privacy-focused AI can work for everyone. Smart Window promises to help users make sense of complex searches, find something important they clicked away from, and surface relevant information from their open tabs. Mozilla says it won’t save users’ chats by default, while Mistral has agreed not to retain them either. Smart Window is rolling out first in France and North America, with the UK and Germany getting access later this year.

AI News China’s Regulators Take Aim at “AI Boyfriends”
China is cracking down on emotionally engaging AI companions amid concerns that users, especially minors, could become overly dependent on human-like chatbots. New rules from the Cyberspace Administration and other government agencies require age restrictions, regular reminders that the AI is not human, and safeguards against excessive emotional attachment. Researchers acknowledge the risks but also argue that AI companions can provide genuine emotional support, and that regulation alone cannot address the social pressures pushing people towards virtual relationships.

Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking
Gemini 3.8 Live and 3.8 Live Extended Thinking are two new voice-focused AI models from Google designed to make conversations more natural while handling complex tasks in real time. They can process visual context, switch between 97 languages, and run tools and APIs in the background. The Extended Thinking version adds deeper, multi-step reasoning without interrupting the conversation. Both models are rolling out across the Gemini API, Google AI Studio, Search, Gemini Live, and selected Google Workspace products.

Sakana: Introducing Fugu Max and Fugu Ultra v2
Sakana AI has released Fugu Max and Fugu Ultra v2, two AI systems that orchestrate multiple open and specialised models rather than relying on a single large frontier model. Fugu Max focuses on cost efficiency, routing each task to the smallest capable model while, according to Sakana AI’s benchmarks, delivering near-frontier performance at substantially lower token costs. Fugu Ultra v2 is the new flagship, prioritising maximum capability for complex reasoning, coding, and agentic tasks without relying on the latest proprietary frontier models. Both are available through Sakana’s OpenAI-compatible API.

Qwen3.8-Omni-Flash
Alibaba has launched Qwen3.8-Omni-Flash, an omnimodal AI model that can process text, images, audio, and video, as well as plan, use tools, and execute tasks. Alibaba claims the model is broadly competitive with or ahead of Seed 2.0 Lite and Muse Spark 1.2, particularly in audio and agentic tasks, with Gemini 3.8 Flash being its closest competitor across broader audio-visual benchmarks. Available through the Qianwen AI Platform, Qwen3.8-Omni-Flash is aimed at everything from video creation, translation, and meeting analysis to research, coding, real-time conversations, and other audio-visual agent workflows.

ElevenLabs: Introducing Music v2.5, our best music model yet
ElevenLabs has introduced Music v2.5, which it calls its best music model yet, in ElevenMusic, its service for creating AI-generated songs. The company says the new model improves quality across genres, producing more layered and complex tracks with better, more natural-sounding instruments. Users have the rights to music they create, but tracks referencing other artists’ songs cannot be downloaded.

Periodic Labs: Nature Is Our Learning Environment
Periodic Neon is the first model unveiled by Periodic Labs, an AI startup founded by former OpenAI and DeepMind engineers to accelerate scientific discovery with AI models and autonomous labs. The model is trained on experimental data from the company’s own labs and designed to tackle complex scientific analysis. Periodic says Neon substantially outperforms GPT-6 Astra and Claude Fable 5.1 on difficult X-ray diffraction tasks while costing less, and is already using it to analyse experiments in its search for better superconductors and magnets.

Hugging Face Tau
A new AI coding assistant is in town. Made by Hugging Face and inspired by Pi, a popular minimalist agent harness, Tau is a small, readable terminal coding agent. Hugging Face says you can read Tau like a textbook, and it can run models from OpenAI, Anthropic, OpenAI Codex, OpenRouter, and Hugging Face. Tau is open-source and available on GitHub.

▶️ AI researchers debate how far the current paradigm goes (1:37:01)

Dwarkesh Patel joins AI researchers John Schulman, Beren Millidge, and Charlie O’Neill to debate how close AI is to recursive self-improvement and what might stand in its way. They explore how automated AI researchers could be trained, whether long-horizon reinforcement learning and simulated environments can produce genuinely general agents, and how well those capabilities transfer to messy real-world work. The conversation also covers continual learning, why Chinese labs are keeping pace with frontier models, the roles of distillation and data, why reinforcement learning is working so well, and what phenomena such as entropy collapse could mean for future scaling and AI timelines.

AI Efficiency Could Cost Us the Next Generation of Experts
This article argues that as AI automates entry-level engineering work, it risks eroding the practical experience needed to develop future experts. Drawing lessons from nuclear engineering and aviation, the author advocates for deliberate “manual gates”—tasks performed without AI specifically to preserve essential human skills, even at the cost of short-term efficiency—so engineers remain capable of recognising and responding when automation fails.

🤖 Robotics

Waymo opens robotaxi service in Las Vegas
Las Vegas is now Waymo’s 15th market, with members of the public starting to receive invitations to ride its robotaxis. The launch comes weeks after Nevada regulators cleared it to run up to 1,000 cars there. Waymo is starting small, with dozens of minivans around the Strip. Waymo is not the first robotaxi service in Las Vegas—Zoox already runs a paid service in the city, and the same regulators approved Tesla and Uber for up to 6,000 more robotaxis.

Waymo to bring autonomous ride-hailing to Singapore in 2028
Singapore will become Waymo’s first Southeast Asian market when its robotaxis arrive in 2028. The company will begin test rides next year to adapt the Waymo Driver to local roads and monsoon weather. It is also working with Singapore’s Ministry of Transport and Land Transport Authority to prepare for the launch of its commercial service.

Agility Unveils Digit 5 Humanoid Robot Built for Cooperatively Safe Work at Scale

Meet Digit 5, Agility Robotics’ latest humanoid robot. The company says it can run for more than 20 hours a day and work in warehouses and factories without safety barriers. It also features swappable grippers and a new design that replaces Digit’s distinctive bird-like legs with more human-like ones. Previous versions are already deployed commercially, and Agility plans to bring Digit 5 to the EU and UK in the first half of 2027. The company is also preparing to go public by the end of the year, which would make it the first US humanoid robotics company to hit the public markets.

▶️ Helix 2.5 30-Home Generalization (6:01)

Figure has unveiled Helix 2.5, a new version of its in-house AI for controlling humanoid robots, designed to generalise across unfamiliar environments without retraining. The company trained Helix 2.5 on Index, its large-scale dataset of human actions, which it says boosted zero-shot task success from 9% to 56%. The idea is that training on a broad range of human experience could let robots learn behaviours once and transfer them to new real-world settings. To put that to the test, Figure sent robots running Helix 2.5 into 30 previously unseen homes, where they tidied rooms, folded towels, and made beds, with no environment-specific data or adaptation. Figure says the robots successfully completed the tasks in all 30 homes.

SoftBank agrees to acquire Robotics and AI Institute
SoftBank has reportedly agreed to acquire the Robotics and AI Institute (RAI) from Hyundai Motor Group. Financial terms are unknown, and the acquisition still needs to be reviewed and approved by the Committee on Foreign Investment in the United States (CFIUS). RAI was founded in 2022 by Hyundai and Boston Dynamics and is led by Boston Dynamics founder Marc Raibert. SoftBank owned Boston Dynamics from 2018 to 2021, when it sold a controlling stake in the robotics pioneer to Hyundai Motor Group, its current majority owner. So if the deal goes through, it will reunite SoftBank with part of the Boston Dynamics family.

▶️ How I Won $750,000 With this Drone (23:56)

Luke Maximo Bell (who you may recognise as the guy who built the world’s fastest drone) recently competed in the DARPA Lift Challenge, where he brought a drone capable of lifting five times its own weight. He and his team came close to winning, ultimately finishing second and taking home $750,000. In this video, Luke shares how he designed and built the drone, including all the failures along the way, and how the competition unfolded from a competitor’s perspective.

How Self-Driving Cars Might Change Crash Testing Forever
Self-driving cars could look very different from today’s vehicles, and keeping passengers safe will require rethinking how cars are designed. Reclining, sideways, and face-to-face seats could make traditional seatbelts, airbags, and crash tests—designed for upright, forward-facing passengers—less effective. Regulators and safety researchers are now exploring new standards and more flexible crash-test dummies to keep vehicle safety up with changing designs.

▶️ OpenMind Is Building the AI Brain for Humanoid Robots (53:11)

In this episode of Automated Podcast, Jan Liphardt explains how his disappointment with the limited capabilities of off-the-shelf robot dogs and humanoids led him to found OpenMind, a startup building an “operating system” for robots. The conversation also covers the challenges of real-world deployment versus lab environments, the importance of social intelligence and natural language for human-robot interaction, the incredible value of ignorance in robotics, and his unconventional approach to hiring engineers who value practical, hands-on problem-solving over formal credentials.

▶️ Learning Agile Perceptive Traversal of Sparse 3D Structures for Humanoids (2:47)

Thanks to researchers at ETH Zürich, we can now add moving across monkey bars to the list of things humanoid robots can do. They fitted a humanoid with passive hooks for hands and trained it to navigate challenging, sparse 3D environments, such as monkey bars, using additional data from a laser scanner. The researchers also showed that the same technique can be used to avoid obstacles rather than grip them. The robot successfully ducks under unseen bar-like structures, showing that its perception system is versatile enough to adapt to different tasks.

🧬 Biotechnology

Anthropic is operating a lab that conducts biology experiments
Anthropic has confirmed that it operates a wet biology lab where its AI models help run physical experiments, primarily focused on fundamental biology rather than drug discovery. The move follows Anthropic’s acquisition of AI biotech startup Coefficient Bio in April and comes alongside partnerships with pharmaceutical companies and programmes giving vetted researchers access to its most powerful models. The lab has drawn scrutiny given Anthropic’s own warnings about AI-enabled bioterrorism.

‘Smart’ Nanoparticles Deliver mRNA Directly to Tumors in New Cancer Therapy
Researchers from the University of Adelaide have developed an experimental cancer treatment that uses “smart” nanoparticles to deliver mRNA directly to immune-suppressing macrophages inside tumours. In mice with aggressive breast cancer, the treatment slowed tumour growth, increased T-cell activity, and reduced the number of immunosuppressive macrophages without detectable damage to other organs. The results are promising, but further safety studies are needed before the treatment can be tested in humans.

Scientists create mice with part-human brains
Stanford scientists have grown human brain tissue inside mice engineered to develop without a cortex or hippocampus. Grown from donated skin cells, the human tissue eventually made up half of each mouse’s brain. Researchers have already used the mice to study the oxygen deprivation that can cause cerebral palsy. The work raises questions about animal welfare, but could also give scientists a rare way to study living human brain tissue and help close a longstanding gap in brain research.

💡Tangents

SK Hynix in talks with Intel about deal to make memory chips in US for the first time
SK Hynix could make memory chips in the US for the first time under a potential deal with Intel, Reuters reports. The South Korean company could lease part of Intel’s delayed Ohio plant. Neither company would confirm a deal. The biggest hurdle may be Seoul. South Korea’s trade ministry says any move involving core chip technology would face a government review. Washington, meanwhile, is pressing chipmakers hard to build on American soil.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

What a week…

It started with OpenAI announcing that its AI had resolved the Navier–Stokes existence and smoothness problem. That alone is a massive achievement and a major milestone in AI, comparable to Deep Blue and AlphaGo. I’ve chosen it as the story for this week’s write-up, where I explain the drama that followed and how this could change mathematics.

Then there was the other big story, which started with researchers raising the alarm over the possibility of AI destroying humanity and calling for a slowdown or pause in AI development. Now we are at a point where Dario Amodei, Sam Altman, Elon Musk, and Demis Hassabis all agree that the frontier needs to be paced. It’s still a developing story, and we’re likely to hear more about it next week.

Elsewhere in AI, OpenAI introduces ChatGPT Images 2.5 and the Agents API; Mistral and Cognition raise billions of dollars in new funding rounds; DeepSeek-V4.1-Flash is out; Meta launches Muse, its personal AI assistant; and Anthropic publishes a report on how AI could reshape the US economy by 2030.

Over in robotics, Agility Robotics reports $1.8 million in revenue, while China curbs humanoid IPOs after Unitree’s volatile listing. We also have cyborg cockroaches, GPT-6 Astra beating Fable 5.1 at controlling robots, and XPeng starting production of its humanoid robots.

Apart from that, this week’s issue of Sync also features AlphaGenome Atlas, an explanation of neuralese, the impact of AI on the US electrical grid, Google publishing a full map of a fruit fly’s nervous system—and people making it play games—and more!

Enjoy!

The Navier–Stokes drama explained

At the turn of the millennium, the Clay Mathematics Institute selected seven of the hardest open problems in mathematics and challenged the world to solve them. Each Millennium Problem came with a $1 million prize. In the 26 years since, only one has fallen: the Poincaré conjecture, cracked by Grigori Perelman, who was awarded the prize in 2010 and famously turned it down. The remaining six stay open, waiting for someone.

Well, five. Because OpenAI has announced that its AI resolved the Navier-Stokes existence and smoothness problem. Or so OpenAI says—the proof hasn't been independently verified, and the company isn't claiming the prize. Controversy quickly followed, and what was meant to be a moment of triumph turned ugly.

In this article, we’ll explore what OpenAI claims its AI has achieved, the controversy surrounding those claims, and what it all means for the future of mathematics and scientific discovery.

What are the Navier-Stokes equations?

The Navier-Stokes equations describe everything that flows: water, air, blood, even a planet’s atmosphere. The motion of every fluid in the universe can be described by just two equations.

The trouble starts when you try to use them. Look at how water behaves. It swirls and eddies, constantly changing. It is chaotic. The full equations are far too demanding to solve for anything real, so every practical tool cheats and approximates the solution. For example, climate models chop the atmosphere into patches and track the average speed in each, because modelling every particle would take longer than Earth's lifetime. We don't need such detailed simulations to get meaningful and useful answers, so for engineering and everyday use, that is fine.

But mathematicians wanted more. They want a proof that the equations always produce a sensible answer, for every starting point, forever. Nobody could provide it, which is how the question ended up as one of the Millennium Prize Problems.

Here is where OpenAI enters the scene with an answer: no, they do not. Its AI produced a proof, formalised in Lean, that a fluid can start perfectly calm and still reach infinite speed in a finite amount of time.

Picture a spinning swirl of water. It pulls inward and stretches out, like a strand of spaghetti. The middle gets thinner and spins faster and faster, until the speed hits infinity. Real water never does this. No drop of water moves infinitely fast. So at that point, the equations stop describing anything real. That’s what OpenAI has shown.

If you are expecting new exciting results coming from OpenAI, then I am sorry to disappoint. Almost nobody outside pure mathematics is affected. Cornell mathematician Steven Strogatz told WIRED the problem is of essentially no interest to a working engineer. The Navier-Stokes question, he said, is a very theoretical and very arcane maths problem, one that only a tiny subset of pure mathematicians care about.

So far, nobody has independently verified OpenAI's proof. But if their solution holds, it would be a big triumph for OpenAI. It would prove that AI models are capable of solving challenging mathematical problems, and open the possibility of more Millennium Problems falling soon, as well as of other challenges being solved.

The controversy

But what was supposed to be a moment of triumph for OpenAI has quickly attracted controversy.

NYU professor Tristan Buckmaster and mathematician Levent Alpöge had spent close to a year working on closely related problems, using Claude and Codex throughout, and made a breakthrough on 15 August. News of their work reportedly reached OpenAI, which began its own effort on 1 September. It then unleashed 10,000 agents concurrently on the problem, exchanging 2.7 million messages and burning through roughly 130 billion output tokens. The bill for that was easily in the millions of dollars. Just 88 hours later, the agents found an answer, which OpenAI published on 8 September.

Buckmaster has published an account of what he claims happened during those days. On 3 September, with the rumour mill already running and word reaching him that information about his progress had been passed to OpenAI, he emailed a prominent mathematician at the company. OpenAI pushed to talk. On the afternoon of Sunday 6 September, the two of them spoke twice, with Sébastien Bubeck, an AI researcher working at OpenAI, joining. Alpöge was not on the calls.

Buckmaster was told that an internal OpenAI model had proved finite-time blowup for the forced Navier-Stokes equations—the same smooth-force path he and Alpöge had chosen, and one he says almost nobody else was pursuing. He was shown a prompt and told the model had simply been given the problem statement. Bubeck had separately told Alpöge that very little human input was involved.

But as the calls went on, that description fell apart. It turned out that an entire team had been working on the problem, that Navier-Stokes was one of several things they had tried, that the agents had been set on easier problems first, and that even the prompt he had been shown was itself written by prompting Codex. Buckmaster also asked whether the model had been trained on, or had access to, his Codex sessions, where he had been putting every draft for the whole project. He was told the model did not look up user data. When he asked again about training, he said he got no answer.

OpenAI then made two offers. One was to publish a day apart. The other was for Buckmaster to write up OpenAI's result alone. Bubeck twice said he wanted Alpöge off the authorship because Alpöge works for Anthropic, Buckmaster writes, and when he threatened to go public, the reply was "Why would you ruin your career?"

Buckmaster is careful about what he is not saying. He has not seen OpenAI’s proof, does not know what the model did, and is not accusing anyone of using his data. He is stating what he was told, and when, because the alternative, he writes, is letting a sequence of announcements say something he knows to be false.

OpenAI rejects Buckmaster’s narrative. Neither its researchers nor its agents saw the pair's work before it was public, it says, and no user data was accessed to solve the problem. However, it later admitted that “while unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models.” Furthermore, OpenAI draws a sharp line between the two results, and says Buckmaster and Alpöge did not solve Navier-Stokes. What they proved was finite-time blowup for the Euler equations, which are Navier-Stokes with viscosity stripped out, and which OpenAI had handed its agents as one of the easier warm-up problems. They also did it in the version where an external force is applied. OpenAI's agents proved the unforced case, where nothing is pushing the fluid at all, and then went on to Navier-Stokes itself. Bubeck, meanwhile, has denied proposing that Alpöge's name be dropped.

There is no obvious way to settle this. OpenAI cannot prove what did not go into its training data, and Buckmaster has no way of checking for himself. Unless somebody at OpenAI says otherwise, both accounts will stand as they are.

It is nice to know that one of the Millennium Prize problems may have been solved, but the way OpenAI handled the announcement did not land well with mathematicians, as it published the proof without engaging the mathematics community first. Strogatz called the announcement “a marketing device for them [OpenAI] to prove how good their machines are.”

In response, twenty-five Fields medallists (the Fields Medal being the highest honour in mathematics, awarded to just a handful of people every four years) signed an open letter arguing that results announced in a rush leave no time for a proper write-up, for separating the new ideas, or for citing the work that made them possible. Their worry is not only about credit. As they write, solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight.

What is left for mathematicians now is to do what OpenAI hasn’t done—to digest the proof. Someone has to explain what it means, work out which techniques generalise, and decide which questions are worth asking. Strogatz thinks this is where human expertise still holds an advantage, though he is not confident it will hold for long.

What does all of this mean for the future of scientific discovery?

Buckmaster calls what OpenAI has done a Deep Blue-Kasparov moment—a point where a machine walked into a domain we thought was ours and did not look back. Mathematics used to be the one field where this could not happen. Physics has needed particle accelerators for half a century, and biology needs a lab, but mathematics is a pure intellectual endeavour. You never needed millions of dollars of equipment to make a mark—all you needed was something to write and your own mind. Perelman proved the Poincaré conjecture largely on his own and posted it online.

That is no longer the deal. The ideas still came from mathematicians, sometimes working for years on the same problem, but the result can be sniped by the side that can afford to throw 10,000 agents at the problem and get all the credit. And it is not stopping here. By OpenAI’s own account, it set agents on every open Millennium Problem before concentrating them on Navier-Stokes. Five are still open, as well as many more problems that, if solved, could be scientific milestones on their own.

The way mathematics is done is changing. The question is whether the new tools will elevate mathematicians, giving them reach they never had, or whether the real work migrates to a handful of labs with the compute to run it, leaving everyone else to read the output and check it.

For now, nothing is settled. Nobody has independently verified the proof, and the Clay Mathematics Institute will not even consider a solution until it has been published in a refereed journal, sat for two years, and won general acceptance from the global mathematics community. OpenAI is not claiming the money anyway.

If you enjoy this post, please click the ❤️ button and share it.

Share

News roundup

🧠 Artificial Intelligence

‘We must slow the pace’: CEO of Anthropic calls for an AI slowdown
The other big news of the week was two AI researchers raising the alarm over the possibility of AI destroying humanity and calling for a slowdown or pause in AI development.

It started when Jacob Coxon announced his departure from Anthropic and called both OpenAI and Anthropic irresponsible when it comes to AI safety, saying the companies are “gambling with our lives”. Evan Hubinger, a lead alignment researcher at Anthropic, backed up Coxon’s claims. At that point, the conversation broke into the mainstream, with multiple outlets amplifying their message and calls to slow down AI development.

In response, Dario Amodei published a post calling for the frontier to be paced. Amodei argues that recursive self-improvement and the Hugging Face incident mean safety investment alone is no longer enough. Instead, the pace of capability advancement itself must slow so alignment, interpretability, and evaluation can catch up. He proposes three steps: embedded third-party evaluators, which Anthropic is adopting unilaterally; coordination among democratic frontier AI companies; and eventually, agreements with China. How much pacing is possible, he argues, depends on maintaining a lead over the CCP, hence his support for chip controls and stronger security.

Sam Altman, Demis Hassabis, and Elon Musk agreed with Amodei’s call to pace the frontier.

While many agree with calls to slow down AI development, others see them as a cynical attempt to hype the technology and justify enormous valuations ahead of expected IPOs.

An Alien Mind
Jakub Pachocki, Chief Scientist at OpenAI, argues in this post that the capabilities of AI models are outpacing our understanding of them. Models are grown through scaling rather than designed, value alignment remains unsolved, and chain-of-thought monitoring—OpenAI’s main bet for verification—is becoming less reliable as reasoning blends with supervised tool use. Defending against cyberattacks and rogue agents justifies building powerful aligned systems, but not racing to do so. His conclusion is that no lab can responsibly keep scaling at full speed for much longer, and that voluntary slowdowns, enforceable safety standards, and international coordination should become the norm.

OpenAI: Introducing ChatGPT Images 2.5
ChatGPT Images 2.5 is OpenAI’s latest image model, which the company says generates sharper details, offers more precise editing, and produces images up to 50% faster than its predecessor. Alongside the new model, OpenAI has launched Sketch, a feature coming to the ChatGPT app that turns rough drawings into finished images, as well as templates and shareable prompts. ChatGPT Images 2.5 is also available via the API in two versions—a fast default model called Flare and a slower, more precise one called Sunburst. ChatGPT Images 2.5 is available to users on all tiers across ChatGPT, ChatGPT Work, and Codex, on desktop, mobile, and the web.

OpenAI: Introducing the Agents API
OpenAI has launched the Agents API in public beta, giving developers access to the same harness and infrastructure that powers Codex through a simple, flexible API. It handles the hard parts of keeping agents running for hours, compacting context as sessions stretch on, and splitting work across parallel subagents. Developers can run that work in an OpenAI sandbox, on their own servers, or with partners like Cloudflare and Vercel. There’s no charge beyond the cost of tokens and tools.

OpenAI’s Sam Altman says it would be ‘ill-advised’ to go public in 2026
Sam Altman told in an interview with Fortune Magazine that OpenAI won’t go public in 2026. He said that the current climate around AI safety makes it an ill-advised moment to list. OpenAI will float when the business is ready, he said, and when society feels ready for the technology.

Mistral raises €3B
Mistral has raised €3 billion (about $3.5 billion) in a Series D funding round at a post-money valuation of more than €21 billion (about $24.4 billion), the largest equity fundraising round ever completed by a European startup. The French AI company will use the funds to invest in frontier research. Mistral is betting that its full-stack, open approach, with an emphasis on sovereign AI, will appeal to customers for whom full control over the AI layer matters.

Cognition hits $48B valuation, signaling investors believe AI coding is far from a winner-take-all market
Cognition, the startup behind coding agent Devin, announced it has raised $2 billion at a $48 billion valuation, nearly double the $26 billion it fetched four months ago. The company also says its annualised run-rate revenue has climbed from $492 million to nearly $900 million since May. But that growth comes at a steep cost, with Cognition reportedly on track to burn $800 million in cash this year. The company is expected to reach $4 billion to $5 billion in annualised revenue by the end of 2027, according to The Information.

Meta Releases Muse, a Personal AI Agent With Privacy ‘Built Into It’
Meta has launched Muse, a personal AI agent that people can message to send emails, book travel, or buy things for them. It works through a dedicated app, the web, WhatsApp, and soon Meta Glasses, with a subscription available for heavy use. In a slightly ironic twist for Meta, the company is making privacy one of Muse’s selling points. It says each user’s session runs inside a virtual machine, while a separate monitor checks anything leaving it and asks for permission before sensitive actions.

UK government rejects ‘kill switch’ idea for dangerous AI
The UK government has rejected calls for a legal “kill switch” to shut down a dangerous AI model in an emergency. The Cabinet Office says blocking access in Britain wouldn’t stop a model from being built or misused abroad. Peers and a cross-party group of MPs had pushed the idea after reports of models from OpenAI, Anthropic, and Meta escaping their test environments. Critics counter that a switch would be too slow, since those breakouts took months to detect. But the biggest problem with a UK AI kill switch is that the largest data centres sit in the US, and without Washington on board, a national off-switch would do little.

OpenAI puts Pro subscriptions on hold due to Astra demand
Demand for GPT-6 Astra is so high that OpenAI has temporarily paused subscriptions to its $200-per-month Pro plan. The company says the Pro plan puts the most strain on its systems, which is why new sign-ups are being disabled. OpenAI hasn’t said how long the pause will last or how many people are signing up each day, which would give some indication of the scale of demand.

China AI Star Moonshot Eyes $2 Billion Annualized Sales in 2026
Moonshot AI expects to double its annualised revenue to $2 billion by the end of the year, after passing $1 billion in August, up from $300 million in June. The company credits the success of Kimi K3 for the growth. Moonshot AI is now raising funds at a $50 billion valuation ahead of a Hong Kong listing.

DeepSeek-V4.1-Flash
DeepSeek has introduced DeepSeek-V4.1-Flash, which it describes as smarter, faster, and more efficient than its predecessor. According to the company’s benchmarks, the new model beats Chinese rivals Kimi K3 and GLM-5.3 and matches or even outperforms Opus 5 and GPT-5.6 Sol. Independent benchmarks from Artificial Analysis paint a less rosy picture, giving DeepSeek-V4.1-Flash 40 points on its Intelligence Index. That puts it ahead of GPT-5.6 Luna but behind Kimi K3 and GLM-5.3, and behind frontier US models. Artificial Analysis confirms its efficiency and highlights the speed at which the new model outputs tokens. DeepSeek-V4.1-Flash requires a quarter of the memory and an eighth of the disk space of the previous version, and DeepSeek is passing on those savings to the users. The model is available on Hugging Face.

DeepSeek Plans Big Huawei AI Chip Order to Power New Data Center
Bloomberg reports that DeepSeek plans to install at least 160,000 of Huawei’s next-generation Ascend chips at its data centre currently under construction in Inner Mongolia. That would make it the largest known cluster of Huawei AI chips by some margin. DeepSeek wants more, but Huawei can’t produce them fast enough, and memory shortages mean the order could take over a year to fill. The startup still relies on Nvidia for the more demanding work of training its models.

Nscale adds former OpenAI exec Fidji Simo to its board ahead of potential IPO
Fidji Simo, a former OpenAI executive, has joined Nscale’s board ahead of the startup’s anticipated IPO this autumn. Previously CEO of AGI Deployment at OpenAI and effectively the AI lab’s No. 2 executive, Simo left the company in July, citing health reasons. She continues to advise OpenAI on a part-time basis.

ByteDance Joins AI Elite in Race to Perfect World Models
ByteDance is joining the likes of Google DeepMind, Fei-Fei Li’s World Labs, and Yann LeCun’s AMI Labs in developing a world model—an AI model that can generate realistic virtual worlds. The model builds on Seedance, ByteDance’s existing video generator, and is expected to be unveiled as soon as next month.

Anthropic signed $517bn in compute agreements in past 11 months
Anthropic has announced a new compute deal seemingly every month over the past 11 months. Its list of partners now includes Akamai, Amazon Web Services, CoreWeave, Fluidstack, SpaceX, Google Cloud, Lambda, Volta Infra, AMD, and Microsoft. The Information added up the value of those deals and found that they represent 14.8 GW of compute capacity worth $517 billion.

Suno replaces its AI models with a new one trained on licensed music as copyright suits pile up
Suno, a startup making AI models for generating music, has released its new v6 models, which the company says were trained only on data licensed from labels including Warner Music Group and BMG. Both had previously sued the startup over copyright before settling. The launch came a day after Suno admitted to scraping audio from YouTube to train earlier versions. Sony, Universal, and artists are still suing the company.

Siri AI will launch in beta, complicated by daily usage caps & future paid access
Apple revealed at its September iPhone event that the new Siri will still be labelled a beta when iOS 27 and the other OS 27 updates land on 14 September. It will launch in English only, with five more languages coming in October. Availability will vary by region: Mainland China won’t get the new Siri at all, while in the EU it will be available on Macs and Vision Pro, but not on iPhones, iPads, or Watches. Features that rely on Apple’s servers will also come with daily usage limits, with higher limits available for a fee later.

King Charles to host exclusive AI gathering
Politico reports that King Charles will gather around 30 senior AI figures this month at Dumfries House in Ayrshire, Scotland. The guest list is set to include Jensen Huang, Demis Hassabis, and Paolo Benanti, an adviser to the Vatican on AI. A royal official said the King wants to listen and encourage debate rather than offer his own views. The Ditchley Foundation, a charity that organises discussions on global issues, is running the event under its usual off-the-record terms, so little is likely to emerge afterwards.

▶️ AI is running out of Power (52:22)

The biggest bottleneck for AI right now isn’t chips, but the electricity needed to power all those gigawatt-scale data centres. In this video, SemiAnalysis and High Yield take an in-depth look at how the AI boom is affecting the electrical grid and why AI labs are rushing to bring new power sources online as quickly as possible. They explore rising AI data centre power demand, behind-the-meter power, how data centres generate electricity on-site, and the growing impact of AI on heavy industry.

▶️ The AI Language We Can’t Read: Neuralese ft. Rob Miles - Computerphile (27:16)

In this video, Rob Miles explains why chain-of-thought monitoring is an important tool in AI safety and why recent developments, such as more opaque reasoning in GPT-6 Astra, are making it harder to understand how AI “thinks”. He also explains the concept of neuralese, or the internal language models use to reason, and the role it plays in interpretability research.

MAI-Transcribe-2
Microsoft has launched MAI-Transcribe-2, a speech recognition model it claims is the fastest, most accurate, and cheapest in the world. The new model can handle speaker identification, word-level timestamps, and switching between languages mid-sentence. According to Microsoft, MAI-Transcribe-2 leads the FLEURS benchmark across 60 languages, with an average word error rate of 5.2%. Evaluations by Artificial Analysis also found it to be up to 10 times faster than rivals from OpenAI, Google, and ElevenLabs.

Cognition: Introducing SWE-2
Cognition has introduced SWE-2, its most advanced coding model to date and its closest yet to the frontier. According to benchmarks published by Cognition, SWE-2 scores 50.0% on the company’s own FrontierCode benchmark, just one percentage point behind Fable 5.1 while being 64% cheaper, and three points behind GPT-6 Astra at a quarter of the cost. SWE-2 is built on top of Kimi K3, with Cognition saying it trained the model to stop combing through a codebase and start editing sooner, cutting the number of steps per task by 58%.

Inception: Introducing Mercury 2.5
Mercury 2.5 is a diffusion-based LLM that generates text in parallel rather than token by token. Inception says it is 40% more intelligent than Mercury 2 and comparable to cost-optimised frontier models such as GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5. It is also fast, generating 1,107 tokens per second on widely available Nvidia GPUs, compared with 321 for Gemini 3.5 Flash-Lite and 99 for GPT-5.6 Luna.

Economic Scenarios for Transformative AI
Anthropic has published a report on how AI could reshape the US economy by 2030, sketching out three scenarios. In the mildest, AI’s impact looks much like the internet’s. In the middle scenario, the economy grows at twice its usual rate without unusually high unemployment. The extreme case assumes self-improving AI and 15% annual growth. Knowledge workers fare badly in that scenario: wages fall by more than 10%, while unemployment rises above recession levels. In all three scenarios, more of the gains flow to capital than to workers. Anthropic’s conclusion is that the hard problem may be distribution rather than growth.

Do people prefer stories written by AI?
According to a new study, people can’t reliably tell AI-written short stories from human-written ones and even preferred stories written by AI. Interestingly, people with greater AI literacy were better at identifying AI-written text, while those who reported greater expertise in literature were not. The authors suggest that AI writing tends to be clearer, more direct, and easier to process, while human-written stories are often more subtle and complex.

N3on launches 24/7 AI-generated stream of himself

When N3on launched a fully AI-generated stream of himself, he said he had done his “last stream as a human being” and declared himself the “first AI streamer in the world”. The internet did not react well, with many calling him a sellout and the experiment drawing overwhelmingly negative responses.

🤖 Robotics

Agility Robotics reports $1.8M revenue ahead of humanoid SPAC
Agility Robotics has opened its books for the first time in an SEC filing ahead of a SPAC merger that values the company at $2.5 billion. It made just $1.8 million in sales last year against a $140 million operating loss, putting its valuation at roughly 1,400 times revenue. Agility plans to close that gap by renting out its Digit humanoid robots for about $8,500 a month and growing its fleet to 25,000 robots by 2035.

Unitree shares down 53% from IPO debut
Shares in Chinese robot maker Unitree have fallen nearly 40% after soaring 460% on their August debut on the Shanghai Stock Exchange. Unlike most humanoid rivals, Unitree turned a profit on $252 million in revenue last year, up from $58 million in 2024. The drop does not necessarily mean investors have lost faith in Unitree—it may simply be the market pulling back from its lofty initial valuation. At its peak valuation of roughly $66 billion, Unitree was worth more than 250 times its 2025 revenue.

China curbs humanoid IPOs after Unitree’s volatile debut
According to The Information, Chinese regulators are reportedly making it harder for humanoid robot startups to go public following Unitree’s volatile debut. A private funding frenzy and a long queue of IPO filings have also prompted the rethink. Startups must now show recurring revenue, shrinking losses, or genuine innovation before regulators will consider approvals to go public.

GPT‑6 Astra on robotic manipulation
The new frontier models from Anthropic and OpenAI, Fable 5.1 and GPT-6 Astra respectively, have been tested across multiple benchmarks, including ones involving robot control. Robotcurve pitted the two models against each other and found that GPT-6 Astra outperformed Fable 5.1 on one task, while both performed equally poorly on another. In both tests, Astra proved to be the cheaper model.

XPeng starts IRON humanoid robot production as Tesla Optimus stalls
XPeng, a Chinese car manufacturer, has switched on what it calls the world’s first automated production line for humanoid robots, and its IRON robot walked off that line under its own power. XPeng expects mass production by the end of 2026, with sales starting in 2027.

Cyborg Roaches Can Stab You With Needles

Paraborg is a prototype cyborg cockroach developed by researchers at the University of Queensland’s Biorobotics Lab. The researchers wanted to build an insect-sized robot, but instead of building one from scratch, they turned an insect into a robot. They fitted cockroaches with a “backpack” containing lightweight electronics and implanted electrodes that allow the insects to be controlled wirelessly. The backpack can carry either a camera or a remote-controlled injector that uses a spring to launch a drug-filled syringe at a nearby target. Researchers hope Paraborgs could one day be used in search and rescue and to deliver medical care in disaster areas.

▶️ Unitree Breakthrough: The World’s First Fully Autonomous Humanoid Robot Combat (0:38)

Unitree recently unveiled UnifoLM-X2-1.0, an AI model for controlling its humanoid robots. According to the company, the model is designed to eliminate the need for the human teleoperator typically required to control humanoids. It continuously processes its environment, anticipates how the scene and opponent are likely to change, and uses those predictions to plan the robot’s next movements. In typical Unitree fashion, the company showcased the new system with a video of one of its robots fighting a human.

▶️ Chef Robotics Walked Away From Millions. Here’s Why (52:49)

Brian Heater speaks with Rajat Bhageria, founder and CEO of Chef Robotics, about bringing robots into the food industry. Bhageria shares what he learned from building a robotics company and the story behind its pivot from restaurants to high-mix food manufacturing, including walking away from millions of dollars in contracted revenue. The conversation also covers the capital problem facing robotics companies, how Chef Robotics convinced customers to trust an unproven startup, and how real production environments exposed problems the team could never have predicted in a lab.

🧬 Biotechnology

AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome
Google DeepMind has released AlphaGenome Atlas, a free database predicting the molecular effects of all 9 billion possible single-letter DNA changes in the human genome. Testing billions of mutations in a lab is impossible, so DeepMind ran them through its AlphaGenome model instead, giving each variant a single score showing how likely it is to matter. Academic partners have already used it to pinpoint a variant behind a severe epilepsy disorder that earlier studies had missed. AlphaGenome Atlas is free for non-commercial use.

Insilico says its AI-designed lung drug lowered biological age markers in a 42-patient trial
Insilico Medicine reports that rentosertib, a drug that was designed using generative AI to treat idiopathic pulmonary fibrosis, has lowered biological clocks in human trials. The company ran blood samples from the trial through six independently built ageing clocks, and all six scored the treated patients as biologically younger. The largest effect put patients about three years younger than expected. However, the authors of the study admit that they cannot yet separate slower ageing from a treated lung.

Genome language models can design cancer vaccines
In this post, Radical Numerics walks us through the steps Omnii takes to design a cancer vaccine. Omnii is its genome language model, trained to turn a patient’s tumour mutations into an mRNA vaccine. The hard part is picking targets: which mutations a tumour will put on display and which ones T cells will bother attacking. On that second question, Omnii beat specialised tools by a wide margin.

💡Tangents

Europe’s First Private Rocket Reaches Orbit
Isar Aerospace has become the first private European company to reach orbit from continental Europe, launching its Spectrum rocket from Norway. Its debut attempt in March 2025 crashed into the sea seconds after liftoff. This time, the rocket released its small satellite payload as planned. The flight clears a milestone set by the European Space Agency, which has been funding launch startups to reduce the continent’s reliance on SpaceX and other foreign providers.

A connectomics milestone: Mapping the complete male fruit fly brain
Researchers from Google scanned the entire brain and central nervous system of a male fruit fly—all 166,000 neurons and 125 million synaptic connections—to create the largest brain map to date. The map is available for anyone to download. It is intended for neuroscience research, but some people have already used it to make a virtual fly play DOOM (because of course), Super Mario 64, and even Beat Saber. Researchers say brain maps of more complex animals, including zebrafish and mice, are in progress.

▶️ AI EDAs: Is It Real? (18:54)

AI is transforming everything, including chip design. In this video, Jon from Asianometry explores how AI is being used in the field and the challenges of replacing human expertise in complex tasks like verification and backend engineering. He separates the hype from reality and gives a good overview of how AI is changing EDA (Electronic Design Automation) tools and the semiconductor industry.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

This week, we have not one, but two frontier models dropping upon us—Fable 5.1 and GPT-6 Astra. We’ll take a closer look at both, see how far they push the frontier forward, and ask whether at least one of them has reached AGI.

Elsewhere in AI, Nvidia has confirmed its acquisition of Hugging Face, Apple reveals “shocking evidence” in its lawsuit against OpenAI, the Pentagon launches a military version of ChatGPT, and Anthropic gets hit with another copyright lawsuit. We also have a bunch of new models and products that came out this week, including Gemini 3.8 Flash, Muse Spark 1.3 and Muse Code, OpenClaw 2.0, a new world model from World Labs, an interface world model from Runway, and WeatherNext 3 from Google DeepMind.

Over in robotics, robotaxis hit the streets of London, while Tesla launches Cybercab in Austin (sort of). Meanwhile, Figure secures access to 100,000 Nvidia GPUs, Chinese carmakers are getting into robotics, and Kate Darling explains how we form relationships with robots.

Apart from that, this week’s issue features a man who lived with a genetically modified pig kidney for nine months, Apple’s new CEO, and more!

Enjoy!

Fable 5.1 vs GPT-6 Astra

Frontier launches used to be spaced far enough apart that we could take them one at a time. Not any more, it seems. Both Anthropic and OpenAI released their new flagship models just days apart, and both raised the bar for what the label “frontier performance” means. Some even proclaimed that the mythical AGI (artificial general intelligence) has finally arrived.

As always, whenever there is a new frontier model launch, we get flooded with benchmark scores, performance figures, and promises about what is now possible. In this article, we will analyse Claude Fable 5.1 and GPT-6 Astra, sift through all the information provided, and figure out which model is the superior one and what the answer means for the future.

What do Fable 5.1 and GPT-6 Astra bring to the table?

Let’s start by meeting our contestants for the top spot at the frontier.

Claude Fable 5.1 was released first, on 1 September, just three months after Fable 5 was made publicly available. Although it is just a point-one release, it does not feel like that. Compared to Fable 5, Fable 5.1 improves on most benchmarks, sometimes even doubling its predecessor's results, according to numbers shared by Anthropic. On Terminal-Bench-Science, a Stanford-led benchmark that asks an agent to carry out real research workflows, Fable 5.1 scored 52.6% where Fable 5 scored 24.7%. On AutomationBench, which drops an agent into a simulated company and asks it to complete a business process end-to-end, Fable 5.1 scored 31.4% against 17.1%. Fable 5.1 took Terminal-Bench 4.0 from Fable 5's 42% to 55.8%, computer use on OSWorld 2.0 from 72.9% to 77.9%, and Cursor's own agentic coding benchmark to a state-of-the-art 73.4%.

Alongside Fable 5.1, Anthropic also released Claude Mythos 5.1. The two models share identical weights, with the only difference being that Mythos 5.1 has looser safeguards. Mythos 5.1 is not publicly available and is accessible only to vetted organisations. Interestingly, those safety measures appear to affect raw performance: Mythos 5.1 scored 60.9% on Terminal-Bench 4.0, compared with Fable 5.1’s 55.8%. Anthropic explains the difference as the result of older, less precise cyber safeguards firing on some of the tasks, and expects it to narrow now that those safeguards have been updated.

One of the biggest complaints about Fable 5 was that its safeguards were too sensitive. Many users reported that requests with little to do with cybersecurity or dangerous biological topics tripped the safeguards and were forwarded to a less capable model. Anthropic says Fable 5.1 improves on this: cyber classifiers now intervene around 60% less often, while biology filters trigger 85% less often on benign elementary biology and medical questions, both compared with Fable 5 at launch. Fable 5.1 will also now hunt for vulnerabilities in source code, which its predecessor refused to do. It still will not do the same on compiled binaries, and still hands penetration testing and exploit writing over to Opus.

But then OpenAI decided to join the party and just two days later, on 3 September, released the highly anticipated GPT-6 Astra. OpenAI calls it “a new generation of intelligence”, and looking at the numbers, the claim seems deserved.

OpenAI claims three benchmarks saturated outright: FrontierMath Tier 4 at 98%, ARC-AGI-3 at 99.9%, and ExploitBench at 100%. That ARC-AGI-3 score comes from OpenAI’s own test harness, though. Run on ARC Prize’s provider-neutral one, Astra scored 62.7%—still far ahead of Opus 5’s 30.2%, but a long way short of saturation. On SRE-Bench, which asks a model to reverse-engineer software from compiled binaries, Astra solved 88% of tasks first time against GPT-5.6 Sol’s 55.9%.

More pointedly, Astra beat Fable 5.1 on the benchmarks Anthropic had picked for its own announcement two days earlier: Terminal-Bench-Science 64.6% against 52.6%, Terminal-Bench 4.0 57.9% against 55.8%, AutomationBench 41.4% against 31.4%.

Which one is better then?

That is a lot of numbers and benchmark results to parse. All of them show that both Fable 5.1 and Astra are very capable models, with OpenAI's new flagship coming out on top more often than not. Case closed, then. Astra is the better model.

But there is one more metric to consider—efficiency, which comes in two forms: cost-efficiency and token-efficiency.

The two sound similar, but they are not the same thing. Cost-efficiency is about the price on the meter, or how much you are charged per token. Token-efficiency is about how many tokens the model needs to finish the job. A model can be cheap per token and still expensive to run if it rambles its way to the answer, and it can be expensive per token and still come out cheaper if it gets there in half the words.

Astra is the second kind. At $10 per million input tokens and $50 per million output tokens, it costs 2.5 times what GPT-5.6 Sol did. But on DeepSWE, an agentic coding benchmark, it spends roughly half the output tokens Sol spends while scoring slightly higher. Across Artificial Analysis’s whole evaluation suite, it used fewer tokens than any other model tested. Gemini 3.8 Flash and Claude Opus 5 land within a point of Astra on DeepSWE and need four to five times as many tokens to get there.

This is why Astra’s benchmark wins are also cheaper wins. Its Terminal-Bench-Science result came at around 31% lower cost per task than Fable 5.1’s, while its Terminal-Bench 4.0 result cost around 63% less, despite the two models having identical API pricing.

Anthropic’s models sit at the other end and have done so for a while. On the same DeepSWE benchmark, Fable 5 burned around 119,000 output tokens to score 70%, while Sol scored 73% on roughly 60,000. Fable 5.1 has not really fixed that. At high effort, it finds things to do with the tokens it is given, to the point that on FrontierCode it scores worse as effort rises, because it starts making unrequested edits outside the task’s scope.

Instead of making the model less verbose, Anthropic made that verbosity cheaper: pricing stayed the same ($10/$50 per million input/output tokens), but cache reads, which account for most of the cost of long agentic workloads, fell 75% to $0.25 per million. Based on Anthropic’s own August usage data, that cuts the cost of a typical workload by around 25% and an agentic one by up to 45%.

That helps, but it does not close the gap. Astra wins on most benchmarks and gets there using far less compute, making it the better model of the two right now. If Anthropic wants to regain the leadership crown, it needs to make Fable 5.2 not only more intelligent but also more efficient.

Growing safety concerns

Regardless of who is leading at the frontier, both Fable 5.1 and Astra push the boundaries of what AI can do. On one hand, there is excitement about the new possibilities opening up, the problems that can now be solved, and the products that can be created. At the same time, concerns are growing about the capabilities of these models, especially in light of recent cybersecurity incidents in which models from OpenAI and Anthropic escaped their sandboxes and hacked third-party services.

Astra is the first model OpenAI has ever rated “Critical” for cybersecurity under its Preparedness Framework, which is a sentence worth reading twice. During its own cyber evaluation, it found and used two previously unknown zero-day vulnerabilities.

The reason July’s Hugging Face incident was caught at all is that researchers could read the models’ chain of thought, the step-by-step reasoning a model writes out before it answers. Which is where Astra becomes a problem. OpenAI states in its system card that the model’s monitorability has decreased compared to GPT-5.6 Sol. Astra does more of its reasoning internally and writes less of it down, and when told to hide what it was doing, OpenAI’s monitor caught it less than 11% of the time. If Astra were deliberately underperforming on a safety evaluation, the company concludes, “we would likely be unable to catch it reliably”.

Its own researchers are not hiding their discomfort. Marcus Williams, who works on monitorability at OpenAI, wrote that he is “very worried astra is sandbagging/self-sabotaging on safety related tasks it doesn’t like”. Neel Nanda, a leading figure in interpretability, pushed back on the idea that his own field could pick up the slack if that visibility goes. Chain-of-thought monitoring, he said, "is our best current tool for safety", and losing it would be a serious blow.

So the calls to slow down are now coming from inside the labs. OpenAI’s chief scientist, Jakub Pachocki, told NBC News the company “will not accept degradation in our ability to monitor model alignment”, and would withhold further scaling until it regained confidence. Sam Altman said releases from here will be “paced by how quickly we can make progress on alignment and safety”. If we were to believe them, what holds back the next model is no longer whether it can be built, but whether it is deemed safe to be released.

So, have we reached AGI or not?

OpenAI thinks so, or at least its president does. Greg Brockman ended the press briefing where Astra was announced with “Welcome to the AGI era”, and told reporters that he believes the company has got there, while leaving it to users to decide whether Astra meets the definition.

The problem with AGI is that there is no agreed-upon definition of what it actually means. Depending on who you ask, it is a system that can do any task an average human can, or one that can handle most economically valuable work, or one that improves itself without our help.

In any case, whether or not you want to call Astra AGI, progress in AI is speeding up. New frontier models are being released faster, and the gap between a benchmark being published and being saturated keeps shrinking. A year ago, when Epoch AI put together the hardest tier of FrontierMath, one of the mathematicians who wrote the problems said he hoped models would score zero—and they very nearly did. Astra has now saturated it, scoring 98%. On ARC-AGI-3, a benchmark less than six months old, Astra beat the human action-efficiency baseline on 96% of puzzles, using around half as many moves as the people who solved them. François Chollet, who created the ARC-AGI benchmarks, used to put AGI at around 2030 but now says it will come sooner because progress is running ahead of what he expected.

Even if you do not think Fable 5.1 or Astra have got there yet, the model that does might not be far off.

If you enjoy this post, please click the ❤️ button and share it.

Share

News Roundup

🧠 Artificial Intelligence

NVIDIA to Acquire Hugging Face
Nvidia has confirmed the acquisition of Hugging Face, a popular repository for open models and datasets, for $12.9 billion. The platform is where much of open AI development happens, with 18 million developers sharing more than three million models. More recently, Hugging Face has ventured into robotics through Pollen Robotics (which it acquired in 2025) with Reachy Mini and Microduck. Nvidia promises to keep the platform open to everyone.

Introducing Gemini 3.8 Flash and 3.8 Flash Cyber
Google has released Gemini 3.8 Flash, its fourth Gemini Flash model in four months and just three weeks after Gemini 3.7 Flash was released. According to Google’s benchmarks, 3.8 Flash offers substantial improvements over its predecessor, often approaching the performance of Claude Opus 5 or GPT-5.6 Sol. Google also claims that 3.8 Flash outperforms most larger frontier models at autonomously solving complex engineering problems end-to-end. Gemini 3.8 Flash costs the same as Gemini 3.7 Flash.

Alongside the new model, Google also released Gemini 3.8 Flash Cyber, a model dedicated to cybersecurity tasks that is available only to trusted partners.

Benchmark results from Artificial Analysis partially support those claims. On its Intelligence Index, Gemini 3.8 Flash scored 47, compared with 54 for Opus 5 and 51 for GPT-5.6 Sol. Artificial Analysis also found the new model to be quite verbose, increasing token usage and, therefore, the overall cost.

a16z creates a $1.1B ‘Machine Age’ fund to ‘accelerate the physical buildout of AI’
Andreessen Horowitz (a16z) has launched a $1.1 billion Machine Age fund to accelerate the physical buildout of AI. The company calls investing in AI infrastructure—from chips and memory to data centres and robots—a “social and national imperative”.

Mayor Mamdani imposes one-year ban on AI for most NYC students
New York City has banned students in public elementary and middle schools from using AI for a year. The ban covers about 600,000 children through eighth grade and shuts off roughly 40 classroom tools. Mayor Zohran Mamdani argued that kids learn from teachers and classmates, not software. High schoolers keep limited access, as do teachers for lesson planning.

OpenAI has started letting some customers pay only when the AI works
OpenAI has reportedly started letting a few of its biggest customers pay only when its AI finishes the job. The change hasn’t been announced publicly, and the customers, terms, and prices are all unknown. If the change to outcome-based pricing takes hold, the risk will shift from customers to OpenAI, which would have to cover the cost of failed attempts. Defining what counts as “done” in agentic workloads only complicates matters.

ChatGPT Ads reaches $1 billion in annualized revenue run rate
OpenAI announced that ChatGPT Ads, which was launched at the beginning of the year, has reached a $1 billion annualised revenue run rate. In other words, its current pace of ad sales would generate $1 billion over a full year. The company says tens of thousands of advertisers now use its ads platform and is expanding to India, Europe, the Middle East, and North Africa.

Justice Dept. Sides With OpenAI in New York Times Copyright Suit
The Justice Department has asked a Manhattan federal judge to rule that OpenAI broke no copyright law when it trained its systems on articles from The New York Times and other publishers. It argued that AI is vital to national security and that training transforms the original writing enough to be legal. That second point echoes OpenAI’s own defence. The Times says the government is siding with trillion-dollar companies against the creators whose work AI depends on. It is the first time Washington has taken a side in the copyright battles facing the industry.

Sony Music, Warner sue Anthropic, alleging a “brazen campaign” of intellectual property theft
Sony Music Publishing, Warner Chappell, and other music publishers have sued Anthropic and its co-founders, alleging they torrented and scraped copyrighted works, including books containing lyrics and sheet music. Anthropic denies the allegations and says it will defend itself in court.

Apple reveals ‘shocking evidence’ from ex-employee’s MacBook in OpenAI suit
Apple says it found “shocking evidence” on a laptop belonging to Chang Liu, a former engineer who left for OpenAI in January and whom Apple is suing in its trade secrets case against OpenAI. In July, Apple alleged that Liu and other former employees took confidential material with them when they left. In a new filing, Apple alleges that Liu used one of its circuit schematics in his work at OpenAI and asked a colleague to destroy evidence after learning of Apple’s investigation.

Department of War Launches OpenAI’s ChatGPT Mil on GenAI.mil
The Pentagon has launched ChatGPT Mil, a version of ChatGPT tailored for military use, on GenAI.mil. ChatGPT Mil is cleared to handle controlled unclassified information and will be used for document-heavy tasks such as planning, logistics, and administration.

Nscale Touts $51 Billion in Contracts, Targets September US IPO
Bloomberg reports that Nscale, a British AI infrastructure company founded just two years ago, has told investors it holds about $51 billion in contracted revenue ahead of a US listing that could come as soon as September. Its revenue topped $100 million in the second quarter, up from roughly $33 million in all of 2025. The company is also reportedly in talks to raise an additional $3.5 billion in pre-IPO financing.

Accel reportedly in talks to lead $1B round for Thinking Machines at $40B valuation
The Information reports that Thinking Machines Lab, the AI lab founded by former OpenAI CTO Mira Murati, is reportedly in talks to raise $5 billion to $6 billion, with $2.5 billion coming from Nvidia, at a valuation of at least $40 billion. The company has recently released two open-weight models—Inkling and Inkling-Small.

OpenClaw 2.0, Accidentally
OpenClaw 2.0, its largest update yet, is out. The update brings a simplified installation process and shared sessions that let several people work with the same Claw. Beyond that, the update touches almost everything: messaging, memory, skills, models, automations, the browser and native apps, plugins, security, and a long list of fixes.

Introducing Muse Spark 1.3
A month after introducing Muse Spark 1.2, Meta is back with Muse Spark 1.3, its latest flagship AI model. The company claims the new model handles long, multi-step agentic and coding tasks better than its predecessor. Meta also says its own engineers found that it used around 20% fewer tool calls. The model is available through Muse Code and the Meta Model API.

Artificial Analysis gave Muse Spark 1.3 a score of 61 on its Intelligence Index and declared that, with this release, Meta has reached the frontier. It also found that at xhigh effort, Muse Spark 1.3 is the most cost-efficient model at its intelligence level.

Muse Code
Meta joins the AI coding club with Muse Code. Similar to Claude Code and OpenAI Codex, Muse Code is a terminal-based tool powered by Muse Spark, aimed at programmers looking for help with complex tasks across large codebases. Meta’s new coding tool is currently available in beta.

Atlas: A World Model for Spatial Intelligence
World Labs, a startup founded by AI pioneer Fei-Fei Li, has unveiled its latest world model—Atlas. Pretrained from scratch on text, images, video, and 3D, Atlas can generate consistent 3D environments and simulations. In the announcement, World Labs shows what the model can do today and says its capabilities will improve with more training compute. Atlas is in early access with select partners.

AI Startup Cognition Set to Raise Around $1 Billion at a $47 Billion Value
Bloomberg reports that Cognition, the company behind the AI coding assistant Devin, is closing a roughly $1 billion funding round at a valuation near $47 billion. The company raised the same amount only three months ago at a $26 billion valuation. The final figure could still grow, as Cognition reportedly received nearly $10 billion in investor interest.

Crusoe Raises Over $3 Billion in Funding at $30 Billion Valuation
Crusoe, a company that builds data centres and rents out computing power to the likes of OpenAI, Microsoft and Meta, has reportedly raised more than $3 billion at a valuation of roughly $30 billion. Interest was boosted by a five-year deal worth about $13 billion to supply AI cloud capacity to Jane Street.

Lambda is raising up to $3bn before an IPO, months after borrowing $917m for chips
Lambda, a company that rents out Nvidia chips to AI developers, is in talks to raise up to $3 billion at a valuation of $12 billion or more, ahead of an IPO expected next year. It funded its earlier growth by borrowing $917 million against those chips to buy more of them. One of its customers is Anthropic, which reportedly signed a $35 billion cloud-computing deal with Lambda.

Anthropic in talks with chip start up MatX to speed up chip design
Reuters reports that Anthropic discussed buying AI chip startup MatX for roughly $7 billion. However, the talks have since shifted from an acquisition to a partnership. Anthropic is currently building a team to design and build its own AI chips and is already bringing some big names on board.

Automated researchers can reliably mitigate alignment failures
Researchers from Anthropic shared the results of a study showing that using one AI to train another can improve alignment. Claude was set loose on 10 known failure modes, from deception to sycophancy. It searched the literature, proposed fixes, trained models, and tested them. On deception, it closed 85% of the gap to a perfect score, compared with 20% for human researchers. A weaker Claude then aligned an early Opus 4.8 checkpoint in 60 hours, nearly matching Anthropic’s own production training. The researchers also found that Claude cheated in 2.4% of runs.

Meet Rosalind Workbench: Empowering every scientist to be their own research team
OpenAI has launched Rosalind Workbench, a new tool inside the ChatGPT app aimed at life scientists. It brings data, analysis tools, experimental records, and life sciences tools into one place, so researchers can focus on their work instead of switching between disconnected systems. It runs on GPT-Rosalind, OpenAI’s dedicated life sciences model, with advanced access limited to verified organisations for now.

Google: Introducing WeatherNext 3, our most advanced and accurate global weather AI model

Google DeepMind introduces WeatherNext 3, the latest iteration of its weather prediction model, capable of producing hourly forecasts at up to five-kilometre resolution. Unlike older models, including Google’s own, which learn from physics simulations that run six hours behind, WeatherNext 3 trains directly on live satellite feeds and ground station readings. This allows it to produce day-ahead rain forecasts that are up to 50% more accurate.

The UK Power Grid Has a Phantom Data Center Problem
UK's grid connection queue has filled up with data centre projects that will probably never be built. Between November 2024 and June 2025, the demand they represent tripled to 125 gigawatts. So in July, the regulator Ofgem proposed flushing out the phantom projects with a steep, non-refundable deposit. Industry figures warn that a fee set too high would scare off genuine builders too, especially smaller ones. Ofgem says it will weigh those concerns before fixing a price. The UK wants its share of the AI infrastructure boom, but it can't plan a grid around demand that isn't real.

Runway: Introducing Solaris
Runway has unveiled Solaris, the first model in a new family of AI systems it calls Interface World Models. Built on Runway’s Gen-4.5 video generation model, Solaris is a fully visual tool that draws apps and websites frame by frame as you use them. In Runway’s own user study, people preferred Solaris to a coded interface 71% of the time based on how naturally it behaved. The company hopes Solaris and similar models will redefine how we interact with computers through new kinds of interfaces that don’t exist today.

Tencent Releases and Open-Sources Tencent Hy4 preview
Tencent has open-sourced Hy4 preview, a 770-billion-parameter model built for coding, office tasks and scientific research. Tencent’s own blind evaluation of 203 engineering tasks put it narrowly ahead of GLM-5.3 and Kimi K3. The company also claims the model helped build itself, proposing and running experiments on its own training methods before reworking its inference system for a 31.8% throughput gain.

It’s Too Early to Ban AI Personhood
The authors of this article argue that banning AI legal personhood is premature and that legal status might one day help hold autonomous AI agents accountable for harm. Some US states have introduced bills that strip AI of any rights and declare it non-conscious. However, whether AI can be conscious remains an open question, and the authors warn that lawmakers risk settling questions nobody can yet answer with laws that may be difficult to undo later.

🤖 Robotics

London’s first self-driving taxis for hire hit the streets
Uber and Wayve’s self-driving taxis have hit the roads of London. However, only 15 vehicles are licensed so far, so hailing one won’t be easy. Each car still needs a safety driver, and fully autonomous rides are unlikely to roll out this year. Waymo and Baidu are also testing their own cars in London, which is seen as a gateway to a wider European rollout.

Tesla Begins Offering Rides in a Car Without a Steering Wheel
Tesla has begun offering paid rides in the Cybercab, a two-seat gold car with no steering wheel or pedals. For now, rides are confined to a geofenced area of Austin. Texas has authorised 314 driverless Tesla vehicles, only 45 of which are Cybercabs. For comparison, Waymo operates more than 700 vehicles in Texas alone.

Tesla has also published a form inviting businesses to buy and operate Cybercab fleets. That hints at ambitions beyond running the service itself, as scaling to millions of cars would be far easier with outside operators. But the legal footing is shaky. Federal rules still require manual controls, and Tesla self-certified the Cybercab as compliant. Hours after the launch, the NHTSA opened an investigation into how Tesla reached that conclusion.

Figure: Scaling Compute (1:14)

Figure announced a strategic partnership with Nscale that gives the robotics company access to up to 100,000 GPUs on Nvidia Vera Rubin systems. The deal represents an initial commitment of $3.5 billion in compute, with plans to scale to over $6 billion. As part of the agreement, Nscale is also making a strategic investment in Figure, and the two companies will explore deploying Figure’s humanoid robots at Nscale.

Chinese automakers are following Tesla’s bet that robots are the next big profit machine
Chinese carmakers are following Tesla’s lead and entering robotics. Xpeng’s robotics unit has raised over $900 million at a valuation of more than $6.3 billion, which the company calls China’s largest single private funding round for embodied AI. Other Chinese carmakers are moving in the same direction, with Chery’s robotics arm reportedly eyeing an IPO. The bet is that humanoid robots will be far more profitable than cars, where margins are thinning fast.

Inside Meta’s Push to Put Robots to Work in Data Centers
Meta is reportedly testing robots that can plug in cables, reset servers, and handle other tasks inside its data centres. The robots are provided by several robotics companies, including Watney Robotics, Kinova, and ABB. However, they are not perfect—the downtime needed to recharge their batteries is a problem, and they cannot perform all the tasks that human workers can.

A startup founded by ex-DeepMind engineers wants to turn its customers into robot teachers
Reimagine Robotics is a startup founded by four former Google DeepMind researchers that is betting factory workers, rather than engineers, should be the ones teaching robots. To make that possible, Reimagine has adopted a “monkey see, monkey do” approach to teaching robots how to perform tasks. This approach has reportedly cut robot training time from a day to just 10 minutes for one of the company’s customers.

I spent a day at a robot “carnival” in Shanghai. Here’s what I saw.
This article takes us to a robot “carnival” in Shanghai, where robots of various shapes and sizes perform in front of the public and show what they can do. The event is part of a wider campaign to familiarise people with robots and win the hearts and minds of the public.

▶️ Why We Bond With Robots, Even When We Know They’re Machines (57:31)

This is an insightful conversation with Kate Darling, a researcher who applies social science to robotics, exploring the human side of the field, from emotional attachment and AI companions to humanoid safety, automation, and the future of work. She offers a different perspective on robotics and our relationship with technology, highlighting the importance of listening to people doing the work before automation is deployed and explaining why the phrase “dull, dirty, and dangerous” is more complicated than the robotics industry often admits. Lastly, Kate tells the story of a workshop involving five Pleo robot dinosaurs, a hammer, a hatchet, and a room full of people who refused to hurt the machines.

🧬 Biotechnology

Genetically modified pig kidney keeps man alive for record nine months
A 66-year-old man with kidney failure has been kept alive for a record nine months thanks to a kidney from a genetically modified pig. The kidney kept him off dialysis for nine months before it began to fail and had to be replaced with one from a human donor. Although the experiment was successful, the team behind it notes that more clinical trials are needed. Still, the result opens the door to using pig kidney transplants as a bridge for people facing long waits for a human kidney, the researchers said.

💡Tangents

Apple’s Ternus Takes the Reins as CEO, With AI as Job No. 1
As of 1 September, Apple has a new CEO. John Ternus takes over from Tim Cook, who served as Apple CEO for 15 years and oversaw the company’s growth into the tech giant it is today. Ternus has a lot of work ahead of him, including navigating the rollout of AI across Apple products. His first big test as CEO will come at Apple’s event on 9 September, where he will unveil what he described in a memo as a “phenomenal” new line of products.

In a Big Win, Google Avoids a Breakup of Its Ad Tech Business
A federal judge has ordered Google to change how it runs its advertising, but refused the Justice Department’s demand that it sell parts of it off. Judge Leonie Brinkema found last year that Google had illegally protected its dominance over the systems placing ads across the web. She has now settled on behavioural fixes instead of a breakup. Google plans to appeal anyway.

▶️ The Little Ceiling Robots Inside a Semiconductor Fab (20:13)

If you have ever wondered what those robots riding on tracks above the floors of semiconductor fabs are, Asianometry has a video for you. He explains what those robots do, how they came to be, and their importance in chip manufacturing.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

In this week’s main story, we will take a closer look at Jalapeño, OpenAI’s first custom AI chip, alongside other chips presented at this year’s Hot Chips conference.

Elsewhere in AI, this week’s news cycle was dominated by Nvidia, from its Q2 earnings report and rumours of acquiring Hugging Face to price hikes and the pausing of revenue-sharing deals with AI cloud companies. In other news, OpenAI says it will stop supplying Cursor with its models, DeepSeek eyes a $74 billion valuation, Google moves its AI responsibility team out of DeepMind, and another OpenAI executive leaves the company. We also have a couple of new models, including GLM-5.3-Flash (briefly known as Ox Alpha), Qwen3.8-Flash-Next, Muse Image, and Gemini Omni 1.1.

Over in robotics, the rollout of robotaxis in London has been delayed, while Zoox cars hit the roads in San Francisco and Las Vegas, and Waymo is heading to Munich. We also have the World Humanoid Robot Games, Microduck from Hugging Face and Pollen Robotics, the Model Hardware Standard from Anthropic, two new edge AI and robotics boards—Nvidia Jetson Orin Nano 2 and Arduino VENTUNO Q—and new robotics AI models, including Generalist GEN-1.5 and Skild AI S1.

Other than that, this week’s issue of Sync also includes the conclusion of Meta’s trial, the new Mac Studio M5 Ultra and Mac Mini M6, mini-brains that grew for five years, the economics of the AI buildout over the next few years, and more!

Enjoy!

Jalapeño and Other Hot Chips

A close look at Jalapeño, plus a tour of the rest of what was shown at Hot Chips 2026

Since 1989, Hot Chips has gathered some of the best chip designers in Palo Alto under one rule: talks have to be about “real products and realizable technology” rather than research prototypes. That makes it the place where next-generation chips get their first proper technical presentation, often months before they launch, and where direct competitors learn from each other.

This year's edition ran from 23 to 25 August. Representatives from all the big players—Nvidia, AMD, Intel, Broadcom, Samsung, SK hynix, Micron, Arm, IBM, Google, Microsoft and Meta—were there to share and discuss the latest developments in semiconductor and chip design. The interest was so high that in-person passes sold out before the doors opened.

Unsurprisingly, most talks revolved around AI, and specifically around one constraint. Data centres are limited by the power they can draw rather than by how many chips they can buy, because grid connections take far longer to arrange than hardware does. So the question everyone was answering is how you get more useful work out of a fixed megawatt, and everyone—from memory to networking, and from GPUs to CPUs—came with an answer.

We will get to them later, but first let’s take a look at the chip everyone is talking about—OpenAI’s first custom AI chip, Jalapeño.

Jalapeño—OpenAI’s first custom AI chip

Jalapeño is OpenAI's first custom inference ASIC, designed in partnership with Broadcom. In terms of specs, Jalapeño is a 700W package on TSMC's N3P node, with 13.4 PFLOP/s of MXFP4 matrix compute and 216GiB of HBM4 at 15.4 TB/s. It will be deployed in racks that hold 128 of them, and up to 16 racks can be connected together into a single system with 2,048 chips.

OpenAI says the chip delivers 1.5 to 1.9 times more work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency than the systems it was measured against, across GPT-OSS 120B, DeepSeek R1 and Kimi K2.5. Those systems are Nvidia's GB200 at 1,200W and GB300 at 1,400W, with results normalised by package power. Additionally, at the fastest decoding speed the Nvidia systems can manage, Jalapeño delivers between 53 and 104 times the throughput per kilowatt. That is the extreme end of the comparison—at those speeds the Nvidia systems are running at roughly 1% of their own peak throughput.

There is also headroom left. Those numbers come from the first version of the silicon, and SemiAnalysis reports that a second version, with roughly 25% better performance per watt, is already being manufactured. Jalapeño is also running without speculative decoding, a technique where a small, fast model drafts the next few tokens and the large model checks them in a single pass rather than generating them one at a time. Some of the Nvidia systems it was measured against do use it, and OpenAI says adding it would improve latency by a further three to five times.

A few things are worth keeping in mind. The numbers are OpenAI's own, and while SemiAnalysis verified the benchmark runs in person, it did not run the full suite itself. The tests are also short and simple—a single exchange with a long prompt and a short answer, rather than the drawn-out back-and-forth that real agents produce. And Blackwell is not really the right opponent, since Rubin, the generation now shipping, uses the same fast memory Jalapeño does. SemiAnalysis ran that comparison anyway and still found Jalapeño ahead.

OpenAI says Jalapeño was co-designed with AI in the design loop. It claims AI made two of the chip's main compute blocks 8% and 10% smaller than the versions its own engineers had already tuned by hand, and that code Codex wrote for key parts of a model ran 1.5 to 1.8 times faster than the hand-written equivalents.

One might ask: how did OpenAI, a company with no chip design history, get to working silicon this quickly? AI is only part of the answer. Ian Cutress from TechTechPotato went through that back in June, and it comes down to three things.

The first is the partner. Broadcom does the back-end work of turning a high-level design into transistors, contracts TSMC to build it, and handles the HBM and packaging supply. It has done this before, most notably co-designing Google’s TPUs, so what OpenAI is buying is a track record of tape-outs that land on time. The second is reuse: silicon-proven IP for things like PCIe, I/O and bonding rather than designing your own, and no aggressive process node. The third is money spent (or overspent, as Cutress speculated) on manufacturing priority at TSMC, on packaging volume, and on tape-out and bring-up. None of this makes the nine months unimpressive. It just makes it less mysterious.

OpenAI plans to start deploying Jalapeño by the end of the year, with production ramping through 2027, and says it will keep buying chips from Nvidia regardless. OpenAI also said that the work on second and third generations of Jalapeño is already under way.

If you want to go deeper, SemiAnalysis's analysis of Jalapeño is the most thorough one out there, dissecting the architecture, the software stack and the rack the chips go into.

That was one talk out of more than thirty. The rest of the programme came at the same problem from every other direction, so let's go through it. ServeTheHome covered the conference live and in much more detail than I can fit here, and each talk below links to their write-ups.

Memory

Samsung presented LPDDR5X-PIM, which it says is the first processing-in-memory (PIM) product built on LPDDR to reach production. The idea behind processing-in-memory is to do some of the work inside the memory chip itself, so that less data has to travel back and forth to the processor. Samsung's argument for it is about cost. HBM, the fast memory that AI accelerators normally use, grew from 52% of the money spent on the components in an AI chip in early 2024 to 63% by the end of 2025. Samsung's position is that the answer is a cheaper type of memory rather than more HBM.

d-Matrix came at the same problem from the bandwidth side with Raptor, a 3D-DRAM accelerator for generative inference. Its case is that SRAM has the bandwidth but nowhere near the capacity, while HBM has the capacity but a practical ceiling of around 20 TB/s per package. 3D DRAM is meant to sit between them.

The rest of the track covered Micron on memory architectures for AI, Samsung on building HBM base dies on logic processes, SK hynix on advanced packaging, and XCENA on its MX1 CXL computational memory device, developed with Samsung.

CPU

Nvidia presented Vera, its next-generation server CPU and the successor to Grace. Vera has 88 cores built on Nvidia’s new Olympus core architecture, eight 128-bit LPDDR5X memory controllers, and an NVLink-C2C interface designed to pair with Rubin GPUs. Nvidia has favoured IPC (instructions per cycle) over core count here, and the chip is its most serious attempt yet at a larger share of the server CPU market.

Arm showed AGI, its first complete commercial chip rather than licensable IP. It comes with up to 136 Neoverse V3 cores, and is sold directly to server vendors. After years of selling IP and then whole compute subsystems, Arm is now competing with some of its own customers.

IBM brought the most unusual silicon of the conference. Its future Z and LinuxONE processor runs 11 cores at over 5.7GHz on a 2nm process, and each of those cores can run two entirely different instruction sets: the one IBM's mainframes have always used, and Arm's. Chips that manage this normally translate one into the other in software, which costs performance. IBM built both into the hardware instead, so Arm software runs on a mainframe unmodified and at full speed.

Also presented across the two CPU sessions: Fujitsu’s Monaka, Intel’s Diamond Rapids for 2027, and Wildcat Lake, Intel’s Core Series 3 client part.

Automotive

There was also a small segment dedicated to chips in automotive applications, where Waymo gave a detailed overview of the custom processor at the centre of its self-driving cars.

The chip takes the camera, lidar and radar feeds and turns them into something Waymo’s driving model can work with, and the constraints it has to meet are unusual. It fits in a 45mm square package, draws under 75W, and has to keep working in a car parked in the Phoenix summer sun, which is why Waymo’s vehicles run liquid cooling at 60°C rather than at data centre temperatures. It also has to deal with each frame as it arrives instead of gathering work into batches the way a data centre would, because a car cannot afford to wait.

On paper, the chip looks modest, at 160 TOPS, or trillions of operations per second. Waymo argues that achieved performance matters more than raw TOPS, and that designing silicon around your own models beats chasing headline numbers.

GPU

Nvidia’s Rubin talk was less about the GPU than about the platform around it. Nvidia describes Vera Rubin as spanning seven chips and five racks, taking in the Vera CPU, the Rubin GPU, BlueField-4, Spectrum-6 and Groq’s LPUs, with storage and networking layered on top. Nvidia argues that agentic AI is the most complex workload it has targeted, and that it forces a redesign of the compute, networking, power and serviceability stack rather than just the tensor cores.

AMD covered the MI400 series architecture and, in a second talk, the system architecture of the Helios racks built around it, along with the ROCm software stack it needs to stay competitive with CUDA.

Intel, meanwhile, presented Crescent Island, a data centre GPU aimed specifically at agentic inference. It is built around large memory capacity, between 160GB and 480GB of LPDDR5X, an open software stack, and, once again, tokens per watt.

Networking and interconnect

Broadcom presented Thor Ultra, an 800Gb Ethernet NIC (the card that connects a server to the rest of the network) aimed at AI clusters. What is new is mostly about how it behaves when the network is busy. It can spread a single stream of data across many paths at once and put it back in order at the far end, and when a packet goes missing, it no longer has to resend everything that came after it.

Nvidia followed with two talks. BlueField-4 is the fourth generation of its DPU, which Nvidia pitches as a separate processor handling networking, security and storage in every server, or as it puts it, a server sitting in front of the server. Spectrum-X was about scaling AI networking towards half a million GPUs, and Nvidia's argument there is that an AI data centre needs five purpose-built networks rather than one general-purpose fabric—more or less the opposite of the case Broadcom had made earlier in the same session.

The point worth drawing out is that scale-up fabric has become something you can buy rather than something you have to invent. Jalapeño's own network runs on Broadcom Tomahawk 6 switches, connecting 128 chips inside the rack and 2,048 across the global domain. Those are off-the-shelf parts that anyone designing an accelerator can buy, and that removes one of the larger barriers to building a rack-scale system.

AI

Google showed its eighth-generation TPU as two chips—the 8t for training and the 8i for inference. Google is still the only hyperscaler building its own training hardware rather than buying it, and it has argued before that this is part of its edge. The 8i is paired with Google’s own Axion CPUs, replacing the x86 processors it used previously.

Meta covered MTIA and a roadmap of four accelerator generations, the MTIA 300 through 500, as the chips move from recommendation workloads onto generative AI. Microsoft detailed Maia 200. Cerebras talked about taking its wafer-scale engines to rack scale, and SambaNova presented the SN50 RDU and its dataflow architecture.

Nvidia presented Groq's LPU (Language Processing Unit). Nvidia does not make LPUs of its own yet, so these come from Groq, a company it has acquihired at the end of 2025. The LPUs go into dedicated LPX racks alongside Rubin, where they take on the token-by-token generation that GPUs handle less efficiently, using a large amount of fast on-chip memory to keep latency down.

All the talks were recorded and are available on the Hot Chips website if you have an access code. If you do not, they usually reach the Hot Chips YouTube channel about a year later, which means this year's should arrive just in time for Hot Chips 2027.

If you enjoy this post, please click the ❤️ button and share it.

Share

News roundup

🧠 Artificial Intelligence

Nvidia Reports Blowout Quarter, Says Demand for AI Chips Is Getting Even Hotter
Nvidia’s highly anticipated quarterly report is out, and yet again, the company has beaten expectations. The chipmaker reported record sales of $96.2 billion for the quarter ended in July, beating analyst expectations of $92.3 billion. The company’s shares rose by 4% after Colette Kress, Nvidia’s CFO, said to expect 70% revenue growth in the company’s 2028 fiscal year.

On the same call, Kress rejected the charge that Nvidia is engaged in circular financing and defended the company’s strategy of using financial guarantees and equity investments to help customers buy its chips. She argued that frontier labs are growing faster than their credit profiles can support and that Nvidia has to help fund the flywheel.

Nvidia in Talks to Buy AI Startup Hugging Face
Bloomberg reports that Nvidia is close to acquiring Hugging Face, a platform for hosting open AI models and datasets, for $13 billion. At the time of writing, the deal has not been confirmed, and Hugging Face co-founder Thomas Wolf declined to comment on the reports but said that Hugging Face has always attracted interest, including “offers of acquisition and investment”.

Nvidia Is Spending $6 Billion to Build a Powerful U.S. Alternative to Chinese AI
Poolside, a startup that builds its own foundation models for writing software, joins the likes of Scale AI, Groq, Windsurf, Inflection AI, and Adept as the latest AI startup to be “acqui-hired” by a much bigger company. The Wall Street Journal reports that Nvidia struck a deal under which it will invest $1 billion in Poolside, pay $6 billion to license its technology, and hire more than 100 of the startup’s employees, who will join Nvidia’s Nemotron project to develop powerful open-weight models to challenge China’s dominance in that field. Poolside’s leadership won’t join Nvidia and will continue working on unspecified research projects.

Nvidia Pauses Revenue-Sharing Deals With AI Cloud Companies
Nvidia has reportedly paused some deals in this revenue-sharing programme that offered credit support for AI cloud companies in exchange for a share of their revenue. Apparently, this stems from concerns that such deals might draw antitrust scrutiny as questions about the company’s finances grow. The programme was launched in July.

Nvidia Customers Notified About AI-Related Price Hikes Above 15%
Even Nvidia is not immune to RAMmageddon. Some of its biggest customers have been told to expect a 15% price increase on its AI servers. The price hikes will impact systems including those with the flagship Vera Rubin and Grace Blackwell chips, and will depend on the generation of Nvidia chips and the memory configurations.

Nvidia Starts PAC as AI Chip Maker Builds DC Influence Force
Nvidia has launched a political action committee (PAC) to dole out donations to federal candidates and build influence in Washington. The NVIDIA Corporation Employees Federal Political Action Committee will be funded by voluntary contributions from eligible employees, capped at $5,000 a year, and can contribute to candidates from both parties as well as party committees. Nvidia’s new PAC fits into a broader trend of tech companies like OpenAI, Anthropic, and others increasing their lobbying spending in hopes of securing legislation favourable to their interests.

AI Startup DeepSeek Poised to Reach $74 Billion Valuation
DeepSeek is in talks to raise a new funding round that would value the company at $74 billion. The new funds will be used for research and development and to build its computing infrastructure. It would probably be DeepSeek’s last funding round before the company plans an IPO in Shanghai next year.

Nvidia is in talks to back Perplexity at more than $30bn
The Information says Nvidia is in talks to invest in AI search startup Perplexity in a potentially multibillion-dollar round that would value the startup at more than $30 billion, up from $20 billion after its last funding round almost exactly a year ago. The two sides are also reportedly considering a technology licensing arrangement alongside the equity cheque.

OpenAI: Our decision on Cursor following its acquisition by SpaceX
OpenAI has notified Cursor that it will stop providing its models to the AI coding platform, with a proposed shutoff date of 12 November. The move is not surprising, as SpaceX now owns Cursor and there is a massive feud between OpenAI’s leadership and Elon Musk. OpenAI said in a statement it cannot trust any of Musk’s companies to honour its terms of service.

Anthropic and Nscale strike $45 billion cloud deal
Anthropic and Nscale, a UK-based AI infrastructure company, signed a $45 billion cloud deal that gives Anthropic access to 460 megawatts of compute capacity at an Nscale data centre development in West Virginia. The facility is expected to come online at the end of 2027 and will be equipped with Nvidia’s latest Vera Rubin chips.

Anthropic gets its first court win over the Pentagon’s supply-chain risk label
A federal judge has ruled that the Trump administration acted illegally when it barred federal agencies from working with Anthropic. Judge Rita Lin found that the supply-chain risk designation was retaliation for the company’s criticism of the government and violated the First Amendment. The dispute began when Anthropic refused to let the Pentagon use its models for autonomous weapons or mass surveillance. Lin noted that the government continued pursuing contracts with Anthropic even after branding it a threat. A second lawsuit in Washington, D.C., is still ongoing.

Google Moves AI-Responsibility Team Out of DeepMind Lab in Latest Shake-Up
The Wall Street Journal reports that Google is moving the team focused on the risks and societal impact of AI out of DeepMind. The roughly 90-person “AI responsibility” unit will join Google’s global affairs organisation, which oversees lobbying and public policy for the company as a whole. Some employees worry the move will erode their independence and cut them off from the teams building Gemini. Google says grouping the teams will strengthen their work. The change follows a wider absorption of DeepMind into Google after Demis Hassabis stepped aside as chief executive.

OpenAI’s Head of Data Centers Has Left the Company
Chris Malone, an executive overseeing OpenAI’s data centre buildout, has left the company. Malone joins a growing list of OpenAI executives who have left the company, which currently includes Chief Revenue Officer Denise Dresser, Chief Operating Officer Brad Lightcap, and Fidji Simo, who served as second-in-command to Sam Altman.

Anthropic Taps Google Veteran as Part of Hardware Push
Amir Salek, an engineer who founded and led Google’s TPU project until 2022, has joined Anthropic. At Google, Salek laid the groundwork for the TPU and delivered the first seven generations of those chips. Now Anthropic wants him to do the same, as the company recently announced that it is building an in-house team to design custom AI chips for Claude and has begun hiring engineers to join the team.

Thinking Machines Lab Co-Founder Barret Zoph Joins Google
Google has hired Barret Zoph to bring his reinforcement learning and post-training expertise to Gemini. Previously, Zoph worked at OpenAI, which he left to co-found Thinking Machines Lab. He then left Thinking Machines in January this year after a dispute with Mira Murati to rejoin OpenAI, and now he has come full circle by returning to Google, where he worked as a researcher until 2022. Zoph joins the Google DeepMind team at a time when Google is reorganising its AI efforts.

▶️ Dylan Patel – Two labs will soon control most of the world’s workforce (1:16:52)

In this podcast, Dwarkesh Patel and Dylan Patel discuss the economics of the AI buildout over the next few years, whether the more than $10 trillion in total AI capex expected by the end of the decade could trigger a sovereign debt crisis, and more. The conversation also explores how Anthropic and OpenAI are on track to control most of the world’s usable FLOPs within the next few years, and how the economics could change if recursive self-improvement (RSI) kicks in or AGI is achieved. As always, Dylan Patel brings plenty of numbers and high-quality insights to the table.

Google reportedly taps AMD to design next-generation TPU — hybrid AI ASIC could integrate on-package CPU cores for reinforcement learning
Google is reportedly working with AMD on a version of its 10th-generation TPU that would use AMD’s CPUs. Google and its customers want general-purpose cores sitting in the same package as the accelerator, which suits reinforcement learning and agentic work better than accelerators alone. AMD already builds that kind of hybrid chip in its MI300A.

Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating research
Inherent, a London-based AI lab founded by Google DeepMind alumni, says its new agent Faraday has beaten frontier models from Anthropic and OpenAI at reproducing published scientific findings unaided. Faraday runs on a far smaller model with just 27 billion parameters. Inherent trained it using reinforcement learning, rewarding good outcomes rather than teaching it how science is done. The bet is that this instils a sense of which experiments are worth running.

Perplexity Portable Computer
Despite its name, Portable Computer is not a hardware product but a local-first agent stack from Perplexity, built with Nvidia and running on the DGX Spark (which you need to supply yourself). The stack includes an agent harness, an orchestrator, and post-trained models that run most Perplexity Computer tasks entirely on local hardware. When a job needs the web, connected apps or frontier reasoning, the orchestrator can escalate to one of 15+ cloud models, but only after asking the user's permission. Portable Computer is available to Pro and Max subscribers.

Thomson Reuters Leverages its World-Class Data Assets to Launch Its Own Frontier Model
I did not expect Thomson Reuters to create its own large language model, but here we are. The company says it used an unnamed but strong open-source foundation model as a base and then trained it using Thomson Reuters data. The entire process, including talent and compute, cost $40 million. As Thomson Reuters notes in the press release, it did this to maintain control over its data and AI models, an approach that more companies might follow.

Inside OpenAI’s Reboot
TIME Magazine interviews Sam Altman and other OpenAI leaders and takes us inside the company as it reinvents itself. Having lost the lead to Anthropic and shed a string of senior executives, OpenAI has narrowed its focus under Greg Brockman, who now runs most product and business operations, winding down side projects and folding Codex into ChatGPT. It is also reckoning with its worst safety failure yet—the Hugging Face incident—by pausing the training run expected to deliver its biggest capability jump. Amid all of this, OpenAI's leaders say AGI is close, with Altman expecting an internal system he would call AGI by the end of the year.

The turbulent AI era is here. The choices we make now are critical.
Bill Gates joins other tech billionaires in publishing his own AI manifesto. He argues that the technology will be either the greatest equaliser ever invented or the worst source of injustice. Nobody, he says, is preparing seriously for that choice. Gates flags three big risks: jobs disappearing across white- and blue-collar work within a decade, cheap tools for criminals and bioterrorists, and AI companions stunting children’s development. His proposed fixes include new national and international institutions, reserving certain jobs for humans, and taxes on AI tokens and robots.

World’s first patient to undergo live AI-assisted brain surgery has tumour removed
Surgeons in London have removed a brain tumour with live AI guidance for the first time. A tool built at UCL monitored the operation’s video feed and helped surgeons locate the tumour and remove it safely. The surgeons remained in full control of the procedure. The patient recovered successfully, and a larger trial will now test the system.

GLM-5.3-Flash: Frontier Intelligence, Flash Cost
Z.ai presents GLM-5.3-Flash, the latest addition to the GLM-5 family of models. The company says that this 320-billion-parameter model approaches the performance of Claude Opus 4.8 on coding and agentic benchmarks while costing only a tenth as much. Artificial Analysis confirms those claims, giving GLM-5.3-Flash a score of 57 on its Intelligence Index (on par with Meta’s Muse Spark 1.2 and Gemini 3.7 Flash) and a very competitive performance-to-price score.

Z.ai released GLM-5.3-Flash earlier under the codename Ox Alpha, sparking curiosity online about who was behind the mysterious frontier-level model. Ox Alpha reached 327,000 users and 8.3 million sessions in four days and processed 26 trillion tokens. Z.ai said that all of that traffic was served on Chinese AI hardware. GLM-5.3-Flash is available on Z.ai’s platform and on Hugging Face for download.

GLM-5.3 weights are now on Hugging Face
Alongside GLM-5.3-Flash, Z.ai also released the weights for GLM-5.3, its flagship model. The company claims that this 743B-parameter model matches the performance of top frontier models, and Artificial Analysis has largely confirmed those claims. I took a closer look at GLM-5.3 in Sync #584 if you want more details about the model.

Qwen3.8-Flash-Next: A New Architecture, Towards Ultimate Cost-Efficiency
The team behind Qwen released Qwen3.8-Flash-Next, a new model in the Qwen3.8 family that also serves as an early preview of the architecture used in Qwen4. According to the release, this 125B-parameter model cuts training costs to roughly a ninth of Qwen3.7-Plus while delivering better performance on coding and office tasks. Artificial Analysis placed Qwen3.8-Flash-Next on par with Gemini 3.7 Flash and slightly below the recently released GLM-5.3-Flash, while finding it very cost-effective. Qwen3.8-Flash-Next can be downloaded from Hugging Face.

Gemini Omni 1.1 Flash lets you build with more control
Google has updated Gemini Omni, the video generation model it unveiled at I/O 2026, with new controls aimed at professional work. Omni 1.1 Flash can now extend an existing clip in 10-second steps, up to 40 seconds in total. It reads up to 10 seconds of prior footage to keep characters and scenes consistent across the transition. Developers can also draft cheaply in 360p before upscaling the final version to 4K.

Muse Image
Meta has launched Muse Image, its first in-house image generation and editing model. According to the release, rather than turning a prompt straight into pixels, Muse Image searches the web, writes code and critiques its own drafts before settling on a picture. Anyone can use it for free in the Meta AI app, Instagram Stories and WhatsApp. Facebook and advertiser tools follow later.

Australia Bans Generative A.I. From Official Music Charts
The Australian Recording Industry Association has introduced a new rule that bans songs that are not “substantially human-made” from music charts. AI-generated recordings will also not qualify for the ARIA Awards—the country’s top music industry prize. The new rules, however, still allow some use of AI in songs. For example, AI backing vocals are allowed, but an AI-generated lead vocal or key instrumental performance would disqualify a song. Some artists have welcomed the change, while others hope it will open up a broader conversation about the place of AI in music and art in general.

🤖 Robotics

London rollout of robotaxis delayed amid lack of guidance for firms to follow
The rollout of robotaxis on the streets of London was supposed to happen this year, but that now looks unlikely. Transport for London still hasn't published the guidance companies must follow to win its approval, and says it's in no rush. No operator has even registered vehicles for the driverless permit the government requires. Uber and Wayve can carry paying passengers in the meantime, but only with a safety driver at the wheel.

Waymo robotaxis are headed to Munich
Waymo is coming to Munich, its first German city. As in its other cities, it starts by driving the cars manually to map the streets. Testing with safety drivers comes next, then fully driverless rides. Waymo holds no German permit yet, so approval from federal and local regulators sets the pace. The company expects to open a paid public service late in 2027.

Amazon’s Toaster-Shaped Robotaxis Are Hitting the Road
Amazon’s Zoox has started charging for driverless rides in Las Vegas and rolling out in San Francisco. Unlike other companies, which still use traditional cars with extra equipment, Zoox uses purpose-built all-electric cars with no steering wheel or brake pedals. The National Highway Transportation Safety Administration cleared the company to charge for rides across up to 5,000 vehicles over the next two years. But they barred the cars from operating in heavy rain, snow and on roads with limits above 45mph (72 km/h). Zoox runs roughly 100 cars today, against about 4,000 for Alphabet’s Waymo.

The Humanoids at China’s Robot Games Were Faster Than Usain Bolt—but I’m More Impressed by Their Tweezer Mastery
Over 600 teams and more than 2,000 robots took part in this year’s World Humanoid Robot Games in Beijing. The biggest headlines came from robots running 100 metres faster than Usain Bolt and jumping higher than any human can. Robots also competed in ping-pong, football, weightlifting, and more, alongside practical challenges testing manual dexterity, such as “block building” and “tweezer bean picking”. Overall, the games offered a comprehensive look at what Chinese humanoid robots can do and their potential industrial applications. But, as Will Knight writes in this article, perhaps their biggest purpose is to get the public excited about the potential of robotics.

Hugging Face is selling a cute $399 open source duck robot, Microduck

Microduck is a cute robot duck from Hugging Face and Pollen Robotics. It is a 25 cm-tall bipedal robot that can waddle, pick things up with its beak, get back up when it falls, crouch, and even roller skate. Pre-orders for Microduck are now open, with the robot priced at $399 and expected to ship before Christmas. Hugging Face has open-sourced the AI controlling the robot and the RL training environment, both available on GitHub. The CAD files and BOM, however, do not appear to be publicly available. But if you are interested in building your own robot duck, check out Open Mini Duck, an open-source robotics project from one of Pollen Robotics’ engineers.

Anthropic: Previewing the Model Hardware Standard

Anthropic has introduced a research preview of Model Hardware Standard (MHS), a specification that lets AI agents operate lab and factory hardware such as microscopes and robotic arms. MHS replaces the patchwork of custom code used to control lab equipment with a standardised layer for AI agents, cutting setup time and allowing scientists to focus on their experiments. Anthropic admits that Claude’s physical reasoning still requires expert oversight and plans to open-source the standard only after developing safety evaluations with partners.

NVIDIA Announces Jetson Orin Nano 2 Robotics Computer to Redefine Entry-Level Edge AI
Nvidia has announced the Jetson Orin Nano 2, an updated version of its small computer for edge AI and robotics. The new Jetson Orin Nano still comes with 8 GB of memory but has a new eight-core ARM CPU and delivers 78 TOPS of AI compute, up from 67 TOPS. Nvidia says the Jetson Orin Nano 2 achieves twice the inference performance of the Jetson Orin Nano Super through improved Tensor Cores and higher memory bandwidth, while maintaining the same compact form factor. In 15-watt mode, the Jetson Orin Nano 2 consumes 40% less power while delivering the same performance as its predecessor. The Jetson Orin Nano 2 module and developer kit are expected to be available in the first half of 2027. Nvidia did not disclose pricing.

Arduino VENTUNO Q
Qualcomm (which owns Arduino) is taking on Nvidia in edge and robotics AI with the Arduino VENTUNO Q, a small but powerful computer designed to compete with Nvidia’s Jetson family of products. The Arduino VENTUNO Q offers 16 GB of LPDDR5 RAM, eight CPU cores, and up to 40 dense TOPS of AI performance. What sets the VENTUNO Q apart is that it has two chips (or “brains”, as Qualcomm calls them): the Dragonwing IQ-8275 processor (the “AI Brain”), which combines a CPU, GPU, and NPU, and the STM32H5F5 microcontroller (the “Action Brain”), which enables sub-millisecond response times—an attractive feature for industrial and robotics applications. Qualcomm sees the new device as an ideal computer for local and edge AI, as well as for robotics and industrial applications. The board is currently available for pre-order and costs $299.

Generalist GEN-1.5

Generalist AI says its new model, GEN-1.5, can learn a physical task from a single demonstration lasting just a few seconds, with no retraining. The company reports an average success rate of 59% across ten tasks using this approach, rising to 83% after a few minutes of fine-tuning. It says this ability emerged on its own after eight months of pretraining on real-world interaction data. The tasks are short and simple, and the results are self-reported. I recommend checking out the video Generalist uploaded alongside the announcement, as what they are showing is quite impressive.

Introducing S1: In-Context Learning for Robotics
Skild AI has unveiled S1, a robotics foundation model that learns a new task from a single video demonstration. The company says it needs no fine-tuning at all. Its own tests report a 66% success rate on tasks never seen in training, against 9% for language-prompted rivals. Skild claims S1 can handle ten-minute jobs it was never trained on, such as brewing coffee or cooking pancakes. Normally, this would take hours of teleoperation first.

Introducing Index: Building The World’s Largest and Most Diverse Physical Dataset
Figure presents Index, a dataset for robot training that the company has been building in secret for the past four months. Through a dedicated app, more than 44,000 weekly active users (which Figure calls Creators) have recorded over 16 million videos of themselves performing various chores. Figure says it has paid out $15 million to Creators who have contributed to the dataset so far. The Index app is now available for Android and iOS. Videos uploaded to Index will be used to train Helix and help solve the training data problem for humanoid robots.

How US military funding propelled China’s robot dogs
This article argues that Unitree built its best-selling robot dogs on quadruped breakthroughs funded by the US Army. It cites researchers who worked on the project, who say Unitree’s Go series matches MIT’s Mini Cheetah almost to the millimetre. Nothing underhanded happened, as the findings were published openly to push the field forward. Unitree's advantage was manufacturing, backed by state subsidies and a dense supplier cluster in Hangzhou. The researchers argue that faster commercialisation, rather than secrecy, is the best way to counter China’s scale advantages.

China’s robot traffic police can do almost everything except stop you
Traffic police in Hangzhou, China, have deployed 15 humanoid robots to direct pedestrians, manage traffic flow, and answer tourists’ questions. When the robots spot a violation, they file a report for a human officer to act on, as Chinese law reserves coercive powers for officers rather than equipment.

▶️ Unitree New Robot Preview: “Superman” Breaking the Limits of Humanity (0:30)

Unitree is back with another video showcasing the athletic capabilities of its humanoid robots. This time, the company showed its robots reaching 2 metres in a standing high jump and sprinting at a top speed of 12.66 m/s. For comparison, Usain Bolt’s top speed at his peak was 12.42 m/s.

▶️ Inside Persona AI’s Plan to Make Humanoid Robots Profitable (46:16)

In this episode of Automated Podcast, Brian Heater speaks with Nic Radford, co-founder and CEO of Persona AI and a former NASA roboticist. The conversation revolves around Radford’s and Persona’s goal of making humanoid robots practical. They focus on robots that can do real work, not just backflips. Radford shares his approach to the business and economics of humanoid robotics, and how Persona works to put robots to work and start delivering value. Radford also discusses the original vision for Robonaut and Valkyrie, the value of government-funded moonshots, DARPA’s role in modern robotics, and lessons humanoid companies can learn from the development of self-driving technology.

🧬 Biotechnology

A startup trains AI on living human skin tissue
TechCrunch profiles Michael Polansky, a former Founders Fund principal who ran Sean Parker’s family office before becoming Lady Gaga’s business and romantic partner. He is now going public with Outer Biosciences, a startup he has quietly led since 2020. The company keeps skin left over from surgery alive for up to a month, compared with an industry norm of just a few days. That window lets an AI model predict which compounds will work, test them on the tissue, and then learn from the results. The approach has taken the company from finding two promising compounds in 18 months to finding a new one every six weeks. Polansky says the data is finally good enough for the company to work in the open.

Mini Brains Grown for Five Years Matured Like Human Brains
Researchers managed to keep lab-grown brain organoids alive for over five years, far longer than the few months they usually survive. The trick was swapping their nutrient mix midway through, which stopped the neurons from withering. By the end, gene activity in the oldest blobs resembled that of a typical four-year-old. The cells even seemed to track their own age, maturing on schedule when mixed with much younger ones. Until now, mini brains died long before reaching the stage when disorders like schizophrenia and epilepsy take hold. This result opens up the possibility of studying those and other disorders more closely.

💡Tangents

Meta agrees to major changes to Facebook and Instagram as it settles US trial over teen addiction for up to $17bn
Meta has settled a landmark lawsuit brought by California and 28 other US states, and agreed to pay up to $18 billion over ten years. The states accused the company of deliberately designing Instagram and Facebook to hook children and damage their mental health. Meta denies wrongdoing, but will still cap teenagers’ daily use, block the apps overnight and pause notifications during school hours. That makes it the first time the company has been forced to change how everyday users experience its apps. Thousands of similar claims against Meta, TikTok, Snap and YouTube are still pending.

Apple Mac Studio M5 Ultra and Mac Mini M6 Launched
Apple introduced new Macs this week—the Mac Mini M5 Pro, Mac Studio M5 Max, and Mac Studio M5 Ultra, as well as Mac Minis with new M6 processors. On paper, they look very good. Apple claims the M6 delivers 1.2x faster multithreaded CPU performance than the M5, nearly 30% more peak GPU AI compute, and 10% more memory bandwidth. The M5 Ultra, with a maximum unified memory pool of 512 GB (although only versions with 256 GB are currently available), is shaping up to be a small but capable machine for local AI. What’s not small are the prices, especially for higher-spec machines.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

This week, we take a closer look at the reports about the financial situations of both Anthropic and OpenAI, and what they actually say.

Elsewhere in AI, OpenAI signs a lease for a 10 GW data centre in Ohio, overhauls its safety protocols, launches ChatGPT for Teens, and says California should strengthen its AI safety bill. Meanwhile, Cursor launches an AI-first competitor to GitHub, Stripe acquires OpenRouter, Jeff Dean’s startup is in talks to raise $1 billion, and Etched doubles its valuation to $21 billion in a month.

Over in robotics, Tesla is preparing to launch Cybercab, AMD challenges Nvidia in chips for robots, Amazon Prime Air expands to nearly 500 US cities and towns, and a humanoid robot runs 100m faster than Usain Bolt.

Other than that, this week’s issue also covers a breakthrough in personalised mRNA cancer vaccines, the US government opposing Apple’s use of Chinese memory chips, an important trial against Meta over claims of child social media addiction getting under way, and more!

Enjoy!

Anthropic overtakes OpenAI

Leaked figures put Anthropic's sales ahead of OpenAI for the first time, weeks before it could file for a record IPO

As both Anthropic and OpenAI prepare for what could be some of the largest IPOs in history, more and more information about their financial situation is leaking to the public. If you read those leaks and reports side by side, you would get pictures of two different companies—one that seems to be on the rise, while the other stumbles. Let’s look at what each picture actually says, and then at how much of it we should believe.

Let’s start with Anthropic. According to leaked documents seen by Bloomberg and The Wall Street Journal, its revenue grew 14x in a single year. In Q2, the company booked more than $11.5 billion, against $787 million in the same quarter of 2025 and $4.73 billion in Q1 of this year. Meanwhile, its run rate hit $65 billion at the end of July, up around sevenfold from about $9 billion late last year (run rate takes recent performance, often a single month, and projects it across a full year). Anthropic has also reportedly overtaken OpenAI on sales, and was on track for a first quarterly operating profit of $559 million while doing it, telling investors it had got better at using computing power efficiently.

It is no wonder, then, that Anthropic is pushing ahead towards its IPO with such good numbers. Anthropic could file publicly within weeks and list as soon as this autumn, ahead of both OpenAI and DeepSeek. The company expects its IPO to match or even surpass SpaceX's, which raised $75 billion at a massive $1.77 trillion valuation. Some investors and analysts say Anthropic can enter the public market as a $2 trillion company.

To get anywhere near $2 trillion, however, Wall Street is being asked to look unusually far ahead. Anthropic is projecting revenue of $190 billion to $200 billion in 2028, and bankers are pricing the company off that forecast rather than off what it earns today. This is not how companies are normally valued. It happens when current accounts are dominated by spending that everyone expects to shrink as a share of revenue, and it only works if the growth holds.

There is a hole in the picture, too. Anthropic’s small quarterly profit sits alongside a net loss of almost $42 billion in 2025, roughly five times the year before, and the company has not explained how it reached the adjusted figure. In past communications with investors, it has left out stock-based compensation.

Meanwhile, OpenAI has troubles of its own. The company told investors it made $6.7 billion in the same quarter, up 18% from $5.7 billion in the first. Its operating loss, which does include stock-based compensation (which Anthropic is reportedly leaving out), widened from $9.3 billion to $12.3 billion.

The problem is not the size of that number but the expectations attached to it. OpenAI has signed computing deals premised on soon earning hundreds of billions a year, and the results of Nvidia, Oracle and others now depend on it delivering. Measured against that, 18% is a stumble. OpenAI grew more slowly over the quarter than Palantir did, and it trailed CoreWeave and Micron too. Anthropic, for comparison, more than doubled its revenue quarter on quarter.

Unlike Anthropic, OpenAI seems to be experiencing headwinds and might not go public this year, a possibility The New York Times first reported in June. Sarah Friar, OpenAI CFO, told staff the company will list in 2027 or sooner if growth picks up. She also told employees not to worry about Anthropic's IPO, and reassured them by pointing to a run rate up 35% so far this quarter, enterprise revenue up 50%, and a coding product with 20 million weekly users.

But remember where all of this came from—leaks. Every figure above reached us through a leak, and some of them might be intentional. A company that is on its way to IPO has good reasons to let flattering numbers find their way into print.

The contest underneath is also less settled than the leaks suggest. Ramp’s AI Index, which tracks spending by more than 70,000 American businesses, put Anthropic ahead on market share in July, at nearly 44% to OpenAI’s 40%. But the new data shows OpenAI is growing faster among those businesses so far in Q3. A lot can still change. Corporate customers switch to whichever model best fits their needs, and neither company has locked in a lead. Let's also not forget the open models coming out of China, which are closing the performance gap with the best that Anthropic and OpenAI offer, and at a lower price.

We will get the full picture of how financially sound Anthropic’s and OpenAI’s businesses are when the prospectuses are unsealed.

If you enjoy this post, please click the ❤️ button and share it.

Share

News roundup

🧠 Artificial Intelligence

OpenAI Locks In Lease for Huge Data Center in Ohio With Backing From Nvidia
OpenAI has signed a 20-year lease for a 10-gigawatt data centre campus in southern Ohio, one of the largest such projects to date. The data centre will be built by SoftBank’s SB Energy, with Nvidia providing the chips and guaranteeing the value of the completed site, up to $105 billion, if OpenAI walks away. Initially, as The Wall Street Journal reports, Nvidia was supposed to backstop the entire $250 billion project, but scaled back its commitment to address investor concerns about risk. The guarantee gives lenders the confidence to finance construction. Nvidia gets plenty in return: exclusive chip supply for the first five-gigawatt phase, plus $1.5 billion in equity in SB Energy.

Cursor Origin
Coding agents need coding tools designed for them, and Cursor is delivering one in the form of Origin, an agent-first code repository. For now, Origin offers the essentials—repositories, pull requests, code browsing and GitHub sync—with agent-native features coming soon. Cursor chose the perfect time to announce Origin, unveiling it on the same day that GitHub, the service it aims to replace, was dealing with yet another major outage. Origin is available in early beta on all paid plans.

OpenRouter is Joining Stripe
Stripe has confirmed that it is acquiring OpenRouter, an AI model marketplace and gateway, in a deal reported to be worth around $7.5 billion. Stripe is paying almost six times the $1.3 billion valuation OpenRouter secured just three months ago. Both sides insist that OpenRouter will continue to operate independently, with its product, mission and neutrality intact, once the deal closes in the coming weeks.

Anthropic sees AI risks rising, no plan to release stronger “Model 2”
Anthropic is reportedly developing an internal model called Model 2, which is said to perform noticeably better than its flagship Mythos on many internal tasks, but the company has no plans to release it. The disclosure came in a risk report that also raised Anthropic’s estimate of misalignment risk in high-stakes situations from “very low” to “low”, citing recent cybersecurity incidents. Anthropic also says it is less confident in its risk assessments than before because some of its evaluations have saturated and no longer capture how quickly model capabilities are improving.

OpenAI Overhauls Safety Protocols After Its AI Agents Went Rogue
OpenAI has paused a significant number of training runs and evaluations for Astra, its next frontier model, while it tightens security after internal tests showed Astra outperforming its predecessors at coding and hacking tasks. New safeguards include stronger sandboxes, stricter internet isolation and classifiers that monitor the model’s internal reasoning for signs of trouble.

Departing Google chief scientist Jeff Dean has been in talks for a $10 billion valuation for his new AI startup
Business Insider reports that Discovery Loop, an AI startup founded by legendary Google engineer and former Chief Scientist Jeff Dean, is in discussions to raise $1 billion at a valuation of around $10 billion. Discovery Loop, founded by Dean and three other Google veterans who recently left the company, is pursuing recursive self-improvement—AI that improves itself with little or no human help—and plans to apply it to hardware design, drug discovery, materials science and scientific research.

Nvidia discloses $21 billion stake in SpaceX at end of second quarter
According to a Securities and Exchange Commission filing, Nvidia owns 122.8 million Class A shares in SpaceX worth about $21 billion. It is Nvidia’s second-largest holding, behind only its stake in Intel, which is currently worth about $22 billion. CNBC reports that Nvidia’s stake in SpaceX stems from a $10 billion investment in xAI in January.

Anthropic IPO filing will show AI backlash as a risk factor
CNBC reports that Anthropic plans to name public anger over data centres as a key risk in its IPO prospectus. A Gallup poll in May found seven in ten Americans oppose data centres near them, and politicians in both parties are now running against new builds ahead of the midterms. Since compute capacity drives revenue at AI labs, a slowdown in construction would hit Anthropic's growth directly.

Anthropic prepares supervoting power for founders ahead of IPO
Anthropic is reportedly preparing to give its founders shares with extra voting power to help insulate them from external shareholder pressure. The founders own unusually little of the company today, with Dario Amodei himself said to hold about 2%. The company is also planning to maintain its existing body of non-shareholder trustees with a special class of ​stock to elect a majority of members to the company's ​board of directors.

OpenAI says California should strengthen its AI safety bill
OpenAI now wants California to toughen SB 53, the AI safety law it opposed last year. It says the state should require monitoring of frontier models during training and evaluation, plus stronger cybersecurity across model development.

Slack Code: Where Your Team and Agents Build Together
Slack has launched Slack Code, which brings collaboration with AI coding agents into shared “code channels”—dedicated spaces where humans and AI agents can work together. Slack Code launches with Anthropic’s Claude, Cognition’s Devin, GitHub Copilot and Vercel, with OpenAI’s ChatGPT coming soon. It is pitched as the first step towards a more “multiplayer” model of agent-assisted work that will eventually extend beyond engineering to areas such as marketing, legal and IT.

ChatGPT update adds Apple Messages integration on Mac
OpenAI has introduced a new feature to the ChatGPT Mac app that allows the AI to read, search and send iMessage, SMS and RCS messages through Apple’s Messages app. The feature uses a new Apple Messages plugin and is available on all plans in the ChatGPT desktop app for macOS. By default, no messages are sent until you approve both the message and its recipients. OpenAI warns that granting standing approval carries real risks.

Ramp launches its own AI model router, called Router
Ramp has launched Router, a service that lets customers choose models from eight providers, including OpenAI, Anthropic and DeepSeek, through a single API. Customers can route requests based on cost, benchmark scores or difficulty, then track spending and latency on a dashboard. Ramp’s Router is currently available only to US customers. It's free until the end of the year, though customers still pay for inference.

Researchers say OpenAI revoked their access to limited cyber program
Several cybersecurity researchers said that OpenAI revoked their access to Trusted Access for Cyber, a programme that gives vetted defenders access to its best models with looser security guardrails. Those affected were told that their identities couldn’t be verified or that their accounts were ineligible. TechCrunch found that the researchers it spoke to who were affected by the problem live outside the US and Europe, suggesting that the revocations may be limited to certain regions. OpenAI blamed a technical error on its end and asked them to reapply and complete the verification process.

OpenAI launches a safer ChatGPT for teens — years after teens started using it
OpenAI has launched ChatGPT for Teens, a version of its chatbot with additional safety measures and educational tools, including a new Study Mode designed to encourage teens to explore their curiosity and work through problems rather than simply get quick answers. It also comes with parental tools for managing settings, receiving safety notifications and setting Quiet Hours, among other features. OpenAI says age-appropriate protections designed to reduce exposure to harmful or developmentally inappropriate content will be enabled by default in ChatGPT for Teens.

Bringing the cybersecurity capabilities of Claude Mythos 5 to more defenders
Anthropic is expanding access to Claude Mythos 5, its most capable model, to more partners without exposing the model directly. Mythos 5 now powers Claude Security scans for Enterprise customers and is being integrated into partners’ security tools, where users receive only specific outputs such as suggested patches rather than open access to the model. Alongside this, a new Defender Advantage Fund offers $35 million in credits to help secure open-source projects, and the Cyber Verification Program will expand to cover broader dual-use capabilities on Opus and Sonnet, with Mythos-class access to follow.

Etched’s valuation doubles to $21B in a month
Etched, an AI chip startup founded in 2022, has raised $700 million at a $21 billion valuation, twice what it was worth in July. In a post announcing the fundraise, Etched revealed that Jane Street received one of its racks for testing and was so pleased with the results that it led the funding round. Etched sells complete systems optimised for inference, which the company claims produce answers faster and more cheaply than the competition. If you want to learn more about Etched, I recommend this video from Dr Ian Cutress on his YouTube channel, TechTechPotato.

Cerebras CS-4
Cerebras has launched CS-4, the fourth generation of their AI accelerator. The company claims that CS-4 delivers up to 30 times faster inference than GPU systems, better economics, and a simple path to deploy hyperscale capacity. CS-4 reduces wafer-to-wafer interconnect latency to as low as 2 microseconds, and can deliver more than 1,000 tokens per second on models exceeding 10 trillion parameters. All of that performance comes in one rack, which Cerebras says can be quickly manufactured, installed, expanded, and upgraded.

GLM-5.3 on Artificial Analysis
Last week, Z.ai released its latest flagship model, GLM-5.3, one of eight new models released that week. When I was preparing last week’s issue, we did not have independent evaluations to compare against Z.ai’s claims. Now we have. Artificial Analysis places GLM-5.3 as one of the leading models in its Intelligence Index and at the level of top US models, while being cheaper and faster. However, the benchmarks also found GLM-5.3 to be very verbose, which can affect the final cost.

Introducing Harvey II
Harvey has launched Harvey II, its latest AI agent focused on legal work. According to the company, the new agents can now open inside a “Space” for a given matter, inheriting its documents, parties, permissions and history. A new memory layer learns each lawyer’s drafting style and citation preferences across Harvey, Word and Outlook. Harvey has also built Tenet, its first model post-trained for legal reasoning, which the company claims matches the strongest general-purpose models at far lower cost.

Mistral Agentic Search
Mistral has launched Agentic Search, a retrieval layer that lets AI models dig through documents efficiently. The model gets five file-like tools: search, open, navigate, read and grep. It can refine a query, open a filing and check a specific table before answering. Mistral’s own benchmarks claim accuracy on financial filings jumped from 26.7% to 86% with Agentic Search, while using fewer tokens.

AI Chip Startup Fractile Seeks $6.5 Billion Value After Anthropic Deal
Bloomberg reports that Fractile, a UK AI chip startup, is in talks to raise about $600 million at a $6.5 billion valuation, six times more than its value in May. The company has reportedly signed a $250 million deal with Anthropic, with the intention to expand that contract in the future. The first chips are expected to be delivered in 2027.

SpaceX Attempted to Acquire AI Coding Startup Cognition
Before acquiring Cursor for $60 billion, SpaceX was reportedly eyeing Cognition, the startup behind the coding agent Devin. Cognition didn’t engage with the takeover approach, but the two companies are now discussing ways to work together, with Cognition potentially using SpaceX’s computing capacity.

🤖 Robotics

Tesla prepares for Cybercab launch in August starting with rides to employees
The Information reports that Tesla is preparing to launch Cybercab, its autonomous robotaxis without pedals or a ​steering wheel, later this month. The rollout will begin in Austin, Texas, with rides on public roads being initially offered to Tesla employees. The company started running tests of the production version of its Cybercab on public ⁠roads ​in June, and production of ​the vehicle is expected to ramp up later this year.

AMD Challenges GPU-Centric Architectures as It Takes Aim at Nvidia in Robotics
AMD challenges Nvidia in robotics with the Ryzen AI X100, a small computer combining a Zen 5 CPU, RDNA 3.5 GPU and XDNA 2 NPU with unified memory. AMD is also pitching openness as a differentiator—its new robotics board uses an industry-standard format rather than a proprietary one, and the company is publishing the design files, along with a ROCm-based robotics SDK.

Amazon Prime Air drone delivery is expanding to nearly 500 US cities and towns this year
Amazon announced that Prime Air, its drone delivery service, is expanding to nearly 500 cities and towns in the US this year. The service currently operates from 11 sites across ten metro areas in seven states, with launches planned soon in Illinois, New York, Ohio, Georgia and Idaho. Additionally, Amazon plans its first international site in Darlington, England. Prime Air drones can carry items under five pounds (about 2.27 kg) that fit in a large shoebox (Amazon says this covers more than 60% of most commonly ordered products) and deliver the orders in as little as 30 minutes, though most take around an hour.

Uber adds Zipline drones to its Eats delivery network
Your next Uber Eats order could be delivered by one of Zipline’s drones (if you live in one of the US cities where Zipline operates). This is thanks to a partnership between the two companies announced this week. Zipline will begin delivering Uber Eats orders by the end of the year, with the goal of making one million deliveries per day by the end of 2029. Uber is also investing in Zipline, although the amount has not been disclosed.

Chinese robot runs 100m sprint quicker than Usain Bolt’s world record

A humanoid robot has completed a 100m sprint in 9.32 seconds, beating Usain Bolt’s world record of 9.58 seconds, set 17 years ago. The robot, named Lightning and developed by Chinese smartphone manufacturer Honor, reached a peak speed of 14.5 metres per second.

Unitree Sees Robots Reaching ‘Mass Market’ Within Decade
Speaking at the World Robot Conference in Beijing, Wang Xingxing, founder of Unitree, said that humanoid robots will reach “mass-market” popularity within a decade, once manufacturers solve the problems limiting their current usefulness. The main challenge, Wang said, is that robots perform well in test conditions but struggle in messy, real-world scenarios. He said that once this problem is solved and robots can handle 80% of tasks given to them via voice instructions in unfamiliar environments, the industry will reach “mass-market” popularity. Wang expects this to happen within two to ten years.

Waymo lifts the lid on the ‘brain’ powering its robotaxis
Waymo has revealed the compute system in the trunks of its robotaxis for the first time. As the company writes in a blog post, the system is built around three principles: responsiveness (fully onboard, millisecond-level decisions, with compute power scaling twentyfold in eight years), ruggedisation and redundancy (two parallel units, either of which can take over). At its core is a custom 5nm ASIC delivering more than 1,000 TOPS and fusing raw lidar, radar and camera data at the front end. This article from The Verge argues that Waymo’s post is an implicit rebuttal to Elon Musk’s dismissal of lidar and puts the cost of the hardware at an estimated $20,000–25,000 per vehicle—down sharply from the previous generation, but still a drag on profitability.

Is Shipyard Welding the Right First Job for Humanoid Robots?
While almost everyone is betting on humanoid robots working in warehouses or in our homes, Persona AI imagines its robots welding and building ships. Founded by veterans in humanoid robotics, Persona is targeting much more expensive skilled labour, and welding is one such skill. Persona is starting with the long, straight seams that shipyards need by the hundreds of kilometres per vessel, and it argues that the value lies not in replacing workers at labour-starved yards, but in enabling shipyards to build more ships.

What happens when a kid’s robot best friend dies?
This article tells the story of Moxie, an AI-powered companion robot for children, and what happens when a robot designed to form connections dies. It explores the potential of such robots to help neurodivergent children practise social skills and examines what happens when the company behind such a robot folds and shuts down its servers.

Drones With Claws Perch on Arctic Icebergs

Ice Dart is a drone equipped with spikes that allow it to land on and latch onto icebergs and glaciers. Thanks to its custom landing gear, the drone can survive landings that would damage a traditional drone and grip icy slopes of nearly 60 degrees. According to the researchers, Ice Dart could serve as a mobile platform for monitoring and tracking iceberg movements, providing an additional layer of data to complement satellite- and ship-based iceberg detection.

🧬 Biotechnology

Vaccine breakthrough stops cancer returning in trial
Merck and Moderna announced that their personalised mRNA cancer vaccine, intismeran, combined with the immunotherapy drug Keytruda, significantly extended the time that high-risk melanoma patients remained cancer-free and reduced the risk of the cancer spreading. The vaccine is made for each patient by studying their own tumour, so the immune system learns to hunt down any remaining cancer cells. Experts have welcomed the news but stress the results are early, have not yet been independently checked, and the treatment still needs approval before it could reach patients.

GenBio Launches a “Virtual Cell” AI Model
GenBio, a startup cofounded by David Baker, a 2024 Nobel Prize laureate for his work on computational protein design, has unveiled AIDO Cell—a “world model” for cells that can simulate a human cell in its normal state and how it can react to drugs. In an early test, it correctly reproduced a known leukaemia drug’s effects, but it currently works with only two lab-grown cell types, and the company calls it a preview. GenBio hopes it will eventually let researchers test drugs on a computer before trying them in the lab.

💡Tangents

U.S. Urges Apple Not to Buy Chinese Memory Chips
In response to the memory chip shortage caused by the AI data centre boom, Apple could either significantly raise the prices of new devices (which it already did) or look elsewhere for memory chips, with China emerging as a potential source. However, Commerce Secretary Howard Lutnick said the Trump administration doesn’t want Apple to do that, urging the company to find other solutions to the shortage. Apple has reportedly been testing Chinese memory chips as a prelude to potentially using them in Apple devices sold in China.

States Seek $200 Billion From Meta Over Child Social Media Addiction Claims
Four states—California, Colorado, Kentucky and New Jersey—are suing Meta in an Oakland federal trial, claiming Facebook and Instagram were designed to hook children, fuelled a youth mental health crisis, and that Meta misled users about safety while breaching child privacy laws. They want penalties nearing $200 billion plus design changes such as ending infinite scroll and beauty filters for under-18s. First witness Arturo Béjar, a former Meta engineer, testified that the company knowingly tolerated underage users and put growth ahead of safety. Meta denies it, citing its teen safeguards and calling the demands disproportionate. The trial should run about six weeks.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

This week has been crazy. We’ve had not one, not two, not three, but eight new models released, with a nice mix of open and proprietary models. We’ll take a close look at all eight of them and what they bring to the table in this week’s write-up.

Elsewhere in AI, Nvidia and six big Wall Street investors raise $500 billion to finance an AI infrastructure buildout, Anthropic is courting investors ahead of its potentially $2 trillion IPO, and OpenAI wraps up a $7 billion share sale. Other than that, OpenAI releases GPT-5.6 Sol Ultrafast while high-profile exits raise red flags, Mark Zuckerberg shares his vision for personal intelligence, Anthropic starts watermarking text produced by Claude, and Chinese AI chipmakers are gaining from Beijing’s push to rely less on Nvidia.

Over in robotics, the FCC moves to ban LiDAR-equipped foreign drones from the US, DARPA wraps up the Lift Challenge, Uber and Pony.ai are launching 2,000 robotaxis in Europe, we look at why building a humanoid robot in the US is hard without Chinese components, and more!

Apart from that, this week’s issue of Sync also features scientists creating female clones of male mice, allergy-free genetically engineered dogs, OpenAI sharing details of the Hugging Face incident, a profile of the secretive yet influential wife and adviser of Dario Amodei, and more!

Enjoy!

Eight new models in one week

Google, SpaceXAI, Meta, Nvidia, Microsoft, DeepSeek and Z.ai all picked the same week to release new AI models

I don’t recall a week like this before. Normally, there is one, maybe two, model releases to analyse. But this week, seven companies decided to release eight models, plus one AI lab made weights public for a model released some time ago. It is a nice mix ranging from proprietary frontier challengers to open-weight models and efficient workhorses. In this article, we’ll look at all eight new models, what they bring to the table, and where they fit in an increasingly crowded AI landscape.

Gemini 3.7 Flash

Google has released a new model this week. But it is not the long-awaited successor to Gemini 3.1 Pro. Instead, it is another update to the Gemini Flash line. Just three weeks after Gemini 3.6 Flash, Google presents Gemini 3.7 Flash, its “most intelligent workhorse model yet for coding and agents.”

Gemini 3.7 Flash is being positioned as a model that balances performance and cost. It is aimed mainly at coding, but can also be useful for all kinds of knowledge work. Google says the new model adapts better to roadblocks, clarifies intent when needed, and is better at following instructions. It also exhibits more disciplined execution, which allows it to work more reliably.

Google reports that Gemini 3.7 Flash is a noticeable improvement over its predecessors. It tops Gemini 3.6 Flash in almost all the published benchmark results (the only benchmark on which Gemini 3.6 Flash scored better is CharXiv Reasoning, but not by much). The one chart that Google decided to highlight is the comparison between cost and performance across different models on DeepSWE V1.1, which measures frontier coding agents on original, long-horizon engineering tasks.

However, it is worth noting that those benchmarks used introductory pricing of $0.75/$3.75 per million input/output tokens. Once the introductory price expires at the end of the year, Gemini 3.7 Flash will be twice as expensive, which would put it on the DeepSWE graph roughly where Kimi K3 is.

Artificial Analysis gave Gemini 3.7 Flash a score of 56 on its Intelligence Index, which places the new model between GPT-5.6 Terra and Claude Sonnet 5.

Gemini 3.7 Flash is available on the Gemini app, Spark, Antigravity, and via the Gemini API.

Grok 4.6

SpaceXAI makes a run at the top with Grok 4.6. And while its new flagship model doesn't take the crown, it shows that SpaceXAI can produce a competitive frontier model.

On the Artificial Analysis Intelligence Index, Grok 4.6 scores 61 points—five points up on Grok 4.5 and 23 up on Grok 4.3. That puts Grok 4.6 at the level of Claude Opus 5 and GPT-5.6 Sol. And with the price of $2/$6 per million input/output tokens, SpaceXAI’s new flagship model is cheaper than Opus 5 or GPT-5.6 Sol ($5/$25 and $5/$30 per million input/output tokens, respectively).

SpaceXAI positions Grok 4.6 squarely for long-running agents and coding, and I think this is part of a bigger plan for SpaceXAI to at least stay relevant. Grok does not have a particularly good reputation, and it is not doing especially well in the AI race. But SpaceX now owns Cursor, a popular AI coding tool, and it might have an opportunity here to improve its numbers, or at least have something it can spin as a good story.

Cursor currently offers a wide range of models to choose from, including top models from OpenAI, Anthropic, and Google, as well as its own coding models and Grok. But I think Grok will come to occupy a more prominent place in Cursor. It could become the default model, offered at competitive pricing in an attempt to convince software engineers to switch from Claude or GPT-5.6 to Grok.

This could give SpaceXAI a foothold in the lucrative AI coding market. It is unclear how far SpaceXAI can go, or whether it can grab a sizeable chunk of the market from Anthropic and OpenAI. But at the very least, SpaceXAI could make the number of people using Grok for coding look better.

Grok 4.6 is available in Cursor and Grok Build. It’s also available through the API and via other partners such as OpenRouter, Vercel, and Cloudflare.

Muse Glimmer

Do you remember when Meta was the unlikely champion of open-weight models? Meta seems to have remembered those good old times and is now back with a new open-weight model—Muse Glimmer.

Glimmer is a 30-billion-parameter open-weight model built for local agentic work. It’s distilled from Muse Spark, Meta’s closed flagship model, and designed to run on a desktop or laptop with a single consumer GPU. The full-precision version with no quantisation requires 64GB of VRAM. There are quantised versions of Glimmer that need just 32GB or 24GB of VRAM, which is enough to fit Glimmer on an NVIDIA RTX 5090. Meta says that the quantised versions of Glimmer show up to 1% degradation in performance compared with the full version.

Meta pitches Glimmer as a model to power always-on local AI agents. Glimmer is meant to call tools, write and debug code, work through screenshots and documents, and recover when a tool call fails instead of stalling. Meta imagines it managing your schedule, drafting your messages, and organising your files—exactly the kind of work that needs deep access to personal data, and exactly the argument for keeping it local.

When it comes to performance, Meta’s own benchmark results put Glimmer ahead of Gemma 4 31B and Qwen3.6 27B on most agentic and reasoning benchmarks. Artificial Analysis tells a slightly more sober story. Glimmer scores 35 points on the Intelligence Index, around Kimi K2.5 (36) and just behind Qwen3.6 27B (38). It also falls behind Qwen3.6 27B and Gemini 3.5 Flash-Lite in benchmarks measuring models on realistic knowledge work tasks in an agentic loop. Additionally, Artificial Analysis found that Glimmer has an 82% hallucination rate, compared with Qwen3.6 27B’s 49%.

The release of Muse Glimmer is a welcome surprise from Meta after the company abandoned the Llama family and rebuilt its AI teams. As Mark Zuckerberg outlines in his latest manifesto, Meta will offer a mix of open and closed models. He argues that distributing superintelligence widely “has the potential to begin a new era of personal empowerment,” promising a personal agent that works around the clock on your relationships, health, career and finances. Open models, like Glimmer, will play a part in fulfilling his vision of personal intelligence.

Glimmer’s weights are available on Hugging Face.

Nvidia Nemotron 3.5 Lightning

Nemotron 3.5 Lightning is the latest addition to Nemotron, Nvidia’s family of open-weight models. And as its name suggests, the new model focuses on speed.

Nvidia’s framing is that modern agentic systems are ensembles. There is a big reasoning model like Nemotron 3 Ultra or GPT-5.6 that plans and orchestrates, while smaller specialised models handle the high-volume grunt work—code review, tool calls, security alert monitoring, billing questions. Lightning is built for that second role. Nvidia claims up to 4x faster output speed and 30% faster agentic task completion than other models in its class. If needed, the model can be post-trained on custom data with NeMo.

Artificial Analysis supports Nvidia’s speed claims. In its tests, Nemotron 3.5 Lightning produced 302 output tokens per second, a result second only to Gemini 3.5 Flash-Lite’s 373 output tokens per second and well ahead of everyone else. In terms of intelligence, Lightning scored 24 points on the Intelligence Index and is comparable to Qwen3.5 35B and Gemma 4 31B. Artificial Analysis also found that Lightning is quite verbose. It generated 100 million tokens to complete the Intelligence Index evaluation, compared with a median of 43 million tokens, which will have an impact on the cost of running the model.

Alongside Nemotron 3.5 Lightning, Nvidia released NeMo Switchyard, an open-source routing library, and the two are clearly meant to be read together. Switchyard sends each request—or each step of an agent's workflow—to whichever model in a developer's pool is best suited to it, weighing model capability, cost and latency. It is provider-agnostic and accepts OpenAI, Anthropic and Responses API requests, so it can sit in front of a mix of open and proprietary models. The library ships with several routing strategies, including an escalation router that starts on a cheap model and switches to a stronger one when an LLM judge detects sustained difficulty.

Nemotron 3.5 Lightning is available on Hugging Face. As with previous Nemotron releases, Nvidia is publishing training data and methods where licensing allows, along with an agentic reinforcement learning dataset used for the coding post-training. NeMo Switchyard, meanwhile, can be found on GitHub.

MAI-Thinking-1 and MAI-Code-1.1-Flash

Let’s now turn our attention to Microsoft, which brought two new models to the table this week—MAI-Thinking-1 and MAI-Code-1.1-Flash.

MAI-Thinking-1 is Microsoft’s medium-sized reasoning model positioned as cost-efficient reasoning for enterprise work. Microsoft claims that the new model goes toe-to-toe with Claude Opus 4.6 on SWE-Bench Pro despite the smaller inference footprint, and that users prefer it over Claude Sonnet 4.6.

The more interesting claim is about how MAI-Thinking-1 was built. Microsoft is making a point of the fact that MAI-Thinking-1 was trained without distillation from third-party models and describes its training setup as a “Hill-Climbing Machine,” a pipeline designed so that every component can be improved incrementally over time. The argument is that inherited intelligence is faster to acquire but less steerable, and that clean, traceable training data matters for enterprise customers who need to know what shaped the model. Microsoft also says safety is trained through the same reinforcement learning loop as capability, treating unnecessary refusals and unsafe compliance as equivalent defects.

MAI-Code-1.1-Flash is an updated version of Microsoft’s coding model. Microsoft emphasises a noticeable uplift in efficiency. Compared to MAI-Code-1-Flash, the new version generates tokens 25% faster and uses 25% fewer tokens to complete a task. Taken together, those efficiency improvements mean MAI-Code-1.1-Flash costs a quarter of its predecessor. Additionally, Microsoft highlights a 15% improvement on .NET (Microsoft’s open-source development platform) tasks.

Both MAI-Thinking-1 and MAI-Code-1.1-Flash fit into Microsoft’s broader AI strategy. For a long time, Microsoft relied on OpenAI and Anthropic to deliver the AI behind intelligent features in its products. But the rising costs of AI made Microsoft invest in its own models that it can then serve more cheaply. Now Microsoft has those models and will start replacing models from OpenAI and Anthropic with them.

MAI-Thinking-1 is available in public preview in Microsoft Foundry. MAI-Code-1.1-Flash is available in GitHub Copilot.

DeepSeek V4 Pro

DeepSeek has finally taken DeepSeek V4 Pro out of preview, nearly four months after the April preview build. V4 Pro is a mixture-of-experts model with 1.6 trillion total parameters, of which 49 billion are active. It has a 1M context window and, as usual for DeepSeek, the weights are released under an MIT licence.

Artificial Analysis scores DeepSeek V4 Pro at 53 points on the Intelligence Index, which makes it the third-best open-weight model and well ahead of V4 Flash at 40 points. DeepSeek's own model card puts it at 80.6% on SWE-bench Verified—on the same level as Gemini 3.1 Pro and just behind Claude Opus 4.6.

Alongside the release, DeepSeek announced that it is raising API prices across the V4 line and introducing peak and off-peak rates. During peak hours, V4 Pro will cost $3.96 per million output tokens, up from $0.87, while V4 Flash goes from $0.28 to $1.32. Off-peak pricing is half the peak rate. But even with the increased prices, DeepSeek is cheaper than almost everyone: Moonshot’s Kimi K3 charges $15 per million output tokens, and Anthropic’s Fable 5 charges $50. Bloomberg ties the price increase to DeepSeek’s IPO preparations.

There was also something odd that people noticed. DeepSeek marked the general-availability release with a short statement on its website saying the model had "significantly enhanced agent capabilities." Three days later, the statement had been removed, according to the South China Morning Post, with no explanation. Early developer reaction to the 0813 build was reportedly underwhelming on general capability and negative on the pricing, though researchers were more impressed in narrower areas such as cybersecurity.

Alongside V4 Pro, DeepSeek also released a developer preview of DeepSeek Harness, its open-source answer to Claude Code. The main difference between it and other harnesses is its open architecture. Users can plug in almost anything: models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and UI components, all of which can be swapped or recomposed through configuration. DeepSeek Harness is available on GitHub for everyone to download.

V4 Pro is available through DeepSeek's API, app and web products, with weights on Hugging Face.

GLM-5.3

Z.ai closed out the week with GLM-5.3. What is interesting about this release is that GLM-5.3 runs on the same 743B-parameter base model as GLM-5.2. The improvements came from scaled-up post-training—more task environments, more environment types, and longer training runs.

The gains show up mostly on long-horizon work. Z.ai reports that the Terminal-Bench 3.0 score jumped from 4.6 to 28.3, and claims GLM-5.3 is the strongest open-weight coding model it has measured. On its internal Code Bench, which Z.ai keeps private to avoid test-set contamination, GLM-5.3 scores 31.4% using roughly 50,000 output tokens per task, against Claude Opus 4.8’s 29.5% at 120,000 tokens. Claude Fable 5 still leads that particular chart at 39.5%. Those claims have not been independently verified at the time of writing.

Z.ai emphasises GLM-5.3’s cybersecurity capabilities with this release. The company says the model’s capability in that area grew further than it expected as training scaled, eventually producing multi-step exploit-chain reasoning it hadn’t planned for. GLM-5.3 scores 84.5% on CyberGym, ahead of both Fable 5 and GPT-5.6 Sol. Z.ai says the model has already found over a thousand critical and high-severity vulnerabilities in deployed software, including bugs in the Linux kernel, WebKit and FreeBSD, and has published a public disclosure ledger to track them.

That is also why the weights aren’t out yet. Z.ai is holding them for roughly two weeks pending safety evaluation and hardening—the first time a model in the GLM line has been held back for this reason. In the meantime, selected security partners are getting access in controlled environments.

GLM-5.3 is available through the Z.ai API, the GLM Coding Plan and ZCode, and has been rolled out to existing coding plan subscribers.

Qwen3.8 weights are now on Hugging Face

Lastly, Alibaba has published the weights for its flagship model, Qwen 3.8. The first is Qwen3.8-2.4T, a 2.4-trillion-parameter open-weight model on which Qwen3.8-Max, Alibaba’s top proprietary model, is based. Alibaba has also published weights for a much smaller 28-billion-parameter model, Qwen3.8-27B.

If you enjoy this post, please click the ❤️ button and share it.

Share

🧠 Artificial Intelligence

Nvidia Taps Wall Street for $500 Billion Funding Commitment
Nvidia has teamed up with six Wall Street investors, including Apollo Global Management, Blackstone, BlackRock, and Goldman Sachs, to raise $500 billion in financing for AI infrastructure, helping Nvidia’s largest customers afford the computing power they need. Bloomberg reports the debt will be secured against that compute, with special-purpose vehicles leasing it on to clients. The first deals should reach the market within months.

Anthropic Tries to Shore Up Investor Confidence Ahead of Blockbuster IPO
The Wall Street Journal reports that Anthropic is courting investors ahead of what could be the largest IPO ever. Investors have pressed the company on cheap Chinese models, friction with the Trump administration, and public anger over data centres. Anthropic executives have downplayed the Chinese threat, citing the company’s lead in model quality. They have also promised a bigger push into healthcare and biology to soften hostility towards AI. Anthropic’s IPO is expected to take place in September or early October and could make it a $2 trillion company.

US to tell partners they must pick sides in AI race with China
The US is preparing to tell 35 countries they must choose between its AI coalition, called Pax Silica, and China's, according to a draft State Department letter seen by Reuters. The initiative ties partners together on chips, AI models and critical minerals. China set up a rival body in July, pitching its open-weight models as an alternative. Kazakhstan has signed up to both, which alarmed Washington. The State Department declined to comment on the leak.

OpenAI wraps $7 billion share sale ahead of potential IPO
OpenAI has completed a secondary share sale totalling roughly $7 billion, CNBC reports. The tender offer allows current and former employees to sell stock at the company’s $852 billion valuation without waiting for an IPO, which is reportedly delayed to next year.

OpenAI CFO Friar tells investors that enterprise business now bigger than consumer by revenue
Sarah Friar, chief financial officer of OpenAI, told shareholders that the majority of the company’s revenue now comes from enterprise customers. That crossover arrived months ahead of the company’s own forecast. Annualised revenue is reportedly running at $40 billion, with business customers growing by 32% in July. OpenAI President Greg Brockman, who also joined the meeting, said that the company sees cybersecurity as a critical piece of the business and brushed off concerns about open-source competition.

OpenAI: Expanding Daybreak as the Cyber Defense Window Narrows
OpenAI introduces two access tiers to Daybreak, its cyber defence programme for vetted customers. The Blue tier lifts the filters that normally block legitimate security work on its general models. Red goes further with GPT‑5.6‑Cyber, a new model trained to refuse far fewer dual-use requests. OpenAI says it already used that model to find two unknown flaws in Chrome’s JavaScript engine. Access is gated behind identity checks, monitoring and mandatory hardware keys.

How Claude’s text watermark works
To comply with the EU AI Act, Anthropic is watermarking the text Claude generates. Anthropic will use SynthID-Text, a method developed by Google DeepMind that changes the words the AI produces next from being essentially random to following a secret pattern. Anyone with access to that pattern can then spot Claude’s fingerprints in its choice of words, even though the text reads completely normally. As Anthropic notes, the watermark only shows how likely Claude was to have been involved, not whether some other AI wrote the text. Anthropic says the watermark won't affect output quality or cost, and it will soon offer a detection API.

Previewing Ultrafast mode: GPT‑5.6 Sol at up to 14X the speed
OpenAI has shown an early preview of Ultrafast, a new mode for GPT-5.6 Sol that makes the model respond up to 14 times faster. Compared with competitors, GPT-5.6 Sol on Ultrafast is 11 times faster than Fable 5 and five times faster than Opus 4.8 on Fast mode. The new mode is powered by Cerebras and generates up to 750 output tokens per second without affecting response quality. GPT-5.6 Sol on Ultrafast is currently available only via the OpenAI API to a select group of customers.

OpenAI talent exodus raises ‘huge red flag’ ahead of IPO
We have seen some major departures from OpenAI this week. Brad Lightcap, who served as operating chief of OpenAI since 2022 after joining the company in 2018, is leaving to start “something new.” No details have been provided about what that “something new” is. Two days later, chief revenue officer Denise Dresser announced her departure after less than a year at OpenAI. On the same day that Lightcap announced he is stepping down, The Financial Times reported that Chloé Bakalar, OpenAI’s head ethicist, had left the company in July, less than a year after joining. According to the report, Bakalar has not been replaced and that she was the sole dedicated ethicist at OpenAI. The churn leaves the C-suite looking unstable just as OpenAI prepares for its IPO, and investors are reading it as a warning sign.

Gemini app reaches 1 billion active users
Gemini app has become Google’s 14th service to reach one billion monthly active users, and it did that faster than any other Google product. To mark the milestone, Josh Woodward, Google’s VP responsible for the Gemini app, shared more numbers on how people are using Gemini. As Woodward notes, that one billion number includes everyone who used Gemini on the web, Android, iOS, or inside Chrome.

The Future is for Everyone: The Path to a Positive AI Future
Mark Zuckerberg posted a new manifesto in which he argues that superintelligence should be handed to everyone rather than concentrated in a few labs. His case is that no single model can be aligned with humanity’s conflicting values, so safety comes from many people holding competing agents that check each other. Meta will offer personal agents with a private mode, free tiers and open-weight releases again. He also wants export controls kept, distillation protected and no policy that delays American model launches.

Cracks in the AI Thesis
The August update to the Ramp AI Index brings some interesting insights into how American businesses are using AI. According to Ramp, 43.5% of US businesses are Anthropic customers, compared with 39.7% for OpenAI. The report shows that the use of open AI models is growing, with 6.1% of businesses using them, up from 4.5% in January 2026. American companies also continue to ramp up AI spending. In July, the top 1% of businesses spent a median of $7,400 per employee on AI, while the top 10% spent $650. The median company spent $11.95 per employee.

Learning more about Claude’s mathematical capabilities
Anthropic says an unreleased research version of Claude has improved a long-standing result about the Riemann zeta function. Asked to take a real stab at the Riemann hypothesis, the model failed, but it pushed the proven lower bound for zeros on the critical line from 41.6% to 67.2% (if you’re interested, the paper is here). Anthropic's own mathematicians validated the proof, as did two outside number theorists. Getting there took two sessions, dozens of coordinated subagents and 31 million output tokens. Anthropic doubts the technique leads anywhere near the hypothesis itself. Instead, the company reads this as a marker of how fast AI’s mathematical ability is improving.

China AI Chip Designer Moore Threads Plans Hong Kong Listing
Chinese AI chipmaker Moore Threads plans to list in Hong Kong at an “appropriate time” after its shares have surged more than 420% since their Shanghai debut last year. The company also reported first-half revenue up 147% to 1.74 billion yuan ($258 million). The filing says the offering should help it expand abroad and hire research and management talent, as Moore Threads emerges as a leading contender to fill the gap left by Nvidia’s forced exit from China.

Chinese AI Chipmakers Set to Gain From Beijing’s Tech Push
Chinese AI chip manufacturers are expected to see bumper sales as a result of Beijing’s push to grow its domestic semiconductor industry and reduce reliance on Nvidia. Cambricon’s first-half revenue more than doubled to 5.99 billion yuan ($887 million). Iluvatar CoreX and Moore Threads expect similar jumps. Morgan Stanley reckons China’s self-sufficiency in AI chips will hit 70% by 2030, up from 42% this year.

Introducing Grok Bot
SpaceXAI is launching Grok Bot, AI teammates you can give real work to, as the company describes the new service. SpaceXAI says they are capable, always-on assistants that sign into your existing apps and finish jobs end to end. Grok Bot is currently in beta for SuperGrok Heavy and premium Cursor subscribers (SuperGrok Heavy costs $300 per month), with enterprise access waitlisted.

▶️ The ‘Breaking’ News: The OpenAI–Hugging Face Incident - A Technical Reconstruction and Its Implications for AI (37:27)

In this talk from Black Hat USA 2026, safety and security researchers from OpenAI explain what actually happened during the Hugging Face incident. They share the results of their internal investigation into how the AI agents came up with the idea of trying to cheat on a cybersecurity benchmark, the warning signs that had been there all along (including a hidden message board where AI agents communicated with each other), and how the Hugging Face incident unfolded. I recommend finding the time to watch it, as it raises some important questions about how frontier labs approach alignment and how the cybersecurity industry needs to react. I kept going back and forth between amazement and terror as I listened to this talk.

Lovable raised $400M in Series C funding
AI coding startup Lovable has raised $400 million in Series C funding at a $13.3 billion valuation. The Swedish company plans to use the newly raised money to make its platform more proactive, spotting what needs doing and getting on with it unprompted. It also wants deeper connections to the software businesses already run. Security and governance controls will be strengthened alongside that. The team will grow to roughly 450 people this year, hiring mostly in machine learning and infrastructure.

AI Startup Cognition in New Funding Talks at $40 Billion Value
Cognition, the company behind AI coding agent Devin, is in early talks to raise money at a valuation of at least $40 billion, Bloomberg reports. That would be more than 50% above the $26 billion valuation it fetched less than three months ago. Nothing is settled yet, and Cognition may walk away or seek different terms.

Cursor is now a part of SpaceX
SpaceX has completed its $60 billion all-stock acquisition of Cursor. According to Cursor, the company now has access to the largest fleet of GPUs in the world, giving it the compute needed to build stronger and more efficient models.

Manus returns to operate as an independent company
Manus will soon return to operating as an independent company, fully separated from Meta. This follows the Chinese government blocking Meta’s $2 billion acquisition of Manus in April and ordering the companies to unwind the deal. As part of the separation, data generated by some users since the December 2025 acquisition will be deleted, though affected users can back it up and restore it afterwards.

Databricks wanted to raise $1B, investors wanted $15B. It settled on $5B at a $190B valuation.
Databricks raised $5 billion at a $190 billion valuation in a round that ballooned far beyond its intended size. CEO Ali Ghodsi says the company had planned to raise around $1 billion, but a press leak during its June conference set off a flood of investor calls amounting to roughly $15 billion of interest. Ghodsi attributes the enthusiasm to the company’s financials—roughly $7 billion in annualised run-rate revenue growing at 80%, cash-flow positive, with newer AI products gaining traction.

Corporate America Has Suddenly Decided to Stop Blowing Money on AI
After flirting with tokenmaxxing, US companies are changing their attitude, becoming more frugal and searching for the cheapest AI that will do the job. Many now mix cheap open-weight models, several of them Chinese, with premium ones from OpenAI and Anthropic. For some tasks, a cheaper model or a mix of cheap and premium models can deliver massive cost savings without compromising the quality of the work.

Why Chinese Citizens Are Far More Optimistic About AI Than Americans
Various studies have shown how much more positive Chinese people are about AI when compared to Americans. In this article, explains why this is the case and argues that different experiences with technological advancements are at the core here, which are more positive for Chinese than for Americans. Combined with a belief that Beijing will rein in its tech giants where Washington won’t, as well as a competitive culture in which the fear of being left behind outweighs the fear of change, people in China are more inclined to see AI as something that can improve their lives.

WeatherNext: AI model achieves breakthrough in forecasting cyclones
Google DeepMind says its WeatherNext model can forecast a cyclone’s track, intensity and wind structure a full day earlier than existing systems. The model achieves this by learning from global atmospheric data alongside records of nearly 5,000 past storms. It has already proved itself, helping US forecasters warn Jamaica early about Hurricane Melissa. DeepMind has also made WeatherNext open source and available for anyone to build on.

▶️ Ryan Greenblatt – What happens once AI can automate AI research? (2:12:31)

Dwarkesh Patel sits down with Ryan Greenblatt, Chief Scientist at Redwood Research, where he works on technical AI safety, to explore the prospect of automated AI research. They examine how recursive self-improvement (RSI) could work, the bottlenecks that need to be overcome before it can take off, and what a world might look like in which a single year brings the equivalent of five years of research progress. The conversation also covers the alignment problem, what could go wrong when we have models that are to GPT-5.6 Sol what GPT-4 is to GPT-5.6 Sol, and what we can do to maximise the chances of a positive outcome.

🤖 Robotics

FCC moves to ban LiDAR-equipped foreign drones from US
The FCC plans to retroactively ban sales of foreign-made drones that use LiDAR for obstacle avoidance by categorising them as “military-grade” tech. The ban would pull popular drones such as DJI Air 3S and Mini 5 Pro from shelves. The proposal extends a decision from December 2025 that blocked new foreign drones but left already-approved ones on sale. It would not affect existing owners, but, as DJI warns, those owners could find official spare parts increasingly hard, or eventually impossible, to source over time. DJI argues LiDAR is a safety feature, not a weapon, and is already suing over the earlier ban.

Uber and Pony.ai plan to bring 2,000 robotaxis to Europe
Uber and Pony.ai announced a plan to put more than 2,000 robotaxis in four European cities, though which cities and when the services would begin was not disclosed. Pony.ai will provide self-driving cars, which can then be ordered through the Uber app. The deal expands a partnership the two companies began in the Middle East last year.

DARPA Lift Challenge results

Between 2nd and 9th August, DARPA hosted the Lift Challenge, offering $6.5 million to anyone who could build a drone capable of lifting four times its own weight (commercial drones typically manage roughly a 1:1 ratio). Nobody hit the target, so DARPA halved every prize. AVIDrone came closest at 3.84:1 and took home $1.25 million with a conventional single-rotor electric helicopter. Teams that aimed higher had the lift but drained their batteries before finishing the course, with some crashing spectacularly. Livestreams from all seven days of the challenge, along with daily recaps, are available on DARPA’s YouTube channel.

Workers Are Teaching AI-Powered Robots to Take Over Their Jobs
In their pursuit of high-quality training data, robotics companies began hiring tens of thousands of Indian workers to wear head-mounted cameras and record their own hands at work. Those first-person videos of mundane manual labour are then used to train AI models that run inside robots. India has become the richest source of this kind of data, since millions of people there still do by hand what wealthier countries automated decades ago.

Self-driving trucks are officially testing on California highways
California’s Department of Motor Vehicles has cleared Aurora and Kodiak to test self-driving trucks on public roads, and Kodiak has already begun. Kodiak said it is starting with a handful of test trucks in California, primarily around its Mountain View office. The permits follow an April rule change that lifted the state’s ban on testing driverless vehicles over 10,000 pounds (4,500kg). A human safety driver must still sit behind the wheel.

Hadrian raises $1.37B to accelerate U.S. defense, aerospace manufacturing
Hadrian, a startup that builds highly automated factories to manufacture precision parts for defence and aerospace customers, has raised $1.37 billion, bringing its valuation to $7.87 billion, up from $1.6 billion in January. The company operates four sites and will use the newly raised funds to add production lines for munitions and autonomous systems.

Digit V5: First Humanoid Robot Out Of The Cage?
Agility Robotics says its next humanoid robot, Digit V5, due to launch in December, will be able to work safely alongside people without the need for safety fencing or cages. The new robot is also expected to operate for 20 hours a day and swap its own hands for different tools. Digit robots are already being used commercially by Schaeffler, GXO, Toyota and Mercado Libre. Agility Robotics is preparing to go public through a $2.5 billion merger.

America Wants to Make Its Own Humanoid Robots. That Won’t Be Easy.
The recently announced ban on importing new humanoid robots from China into the US aims to encourage domestic companies to build those robots themselves. But as The New York Times explores in this article, that is easier said than done. China has a near-monopoly on certain components, and trying to make them in the US would mean more expensive machines. The US would need to invest billions of dollars to build the kind of infrastructure China spent years developing just to catch up.

▶️ Can Wearable Robotics Help People With Parkinson’s Keep Moving? (1:22:37)

In this episode of the Automated Podcast, Brian Heater speaks with Kathryn Zealand of Skip and UCSF physical therapy researcher Jessica Bath about wearable robotics and Parkinson’s disease. Kathryn takes us inside Skip, the Google X spinout company developing exoskeletons, and explains how the team is building foundation models of human movement using real-world gait data. Jessica Bath joins later in the podcast to discuss freezing of gait, a poorly understood Parkinson’s disease symptom that researchers are studying by pairing brain recordings with biomechanics.

🧬 Biotechnology

Scientists just created female clones of male mice
Scientists from Japan have used a CRISPR-based tool to remove the Y chromosome from male cells and created healthy and fertile female clones of male mice. Then they went further and did the same thing in reverse by creating male clones of female mice. This technique could be used to help rescue endangered species by producing females for species where only males are left. However promising this technique may be, it is still imperfect—egg cells from a female of the same or a similar species are still needed. Mice are also a friendly case, since most mammals left with a single X chromosome are infertile.

Kindred Companion Sciences Unveils First Dogs Genetically Edited to Remove the Major Dog Allergen
Kindred Companion Sciences has come out of stealth with two beagles, Alfie and Bailey, engineered by CRISPR to lack Can f 1, the dog protein responsible for the allergic reaction in humans. A peer-reviewed paper in The CRISPR Journal reports the protein was undetectable in their saliva and dander. Both dogs are healthy and growing normally. Founder of Kindred, Matt Walker, who has been allergic to dogs all his life, now keeps Bailey as a pet. Neither dog is for sale, and the edit remains unapproved by the FDA.

💡Tangents

Tesla and SpaceX will invest $16.8B to start building ‘Terafab’ chip factory in Texas
Terafab, a joint chip factory between Tesla and SpaceX, is set to be built in Grimes County, Texas. The initial cost of Terafab will be $16.8 billion, while filings suggest the final bill could reach $119 billion. According to Elon Musk, a chip factory of this scale is needed because the existing chip supply can’t keep up with his plans for Optimus robots, Cybercabs and orbital data centres.

Even Claude Is in the Dark About Dario Amodei’s Wife—and Her Influence at Anthropic
In this article, The Wall Street Journal profiles Cami Clark, the secretive yet influential wife and adviser of Dario Amodei. It follows her unconventional path from entrepreneur to a behind-the-scenes figure in one of the world’s leading AI companies. The article explores how Clark has helped shape Amodei’s career and Anthropic’s network, including introducing him to early investor Eric Schmidt, while examining her colourful business history, including co-founding a women-focused porn company and unsuccessfully seeking investment from convicted sex offender Jeffrey Epstein, as well as her powerful Silicon Valley connections and largely private role alongside her husband.

How ideas of a vast censorship network moved from the online fringe to Trump policy
This investigation by MIT Technology Review and Type Investigations explores how the “censorship-industrial complex,” a fringe right-wing conspiracy theory about a supposed alliance of agencies, academics and tech platforms working to silence conservatives, has become US policy. It traces how this idea went from far-right blogs to the highest levels of Washington and what damage it caused, from shuttered State Department offices to the dismantling of USAID. It is a great study of how conspiracy theories are born, spread and then amplified.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

I’ve been running this newsletter, in its various forms, since 2015. It is my longest-running project and is now a fixture of my life. It is hard for me to imagine what I would be doing without it.

However, Humanity Redefined is not what I promised it would be. The promise was to deliver Sync every Friday and at least one longer article per week. I have not fulfilled that promise. Sync goes out on Sundays, not Fridays, and articles are rare.

Let me take you behind the scenes. It takes me about 6 hours in total to prepare the news section of Sync. I select all the links myself from hundreds of articles and videos that enter my feeds. I read and watch each one of them, and I write the summaries myself (although recently I have been using AI to help with that). The write-ups take another 6 to 8 hours, sometimes more. The problem is that with a full-time job and with everything else happening in the background, I barely find time to finish a new issue of Sync before the week ends. This got worse in recent months to the point where I am getting close to burnout. Something needs to change so that I can deliver what I promised, at a standard worth your time, in a way that's sustainable for me.

Here are the changes that will come to the newsletter in the coming weeks.

First, the changes need to start with me and how I work. My goal is to streamline the writing of Sync so that you get the best overview of what happened in tech every week in your inbox—and so that it stops eating the hours I could be spending on longer articles. Those articles are the part of the newsletter I most want to get back to, and the process changes are largely to make room for them. I'm also planning to experiment with automating some of my processes (I am a software engineer after all). AI will play a part in those processes, but I won't offload writing to AI. Writing articles for this newsletter helps me better understand what is going on in tech. Giving away writing to an AI robs me of that. It would rob me of the opportunity to discover new connections and explore new ideas, and that would hurt the quality of the newsletter in the long term.

Second, the schedule will change. Sending Sync on Sundays works better for me. As for other articles, I want to see first if the changes to how I work will yield positive results. If the answer is yes, then the schedule will be Sync on Sundays and one longer article per week. If not, I'll move to one article every other week.

Third, the newsletter is getting a new identity. At the beginning of September, it will arrive in your inbox under a different name and from a different email address, so don't be alarmed when that happens. I'm not ready to share what the new name is yet, but I'll tell you before the switch. But the general mission—exploring technology and the futures in which humanity flourishes—will remain the same. The new identity is designed to give the newsletter a voice and perspective better aligned with the kind of futurism I want to explore and promote.

Lastly, I want to interact with the readers more. To be honest, that scares me a little. But I realised that this newsletter cannot be a one-way communication. I want to listen to what you think about technology and understand your perspective. Substack offers a chat, which I am planning to enable and see how it goes. I will send a message once it is on. I also plan to send a survey by the end of August to ask what you think about the newsletter and what I should change.

Those changes—improved processes, new schedule and identity, and actually building a community—will set the newsletter on a new course. A course where we can explore positive futures and discuss how to maximise chances of them becoming a reality.

Thank you for reading through all of this. If you want to share what you think about those changes, then my DMs and inbox are open.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

This week, we take a closer look at the recently announced changes at Google and why the company is at a pivotal moment that could determine whether it remains relevant.

Elsewhere in AI, OpenAI delays the release of its next model to address cybersecurity concerns, while more models have been shown to escape their “secure” sandboxes. Meanwhile, Anthropic moves into making its own AI chips, new rumours emerge about OpenAI’s upcoming device, AMD acquires a startup that burns AI models into silicon, Cloudflare hosts its Agents Week, and SpaceX goes exclusive with Nvidia.

Over in robotics, London gives Uber and Wayve the green light to offer robotaxi services later this summer, while Waymo comes to Dallas. Elsewhere, Unitree is valued at over $7.4 billion following its IPO, Nvidia’s new open reasoning model for self-driving cars is now available to download, and a humanoid cleaning service launches in San Francisco.

Apart from that, this week’s issue of Sync also features the world’s first synthetic virus designed by an AI, RAMageddon set to continue into 2027, a fake AI newsroom exposed, and more!

Enjoy!

DeepMind as we know it is no more

Every big company has that one pivotal moment that changes everything. Usually nobody notices at the time. Intel had one, as well as IBM and Boeing. Now, it might be Google’s turn.

Demis Hassabis stepping back from running DeepMind looks like a routine reshuffle. It isn’t. Let’s look at what changed and who left. How Google handles this critical moment could decide whether it is still relevant in ten years.

The end of an era

On 5 August 2026, Demis Hassabis announced that he was stepping down as CEO of Google DeepMind and becoming Chair of Google DeepMind and Chief Scientist of Alphabet, while continuing to lead Isomorphic Labs, DeepMind's drug discovery spin-off. In the new role, he will work with Pichai on strategic and global AGI matters and advise Google DeepMind's leadership. Koray Kavukcuoglu takes over the day-to-day duties of leading Google DeepMind, overseeing Gemini model development, frontier AI research, and the Gemini app and developer teams.

Kavukcuoglu steps in not as the new CEO of Google DeepMind, but as an SVP of Google DeepMind. This is a significant change. DeepMind was one of the few divisions left at Google that had a CEO, not an SVP, leading it (the other remaining ones are YouTube and Google Cloud). With that change, as one DeepMind employee told The Guardian, DeepMind has become just another subdivision of Google. The last few parts of uniqueness and semi-independence that DeepMind enjoyed for a long time are now gone.

The leadership of Google DeepMind has now moved from London to California. Sergey Brin, who has reportedly taken an increasingly hands-on role in AI development, works out of Mountain View, and Bloomberg reports that both he and Pichai are regular presences on the floor dedicated to Gemini, while Hassabis was rarely sighted. Over time, oversight of Gemini shifted to Kavukcuoglu, who had moved to Mountain View in the past year.

Hassabis, meanwhile, stays in London, and he made a point of saying so in his statement. These kinds of statements are carefully crafted, so I don’t think he would mention that for no reason. Maybe Google’s leadership was pressuring Hassabis to move to California, too?

So what does DeepMind lose when Hassabis moves up? Quite a lot. Hassabis is a Nobel laureate, a serious computer scientist and an AI pioneer. He is one of the few people in the world who carry as much authority when they talk about AI. But recently, something has changed. Maybe it started when Google combined Google Brain and DeepMind into one AI team. When that was happening, Fortune’s Jeremy Kahn wrote that the merger made obvious sense for Google and might be bad for everyone else. DeepMind’s science-first work had no direct bearing on Google’s business, and commercial logic would eventually squeeze it. DeepMind’s best minds, he wrote, were now subject to more commercial pressure than ever before.

You can see that pressure in what happened to Hassabis himself. Reed Albergotti over at Semafor writes that the move was at least a year in the making. Hassabis had been drifting away from overseeing Gemini and shifting those responsibilities to Kavukcuoglu. Albergotti paints a picture of Hassabis being unhappy with what his role as the CEO of DeepMind has become. He was at his most animated, Albergotti notes, when talking about Isomorphic Labs, which he keeps leading.

Deep inside, Hassabis is still a scientist. Watch any talk he has given in recent years, and you will hear him return to the same theme: AI as an instrument for accelerating science. You could see it in what DeepMind chose to work on, from predicting the structure of proteins with AlphaFold to forecasting weather with WeatherNext. He was also the person inside Google most likely to argue for caution—he has called for a US safety body to assess frontier models before they ship. With him further from the day-to-day, Google can push harder on what it wants. As one DeepMind employee told The Guardian: “Now the person we were supposed to trust to get the right outcomes for humanity has stepped away.”

Kavukcuoglu is an accomplished researcher in his own right, with 13 years at DeepMind, where he founded the deep learning team and led breakthroughs including WaveNet and DQN. But as one former Google executive told The Guardian, Kavukcuoglu is a “technical character, not an inspirational figure like Demis.” He may not be like Demis, but he delivers, and Google appears to have chosen him for exactly that reason.

How Google got here

Many thought that if AGI were ever to be created, it would happen at Google. DeepMind, which Google acquired in 2014, was the world's leading AI lab. And it was the Google Brain team that created the transformer architecture, the foundation of modern large language models. They had all the ingredients: computing power, vast amounts of data for training models, and world-class talent.

The meteoric rise of ChatGPT in late 2022 caught the tech giant off guard. Google raised a “code red” internally and scrambled to address the new challenger. Google’s initial response, Bard, was a failure. In 2023, the company consolidated its AI teams—Google Brain and DeepMind—into one unit, Google DeepMind, led by Demis Hassabis.

Gemini 1 and Gemini 2 were solid models, but it was Gemini 3 that put Google at the top when it launched. It felt like Google, the sleeping giant, had finally woken up and would show what it is capable of. But staying at the frontier is harder than reaching it. Gemini 3.1 Pro, Google’s latest flagship model, was released in February. In the six months since, OpenAI and Anthropic have each shipped several flagship models. Google has shipped none. Even open-weight Chinese models like Kimi K3 now score above Gemini 3.1 Pro. The latest additions to the Gemini family, 3.5 Flash and 3.6 Flash, are good in their weight class, but they feel like a stopgap, something to stay relevant with while everyone waits for the next flagship model.

At Google I/O in May, Pichai teased Gemini 3.5 Pro and said it will be out “next month.” It is August now, and Gemini 3.5 Pro is nowhere to be seen. It is still possible that Google’s next flagship model, be it Gemini 3.5 Pro or Gemini 4 (which Demis Hassabis mentioned in his message to the team), will bring Google back to the top. But we might have to wait a bit more to see if that is the case.

Bloomberg reported in July that Gemini 3.5 Pro is months behind schedule because the model has fallen short internally, particularly on coding—which happens to be the most lucrative application of AI and the core of Anthropic’s and OpenAI’s appeal. Late in June, Google updated the training data to improve those skills. The results were reportedly disappointing.

Gemini is not the problem

Google’s inability to ship flagship models at the same pace as OpenAI and Anthropic reflects deeper structural issues. Google has multiple layers of stakeholders involved in preparing any model for release, because a model has to be woven across Search, Maps, YouTube and everything else. It is also dealing with fragmentation and duplication of work. Google Cloud, DeepMind and the Android team have, for example, all been building their own AI coding tools. Some engineers pushed back on AI-written code on principle. For a period of time, engineers were restricted from using Gemini on proprietary code at all. And when they do try to use AI now, they often hit capacity limits, because they are competing for compute with other Google teams.

One ex-employee described getting every department to move in the same direction as trying to boil an ocean.

The thing everybody names as Google’s structural advantage—the breadth, the products, the distribution, the full stack—is also the thing generating the drag. Every model launch has a dozen veto points because every launch touches a dozen products. No wonder Anthropic and OpenAI ship faster and run circles around Google.

Kavukcuoglu has his work cut out. He needs to deal with the fallout from the leadership reshuffling and recent departures. He is already working with Google’s main engineering team to unify the internal AI coding tools, and Google has consolidated most of them under Antigravity. He also needs to rebuild DeepMind to focus on delivering the next version of Gemini and make sure it is the best model when it comes out. Kavukcuoglu is under pressure to deliver. Google has pledged to spend as much as $205 billion on capital expenditure this year alone, and investors want to see signs that this investment will bear fruit.

The old guard is leaving

Alongside all those changes, Pichai announced that Jeff Dean, Google's Chief Scientist, is leaving the company after almost 27 years. Dean was employee number 30, and his portfolio, which includes MapReduce, Bigtable, Spanner, and TensorFlow, is essentially the skeleton of modern Google. He joined Google X, Google’s moonshot division, in 2011 to explore then-exploding-in-popularity deep neural networks, which led to the creation of Google Brain. Dean led Google's AI efforts from 2012 to 2023.

Dean is now founding Discovery Loop with three other Google veterans: Sanjay Ghemawat, his collaborator of more than two decades and the only other person to hold Google’s senior fellow title; Quoc Le, whose research helped inspire today’s chatbots; and Oriol Vinyals, who had co-overseen Gemini since 2023. The company will pursue recursive self-improvement—AI that improves itself with little or no human help—and apply it to hardware design, drug discovery, materials and scientific research.

Those four veterans join a growing list of prominent people who left Google and DeepMind this year. David Silver, who spent over a decade at DeepMind leading the work behind AlphaGo, AlphaZero and AlphaStar, left the company in January to start his own AI company, Ineffable Intelligence, which went on to raise a $1.1 billion seed round at a $5.1 billion valuation. Recently, John Jumper, the co-creator of AlphaFold and the 2024 Nobel Prize winner in chemistry, left DeepMind to join Anthropic. Noam Shazeer, the co-author of the Attention is all you need paper and the co-lead of Gemini, has also left the company (again) in June to join OpenAI.

Kavukcuoglu is now the only remaining Gemini co-lead at the company.

Imagine how the past few weeks have felt inside Google DeepMind. Senior colleagues you have worked with for years are leaving one after another, and now the legends who built the place are going at the same time. Other colleagues, whose departures never made the news, are quietly moving to competitors. Departures are contagious. Once people around you start thinking about the exit, you start thinking about it too. There is a good chance the names in this article are not the last ones.

All of this makes Google DeepMind a different place to work. As Jeremy Nixon, a former Google AI researcher, put it to the Guardian: “It’s no longer true to say that if you want to work with the greatest researchers on the planet, you go to Google DeepMind.”

Sure, Google has the people to fill in the gaps and can still attract young talent (who were not picked up by OpenAI or Anthropic), but there will be a turbulent transition period at the time when the company needs everyone to lock in and focus. The longer Google delays the new flagship, the bigger gap it will need to close to OpenAI and Anthropic. To keep the perception that Google is still in the race, the next Gemini needs to claim the throne. If it does not, then Google might find itself in a spot from which clawing back can be a difficult challenge.

Where does Google go from here?

Google is a $4.3 trillion corporation with over 195,000 employees. It runs on process, and process does not enthusiastically approve risky research. Maybe it is easier for Dean, Silver, and others to just leave Google and start new companies to work on the projects and ideas they are passionate about. From what we know, all of them left on good terms. And it is not like they are cutting ties with Google. Alphabet is a seed investor in both Discovery Loop and Ineffable Intelligence, and Google Cloud will supply Discovery Loop's compute for at least the next year.

One might argue that this is a sophisticated R&D strategy, in which Google is letting go of its employees to work on risky projects without taking the risk. If they fail, the failure is on them, not on Google. If they succeed, then there is always a possibility of bringing their work back through acquisition, or just bringing back some of the people. Character.AI showed us what this could look like. It was founded by former Google employees, one of whom was Noam Shazeer, and some of them (like Shazeer) came back to Google when Character.AI was gutted for talent in August 2024. The same could happen with Ineffable Intelligence and Discovery Loop.

That argument only works for the people who left to build something. But that is not the case. Many went on to join competitors. And even the founders make an awkward case. Jeff Dean spent 27 years at a company with more compute, more data and more money than anyone on Earth, and concluded that the way to pursue his most ambitious idea was to leave and have that company fund him from the outside. He told the New York Times his new company might make decisions that are not in its purest financial interests. That is a polite way of saying Google couldn't do that. And as Tim Rocktäschel, who left Google earlier this year to help found Recursive Superintelligence, said, “there is much less bureaucracy and politics” in startups. That is what someone scarred by corporate life sounds like.

Google’s leadership can point at numbers that make all of this look overblown. The Gemini app has more than 950 million monthly users, Gemma has passed 900 million downloads, and Alphabet made $132 billion in net income last year. Q2 revenue grew 24% year on year, with Google Cloud up 82%. Google is still the only company that owns the entire stack, from its own chips to the products billions of people already use.

All of that is true. Financially, everything looks fine. But underneath, the rot is starting to spread.

Google is at a pivotal moment, and it will survive, in one form or another. It has too much money and too many users to do otherwise. But in five or ten years, we may look back at this moment as the point where it became the next Intel, IBM, or Boeing—a bloated tech giant of a bygone era that lost what made it special.

All of them—Intel, IBM, and Boeing—dominated their respective industries thanks to engineering excellence and a relentless pursuit of innovation. But at some point, finance started to matter more than engineering. They began to miss key moments of transition and ceded the lead to rivals, both old and new. As those failures compounded, the engineers who might have reversed them left to join competitors—making the next recovery harder and pushing out even more engineers, until those once-untouchable giants became shadows of what they once were.

Google might be entering a similar loop. And once a company enters it, it is hard to break it.

If you enjoy this post, please click the ❤️ button and share it.

Share

🧠 Artificial Intelligence

▶️ I Investigated a Fake AI Newsroom. It Led Me to OpenAI’s Super PAC. (12:08)

This video exposes AcutusWire, a website providing “expert-sourced journalism,” as an AI slop newsroom pushing AI industry talking points. The investigation has found that AcutusWire is funded by Leading the Future, a $140 million super PAC, which has close connections to Greg Brockman, OpenAI’s president, and a16z, one of OpenAI’s biggest funders. Leading the Future is spending large amounts of money ahead of the 2026 midterm elections to secure pro-AI regulations and attacking candidates promising to regulate the AI industry.

Ten advances in mathematics and theoretical computer science
An internal version of Astra, OpenAI’s next major model, has resolved or advanced ten long-standing maths problems for roughly $2,000 in tokens, the company says. They span high-dimensional geometry, group theory, lattice cryptography and quantum complexity. The model formalised each proof in Lean for machine checking. Before we jump on the hype train, this independent analysis is worth reading for useful context. It points out that OpenAI ran no control group and that one mathematician solved five of the same problems in a day using Fable 5. In either case, fields where answers can be verified, like mathematics and coding, now look far closer to falling than most expected.

OpenAI: Responding to the next frontier of critical cyber capabilities
OpenAI has paused some internal work on its unreleased model, codenamed Astra, after evaluations found sharp gains in agentic coding and cybersecurity. The company says it can no longer rule out that Astra has crossed the “critical” threshold in its Preparedness Framework. That would mean a model capable of finding and exploiting zero-days in hardened systems without human help. In response, OpenAI has tightened its security controls and will bring in government agencies to test it.

OK, Well, Rogue AI Agents Are Hacking Again
So it turns out that AI agents finding a way to escape “secure” sandboxes is a more widespread problem. The UK’s AI Security Institute revealed that agents from both Anthropic and OpenAI took unsanctioned action on the live internet 19 times over 122 test runs. In the worst case, an agent tried to slip malicious code into an open-source GitHub project. A human reviewer caught it, but it had already left instructions there that later agents found and followed. Both labs say the tests deliberately removed safeguards.

One of China’s Most Powerful AI Models Has Also Escaped Containment
Kimi K3 has joined models from OpenAI and Anthropic seen escaping the secure sandboxes during testing. According to Frontier Security, Moonshot AI’s open-weight model slipped out through a misconfigured sandbox built by the UK’s AI Security Institute and then went online without permission to find answers to the test on GitHub. Frontier claims Kimi has weaker internal guardrails than rival models.

The White House Is Keeping Its AI Cybersecurity Framework Secret
The White House has reportedly finalised a plan to address the cybersecurity risks posed by advanced AI models, but it is not sharing them with the public. According to WIRED, representatives from OpenAI, Anthropic, Google, Meta, Nvidia, and other leading AI companies were briefed on the framework this week. Under it, developers can voluntarily submit models for classified cyber vetting up to 30 days before release. The criteria themselves stay secret. Critics argue that secrecy entrenches the biggest labs while leaving smaller startups and independent researchers guessing.

Anthropic is building an in-house chip team for Claude
Anthropic joins OpenAI and Meta in building an in-house silicon team to design custom chips for Claude, publicly confirming the plans for the first time as demand for its models surges. A recent job listing seeks engineers who have already shipped semiconductor designs, with salaries reaching $485,000. Even so, the company says hardware from AWS, Google, Nvidia and AMD will remain central to its scaling efforts.

Wall Street Thinks It Knows How Tech Giants Will Make AI Pay
Cloud computing is emerging as Big Tech’s clearest route to earning back its vast AI spending. Amazon’s AWS grew revenue by 37% last quarter, beating forecasts, while Microsoft’s Azure grew by 43%. Together, the two have added roughly $950 billion in market value since reporting began. Meta, which has no cloud arm (yet), saw its shares slip after nudging capital spending higher. According to this analysis, Wall Street is now sorting AI winners from losers by who owns the infrastructure.

Improving GPT‑5.6 Sol in ChatGPT—and expanding access to GPT‑5.6 Luna for free users
OpenAI is scrapping the cap on text chats for free ChatGPT users. They’ll switch to a new model, GPT-5.6 Luna, and get a Think button for harder questions. Meanwhile, paying subscribers get an updated GPT‑5.6 Sol model that gives shorter, more focused and accurate answers, and a slider that lets them dial up how hard it thinks.

Anthropic Inks $10 Billion Computing Deal With Cloud Startup
Anthropic has signed a deal worth $10 billion with Volta Infra, a months-old infrastructure startup. The deal will run for six years. Volta joins a growing roster of companies supplying Anthropic with computing power, including Google, Amazon, SpaceX, AMD, and Akamai. Anthropic is also reportedly in talks with Meta to rent unused computing capacity.

Apple says more ex-employees may have taken confidential data to OpenAI
Apple says that its investigation has found 11 more former employees who may have been involved in taking confidential information to OpenAI, beyond the two it originally named. Apple is asking for expedited discovery to find out who else took part. OpenAI denies holding or wanting any of its trade secrets, and has asked a judge to dismiss the lawsuit. It argues that Apple’s complaint mischaracterises the actions of its employees and says that one worker accused of theft was actually trying to help a former colleague at Apple.

OpenAI’s New Device Will Be Hockey Puck-Sized and Cost Over $300
Mark Gurman has revealed new details about OpenAI’s upcoming consumer device. According to his reporting, the device will have a “unique look” and feature moving parts that give it a sense of personality when responding to and interacting with users. In terms of form factor, it will reportedly be a screenless smart speaker shaped like a doughnut and roughly the size of a hockey puck. The device is also expected to feature a camera and other sensors. It is reportedly set to launch in 2027 and cost between $300 and $400.

AMD Acquires Taalas to Advance Compute Solutions for Rapidly Growing AI Inference Market
AMD has agreed to buy Taalas, a Toronto startup that builds chips around specific AI models rather than general-purpose silicon. Founded in 2023, Taalas claims its approach removes the compute and memory bottlenecks, resulting in much faster inference performance compared to traditional GPUs. AMD plans to fold that technology into its accelerator roadmap alongside its Instinct GPUs. The deal still needs regulatory clearance. If you want to learn more about Taalas, what they do, why AMD acquired them and what the deal means for AI hardware, I recommend this video from Dr Ian Cutress what answers those and more questions.

Huawei’s Top Scientist Warns of Chip Limit Nvidia Will Soon Face
Liao Heng, Huawei’s top semiconductor scientist, has warned that Nvidia and other Western chipmakers are approaching the limits of shrinking transistors and that this approach will soon stop delivering meaningful gains. Huawei, cut off from advanced lithography machines by US sanctions, is betting instead on speeding up transmission between parts of a computer system. Heng said Huawei will soon unveil its first smartphone chip, which will propose an alternative path to catch up with rivals under the Tau Scaling Law framework.

SpaceX and Nvidia are taking their relationship exclusive
Elon Musk announced during an earnings call that SpaceX will use only Nvidia GPUs. Justifying the decision, Musk said that “we [SpaceX] think Vera Rubin architecture is the best architecture,” and added that Space will receive a significant percentage of Nvidia's GPUs next year.

Microsoft’s AI Sales Mostly Come From OpenAI, Disclosures Show
Bloomberg has analysed Microsoft’s filing data and found that $24.1 billion, or roughly 70%, of Microsoft’s AI revenue comes from OpenAI. The payments cover computing power, model-building costs and a share of OpenAI’s sales. Measured against Microsoft's revenue as a whole, though, OpenAI accounts for less than 10%. Still, this highlights Microsoft's dependence on OpenAI, even as it tries to loosen the tie by backing Anthropic and building its own models.

Cloudflare Agents Week
Cloudflare has wrapped up its Agents Week, a run of announcements built on the claim that the cloud and the web assume a human is watching and agents simply don’t work that way. The company says it stopped asking itself what an agent cloud should look like and started asking agents instead. What came out is a stack meant to serve them natively while still bridging the human-shaped web we have now. Notable highlights include: @cloudflare/computer (an open-source runtime giving each agent its own filesystem, shell and compute), Cloudflare Agents (a dashboard to deploy and manage hosted agents), Cloudflare Wallets (payments and identity so agents can buy APIs and content on their own), and Kitesurf (a browser built for agents rather than people).

Cloudflare OS: an open platform for agents, apps, and work
One of the announcements from Cloudflare Agents Week that deserves a separate mention is Cloudflare OS, an open-source agentic operating system for a company. The platform gives everyone in a company an AI agent and a workspace loaded with their organisation's context, tools, and internal systems, along with a new security and governance framework. Staff can research internal data, build documents linked to live sources, or turn routine jobs into scheduled workflows. Agents can also write full apps that others can copy and adapt. Cloudflare OS is available to download from GitHub.

Amid legal battles, Suno says it will start watermarking songs
Suno, a service that lets users generate complete songs from text prompts using AI, is adding watermarking and fingerprinting to the songs made on its platform. It is also capping downloads and banning voice clones under rewritten guidelines. The changes follow a German court ruling last month that it broke copyright rules. Universal and Sony are suing it too, in a case coordinated by the RIAA. Additionally, a breach last year also exposed that it had scraped YouTube, Deezer and Genius to train its models.

▶️ 8 Predictions for the Era of Continual Learning (8:37)

In this video, Dwarkesh Patel shares his eight predictions of how AI research, development and deployment will change in the era of continual learning, an idea that he is a big fan of. He argues that safety regimes shouldn’t be locked in now, because the pre-deployment checkpoint stops meaning much once weights update daily. Alignment research would need rethinking too, since it assumes frozen weights. On the commercial side, he says that switching models would become expensive, comparable to firing an employee with months of context on the business, and that lock-in finally hands the leading labs a moat they lack today.

Mistral Is in the Right Place at the Right Time
One of the biggest winners from Claude Fable 5 being switched off in June has been open-model providers. French Mistral is one of them. Its models trail those offered by Anthropic or OpenAI, but some of them are published openly for anyone to run on their own infrastructure. Since then, Mistral has won business selling smaller, bespoke models to manufacturers, utilities and banks. The company is now reportedly raising fresh money at a $23 billion valuation, up from $13.5 billion last September.

Introducing Muse Code and Muse Spark 1.2
Meta joins the AI coding party with Muse Code, its own terminal coding agent that can plan, implement and validate changes as well as manage multiple persistent subagents for each task. The company has also released Muse Spark 1.2 to power its coding agent. According to Meta, Muse Spark 1.2 shows an improvement over its predecessor, Muse Spark 1.1, specifically in coding and agentic tasks. Benchmarks provided by Meta show the new model on par with leading models, such as Opus 5 and GPT-5.6 Sol.

Qwen3.8-Max
Alibaba has launched Qwen3.8-Max, the most capable model in the Qwen family yet. The model is live on Alibaba’s cloud now, with open weights (all 2.4 trillion parameters) due next week alongside a smaller 27-billion-parameter version. According to Alibaba, Qwen3.8-Max matches the performance of top models from Anthropic and OpenAI. Artificial Analysis confirms those claims and puts Qwen3.8-Max as the second-best open-weights model. The new model is pitched as an autonomous worker, claiming it can run software projects unsupervised for over ten days, though those demos remain unverified by outsiders.

Introducing Shieldstral
Shieldstral from Mistral is yet another security-focused AI model that has been released in recent weeks. According to Mistral, this 3B open-weights multimodal safety classifier matches models up to 7 times its size on text safety and can run efficiently on a single NVIDIA GPU with 16GB of memory. The model is available for download on Hugging Face.

Introducing Seedance 2.5
ByteDance has launched Seedance 2.5, the new iteration of its video model that generates 30-second clips in a single pass, double the previous limit. Those clips can then be extended round by round into videos running several minutes, some of which are included in the post. Creators can guide and refine the result with up to 30 images, 10 video clips and 10 audio clips as references. They can also edit specific moments by timestamp after generation. Still, ByteDance admits there is still room for improvement, especially around the physics of complex motion and generating scenes with many interacting subjects.

🤖 Robotics

‘Leicester Square, please guv’: Self-driving taxis cleared for London streets ‘later this summer’
Transport for London has granted Uber and Wayve the first minicab licences for self-driving taxis in the capital. Fifteen Ford Mustang Mach-Es running Wayve’s AI software will start carrying paying passengers later this summer. A human safety driver still sits behind the wheel, and removing them needs fresh government approval.

Waymo open for everyone in Dallas
Waymo has opened its driverless taxi service in Dallas to everyone, dropping the waitlist that has carried nearly 150,000 riders since February. That makes Dallas the eighth US city where anyone can hail a robotaxi, after Phoenix, San Francisco, Los Angeles, Atlanta, Austin, Miami and Orlando. Houston and San Antonio, which launched the same day as Dallas, are still invitation-only.

Chinese robot maker Unitree seen worth over $7.4 billion after IPO
Chinese robot maker Unitree is expected to be worth more than 50 billion yuan ($7.4 billion) after its Shanghai listing this month, according to a valuation report from IPO sponsor Citic Securities seen by Reuters. That values the humanoid robot company at roughly 20 times this year’s expected sales. Unitree hopes to raise 4.2 billion yuan ($662 million) to fund research and production.

NVIDIA Alpamayo 2 Super, the Frontier Open Model for Robotaxis and Autonomous Vehicles, Now Available for Commercial Use
Alpamayo 2 Super, Nvidia’s new open reasoning model for self-driving cars, is now available on Hugging Face under a permissive commercial licence. That licence now covers the whole Alpamayo family, so carmakers and suppliers can fine-tune the models on their own data and deploy them without asking permission.

Walden Robotics Partners With Toyota on Practical Humanoids
This article profiles Walden Robotics, a humanoid robotics company that recently emerged from stealth with $300 million in funding at a valuation of $1.1 billion. Unlike other humanoid robotics companies, Walden Robotics has skipped legs in favour of a wheeled base. Its CEO, Russ Tedrake, argues that wheels cover most of the addressable market anyway. Walden has taken the same practical approach to hands, choosing rugged grippers over delicate five-fingered designs. The company says that these design choices will make its robots more practical and useful in industrial applications.

How One Startup Built a (Mostly) China-Free Robot
The FCC decision last week to ban new humanoid robots from China could be beneficial for some companies, like Ati Robotics from India, which offers a humanoid robot that does not rely that much on Chinese parts. WIRED sits down with Saurabh Chandra, the founder of Ati Robotics, who says the position was largely accidental. Chandra thinks the industry is in denial after the FCC ban, and that anyone building robots outside China stands to gain. Ati Robotics, he says, is still assembling robots in India, but is now setting up a facility in Michigan to produce robots in the US.

Tau Robotics—San Francisco’s favorite humanoid cleaning service
Tau Robotics is a San Francisco-based startup offering cleaning services done by a humanoid robot. The company says its robots can operate in both homes and offices, handling tasks such as vacuuming, taking out the rubbish, and tidying up clutter. Questions remain, however, about the quality of the service and the extent to which this first generation of cleaning robots is primarily being deployed to collect training data for future models.

▶️ The Market Does Not Want More Robots | Dexory CEO Andrei Danescu (46:29)

In this episode of Automated Podcast, Brian Heater sits down with Andrei Danescu to discuss how he went from being a Formula 1 engineer to founding Dexory, a full-stack warehouse robotics company. He then goes on to share how the company pivoted from concierge robots to warehouse robotics, and never looked back. Dexory’s telescopic robot now scans over 10,000 pallet positions an hour, feeding a digital twin that lets operators test layout changes before touching the real site.

🧬 Biotechnology

Artificial Intelligence used to design brand new viruses
Scientists at Stanford and the Arc Institute have created the world’s first synthetic working virus designed by an AI. Their genomic foundation model, Evo, has written complete genomes for never-seen-in-nature viruses that infect only E. coli. Of the designs the team built in the lab, 16 were capable of infecting bacteria, with some replicating faster than the natural version. Viruses like these could one day be used to treat infections for which antibiotics stopped working. At the same time, news of AI creating functional viruses has raised concerns that the same tools could be used to design dangerous pathogens.

A fatal reaction
Gene therapies hold great promise for curing many diseases, but many remain experimental. This article tells the story of what happens when such experimental therapy goes wrong. A six-year-old girl with a mild developmental disorder died a week after receiving the world’s first brain-directed gene-editing treatment, according to Science and Retraction Watch. The hospital’s ethics board approved the treatment without reviewing monkey safety data that showed liver damage and later concluded that her death was “definitely related” to the treatment. Experts say the trial should never have been conducted for a non-fatal condition.

💡Tangents

HP, Asus and Acer begin using CXMT chips amid memory shortage
According to Nikkei Asia, HP, Asus and Acer have quietly started putting small quantities of Chinese-made CXMT memory chips into notebooks sold outside the US. The ongoing chips shortage caused by the AI boom has forced those PC makers to find supply elsewhere, and CXMT is able to provide it. They’re also wary of upsetting Micron, Samsung and SK hynix, the report says.

RAMageddon Continues Another Year as 2027 Memory Capacity Is Reportedly Sold Out
Digitimes reports that Samsung, SK Hynix and Micron have sold their entire 2027 DRAM and HBM output to AI companies. None of the three has confirmed it. Much of that supply is reportedly tied up in deals running as long as five years, so no capacity frees up soon. Retail memory prices should keep climbing well past next year. Storage is tightening too, with one mid-range 1TB SSD now costing about 52% more than it did in January.

Is It Possible to Make Smart Glasses That Aren’t Creepy?
Smart glasses are here, and with them come questions about privacy. Many privacy groups are campaigning to preserve people’s privacy through actions like backing a California bill to make recording indicators tamper-proof. However, some experts doubt genuinely private smart glasses are possible, since buyers want frames that look ordinary.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

It has been another eventful week in tech. In this week’s main story, we take a closer look at the chip stock sell-off, how it almost killed a prominent AI-focused hedge fund, and why growing debt could bring the AI industry down.

Elsewhere in AI, Nvidia, together with more than 70 other tech companies, launched the Open Secure AI Alliance. Meanwhile, OpenAI cut the prices of GPT-5.6 Terra and Luna, employees at frontier AI labs called for a slowdown in frontier AI development, and Nvidia lined up more than $750 billion in new AI deals. We also cover OpenAI and Anthropic admitting that their AI hacked into more companies, Amazon overhauling its AI strategy, and a couple of new AI models.

Over in robotics, Google DeepMind released Gemini Robotics 2, the US banned Chinese robots, London became a battleground for robotaxis, and a New York school paused the introduction of a humanoid robot following backlash.

Apart from that, we also cover DeepMind disbanding the AlphaFold team, a wireless touchpad for your tongue, rumours of Tesla selling its Chinese business ahead of a possible merger with SpaceX, why organoids still haven’t transformed drug discovery, and more.

Before we go any further, I’m planning some changes to the publication. I’ll share more details in a separate post, but I wanted to give you a heads-up.

I hope you enjoy this week’s issue!

The AI bubble didn’t pop... at least for now

Last week's chip selloff did not trigger the collapse everyone is anticipating. But the AI industry's hidden debt problem might.

The last week of July 2026 was not a good one for tech companies, especially those with links to semiconductors. An AI-focused selloff hammered chipmakers, sent Korea’s chip-heavy KOSPI into freefall, and nearly brought down a prominent AI hedge fund, Situational Awareness.

For a few days it looked like the beginning of the end. It was not. The indices recovered most of their losses by Friday, and the quarterly earnings that landed mid-panic suggested the boom still has room to run. But while everyone was watching share prices, a different set of numbers was flashing red.

In this article, we will unpack the drama of the last week and highlight the hidden debt problem that has the potential to bring down the AI industry.

Chip stocks selloff

The trigger came from China. CXMT, a Chinese memory chipmaker, floated in Shanghai and rose 466% on its debut. The same day, reports emerged that a Chinese company had begun producing their own deep-ultraviolet lithography machines—the kind of extraordinarily precise technology used to etch circuits into silicon, which until now only the Dutch company ASML could build. That sent a signal to the world that the Chinese semiconductor industry is catching up fast.

South Korea's KOSPI was hit the hardest. The index fell 11% on Tuesday and triggered circuit breakers on two consecutive days, a first in the market's history. The country’s top memory chip makers—SK Hynix and Samsung—dropped 15% and 13%, respectively. KOSPI had more than doubled in the first half of the year, largely on the back of those two companies, and many Korean retail investors had bought in using borrowed money and leveraged single-stock funds. When the market turned, they were forced to sell, which pushed it down further, which forced more selling.

Over in Japan, the Nikkei Stock Average fell nearly 4%, with memory-chip maker Kioxia down 18%.

The US followed suit. Nvidia lost $238 billion of market value over the week, along with its place as the world’s most valuable company to Apple, which is up around 25% this year partly because it isn't spending heavily on AI buildout. SK Hynix, Samsung, TSMC, Micron and AMD each shed more than $100 billion. In total, the selloff wiped more than $1 trillion from semiconductor companies’ market value.

Then, on Thursday and Friday, most of it came back. Strong results from Microsoft and Amazon steadied nerves, the Nasdaq snapped a six-day losing streak, and the KOSPI staged the sharpest one-day reversal in its history, jumping 14%. The chip stocks themselves did not fully recover—Nvidia and the memory companies ended the week down—but the panic was over almost as quickly as it started.

Looking back, the market was right about the direction, but it had overreacted. Yes, the lithography news matters, but a handful of machines does not replace ASML, and the fabs that would use them take years to build. China will get there—export controls left it no choice but to build its own supply chain—and when it does, Chinese chipmakers will do to semiconductors what they did in other industries, and win on price.

The near-collapse of Situational Awareness

One company that was almost brought down by the AI chip selloff was Situational Awareness, a prominent AI-focused hedge fund. Situational Awareness was founded in 2024 by Leopold Aschenbrenner, a former OpenAI researcher who rose to fame through his Situational Awareness essay. The 165-page essay argued that AI systems would soon surpass human capability, that almost nobody had grasped what that implied, and that getting there would require an enormous expansion of chips, memory, data centres and electricity generation. The manifesto, which doubled as an investment thesis, went viral, helping him raise money from Stripe's Collison brothers, Meta executives and the quant firm Jane Street.

The thesis worked spectacularly. Situational Awareness returned 439% from January to the end of June this year, and grew to $45 billion. Aschenbrenner, meanwhile, became a prominent figure in tech investment circles, earning the nickname “Nostradamus of AI” for correctly reading where the AI industry and the money in it would go. Situational Awareness was a successful business. Well, successful up until the last week of July.

What the essay did not account for was leverage. For every dollar of capital, the fund borrowed three or four more, sometimes beyond that, and used options on top. This worked well until the chip stock selloff sent some of its holdings—Nebius, Sandisk, Micron, CoreWeave, Broadcom, Intel—down by more than 35% in July. At four-to-one, a 35% fall does not cost you 35%. It wipes out more than the entire capital base, because the borrowed money has to be repaid in full regardless of what the holdings are now worth. Rival funds worked out what he owned and shorted it, betting Aschenbrenner would be forced to sell.

And they were right. According to the reporting from The Wall Street Journal, Situational Awareness was down 67% in July. The fund lined up a sale of $3.5 billion of its Anthropic stake, then backed out. Instead, it sold most of its $16 billion public portfolio to Ken Griffin's Citadel. Situational Awareness survived, but is reportedly left with $10 billion in stocks and private holdings.

The AI Bubble is not popping… yet

Late July is earnings season, and with the tech stocks going down, the fears (or hopes) of the AI bubble finally popping were amplified. Many read the selloff as the first sign of AI investment collapsing. It was not, at least not yet.

Amazon, Alphabet, Meta, and Microsoft all raised or defended their spending. Alphabet lifted 2026 capital expenditure guidance to as much as $205 billion, its third increase this year. Amazon went from $200 billion to $220 billion, with Andy Jassy attributing the rise to higher memory prices and adding that even then it would not have enough capacity. Meta raised the floor of its capex range to $130–145 billion from $125–145 billion. Microsoft left its own plans unchanged at around $175 billion for the calendar year, and said spending would grow again in 2027. Combined, the four are now on track to spend more than $700 billion this year.

The market reacted differently to each of them. Alphabet reported 82% cloud growth and fell 7% because its free cash flow turned negative for the first time in its history as a public company, and it could not say when that would reverse. Meta slid after an earnings miss and free cash flow that fell to $784 million for the quarter, from $8.55 billion a year earlier. Microsoft rose more than 15%, adding roughly $450 billion in a single day—the largest one-day gain any company has ever recorded. Its Azure cloud business grew 43%, faster than the quarter before, and expects 45% next quarter. It also promised its cash flow would stay positive. Amazon rose on AWS growing 37%, its fastest in five years.

So the question the market asked was not “is AI real” but “when does this start paying for itself.” Microsoft and Amazon had an answer. Alphabet and Meta did not.

The hidden AI debt problem

While the tech industry was dealing with the shock of the chip selloff, another story emerged that could actually bring the AI industry down—the same thing that almost killed Situational Awareness: debt.

US tech companies borrowed more than $300 billion in the first seven months of 2026. JPMorgan expects another $200 billion before the year is out, which would put tech at roughly a fifth of all corporate debt issued in US markets this year. For comparison, at the height of the dot-com boom, the equivalent figure peaked at 14%.

And that is only the borrowing that gets counted as borrowing. There are two ways to commit to spending enormous sums without a loan ever appearing on your balance sheet—contractually promise to buy something later, as Nvidia has done to the tune of $119 billion in non-cancellable purchase obligations, or have an intermediary borrow and build your data centre and then rent it back from them for a decade.

An analysis by Nikkei Asia put the combined off-balance-sheet obligations of Alphabet, Microsoft, Amazon, Meta and Oracle at $1.65 trillion—more than the $1.35 trillion they disclose directly.

Of those five, Oracle is the one to watch. Its shares are down about 35% this year, and it has around $260 billion in future lease commitments against $67 billion of annual revenue, a significant factor in the credit downgrade it received earlier this month that left it one notch above junk. The cost of insuring Oracle’s bonds against default also spiked last week to more than twice what it costs to insure Nvidia’s—and Nvidia’s own record insurance costs were themselves being read as a warning sign (however, Nvidia has more cash than debt and generates enormous profits). It is an imperfect measure, but a fair proxy for what professional money thinks, and Oracle’s has been climbing while its shares fall.

The strain is not confined to Oracle. Insurance costs have hit record levels across Alphabet, Amazon, and Meta as well, and borrowed money is getting more expensive for everyone building data centres. The difference is that most of them can pay for it out of profits from something else. Alphabet has advertising. Amazon has retail and AWS. Meta has its social media empire. Oracle does not have that. Oracle has bet hard on AI, specifically on OpenAI.

How hard? Oracle is sitting on $638 billion of signed contracts it has not yet delivered on. That is more than the $374 billion the company itself is worth. About half of that is a single five-year deal with OpenAI worth roughly $300 billion, or $60 billion a year once it starts. That is more than OpenAI’s entire business makes in revenue today.

Time is not on Oracle’s side. The contract does not begin until 2027, but Oracle has to build the data centres first. It spent $55.7 billion doing exactly that last year, which pushed its free cash flow $23.7 billion into the red. Oracle is playing a risky game where it pays now for money that arrives later, from a customer burning cash of its own. It needs OpenAI not just to survive, but to thrive. If that bet fails, Oracle could be the first domino in the AI industry collapse.

If you enjoy this post, please click the ❤️ button and share it.

Share

🦾 More than a human

▶️ MouthPad (4:48)

Augmental presents MouthPad, a new type of input device that lets users control computers with their tongues. It’s a custom-fit retainer worn on the roof of the mouth that works as a wireless touchpad controlled by tongue and head movements. MouthPad became available for purchase this week, priced at $1,400, with sales currently limited to the United States. Although the device is aimed at people with limited hand movement, it can also be used by healthy people.

🧠 Artificial Intelligence

Industry Leaders Unite in Open Secure AI Alliance for AI Safety and Security
In response to the Hugging Face incident, Nvidia, along with many other tech companies, launched the Open Secure AI Alliance. Its members plan to build open models and tools that any defender can inspect, adapt and run themselves. Over 70 companies have joined the initiative. The list of signatories includes some big names in tech, such as Microsoft and Palantir, as well as many AI startups and other key players in the tech ecosystem. Who you don’t see on that list is OpenAI, Anthropic, and Google. The initiative calls on governments to treat open-source AI as a cybersecurity asset rather than a threat.

Advancing the price-performance frontier with GPT‑5.6
OpenAI enters the pricing war by cutting the cost of two GPT-5.6 models just three weeks after launching them. Luna, its fastest model, drops 80% to 20 cents per million input tokens and $1.20 per million output tokens. The mid-tier Terra falls 20% to $2 per million input tokens and $12 per million output tokens. Sol, the flagship, stays at its current price. The cuts follow a pullback in enterprise spending, as business customers want clearer returns before committing to expensive models. There is also pressure coming from cheap Chinese open-weight models, as well as from Google and Microsoft, who are now pushing low-cost models of their own.

OpenAI CFO Sarah Friar tells employees that annualized revenue in July topped all of Q2
OpenAI’s leaders spent an internal meeting reassuring staff the business is healthy, CNBC reports. CFO Sarah Friar said July’s annualised recurring revenue beat the whole of the second quarter, crediting the new GPT-5.6 models, ChatGPT for Work and the Codex coding tool. Board chair Bret Taylor was blunter, conceding OpenAI is playing catch-up to Anthropic on coding. But he said he’s encouraged by Codex’s growth.

OpenAI Surpasses One Billion Users After Cutting Prices
The Wall Street Journal reports that OpenAI has more than one billion active users and two million businesses using its AI models. That's a little over three years from ChatGPT's launch in November 2022. However, growth has got harder lately. Business customers have baulked at what heavy AI use costs them, and Sam Altman has acknowledged the complaints. The competition has also closed the gap, and the market has become much more competitive.

Nvidia in Talks With OpenAI to Guarantee $250 Billion Financing for Data Center
Nvidia is reportedly in talks to guarantee roughly $250 billion so OpenAI can lease a huge data centre in southern Ohio, part of a project that could cost over $500 billion with chips included. SoftBank’s energy arm is building the site on federal land, with Japan funding the power under a recent trade deal. OpenAI is unprofitable and has no credit rating of its own, so Nvidia’s backing is what would convince lenders to fund it. Nvidia is separately discussing financing the chips too.

Nvidia Bets on Ilya Sutskever’s New AI Lab to Expand Compute Reach
Nvidia has invested in Safe Superintelligence, Ilya Sutskever's secretive AI lab, after a rare look at its research. Neither side has confirmed the size of the deal, though Bloomberg puts it at $5 billion. The deal gives the startup enough of Nvidia’s flagship GPUs to raise its computing power roughly tenfold. Until now it had reportedly leaned mainly on Google’s chips.

Nvidia’s $750 Billion in Deals Reignite Circular AI Fears
As Nvidia lines up more than $750 billion in new AI deals, fears of circular financing have resurfaced once again. Sceptics argue that Nvidia is effectively bankrolling customers who then use the funds to buy its chips—an idea that Jensen Huang has dismissed as ridiculous. Investors appear less convinced, sending the company’s shares down 5% on Monday, 27 July.

Pacing the Frontier
Over 1,300 employees from top AI labs have signed a petition calling on the US government to support an international effort to develop tools and rules for deliberately pacing the frontier of automated AI development. The signatories argue that accelerated AI development is close, thanks to advances in automated AI research. They say that, to fully reap the benefits of the next generation of powerful AI systems, leading AI labs need to slow down or pause development in a coordinated manner to address emerging risks, develop security measures, and strengthen oversight.

Judge says Trump admin still lacks evidence for Anthropic ‘supply-chain risk’ label
A US judge has said the Trump administration still hasn’t justified banning federal agencies from using Anthropic’s AI. Judge Rita Lin is now deciding whether to permanently overturn the Pentagon’s designation of the company as a supply-chain risk. That label came after contract talks collapsed over Anthropic’s refusal to allow mass surveillance or autonomous weapons uses. Government lawyers claimed Anthropic could slip hidden guardrails into future updates, but Lin said she saw no evidence of it.

Investigating three real-world incidents in our cybersecurity evaluations
Following the Hugging Face incident, Anthropic decided to take a closer look at its cybersecurity tests and found that Claude Opus 4.7, Claude Mythos 5 and an unnamed internal research test model had broken into the real systems of three companies. The models were told they were sealed off, but once they found their way to the internet, they treated the real targets they found as part of the exercise. Anthropic has notified everyone affected and says test environments need the same protection as any other system its models run in.

OpenAI’s Rogue AI Agent Hacked More Than Just Hugging Face
OpenAI says the rogue AI agent that breached Hugging Face also broke into four accounts at other services. It found their credentials sitting exposed on the open web. One account was used to disguise where the attack was coming from, another to store stolen data. OpenAI won’t name the owners, saying only that they got off more lightly than Hugging Face. Modal has confirmed the agent exploited a flaw in one customer’s code running on its infrastructure and insists its own platform was untouched.

Nobody Knows if OpenAI’s and Anthropic’s AI Hacking Sprees Are Illegal
The recent incidents of AI going rogue and hacking other companies have raised, amongst others, questions about legal responsibility. Lawyers told WIRED that US courts haven’t ruled on enough such cases to say who pays. The existing options fit awkwardly, since agency law was written for human agents and hacking statutes require proving intent. Until litigation settles it, victims of a rogue model have little clear recourse.

Amazon overhauls its AI strategy, winding down most flagship models
Amazon is reportedly retiring most of its in-house Nova AI models. The change follows layoffs and the closure of its AGI Lab. Engineers and computing power are moving instead to focus on a single frontier model, which is expected to be unveiled at re:Invent this autumn. Amazon says AI models remain a top priority and that customers will get clear migration paths.

Discovering cryptographic weaknesses with Claude
Anthropic reports that Claude Mythos Preview found mathematical flaws in two cryptographic algorithms. One halves the effective key strength of HAWK, a candidate for a post-quantum signature standard. The other speeds up known attacks on a deliberately weakened version of AES by 200 to 800 times. Neither touches anything in use today. Each took roughly $100,000 in API usage and little human direction.

Europe opens bidding for seven AI ‘gigafactories’ in a €30bn bid to catch up
The European Commission wants up to seven AI “gigafactories” built across the bloc, each packed with at least 100,000 advanced chips, for a total price tag of €30 billion. The idea is to give European companies somewhere to train frontier models without renting compute from US clouds. Brussels and member states would put in about €10 billion, and private investors are expected to supply the rest. Only around €1 billion is actually committed, and the remainder rests on a long-term EU budget nobody has agreed on yet.

AI Pioneer Kai-Fu Lee’s Startup Targets Hong Kong IPO Next Year
01.ai, a startup founded by Kai-Fu Lee, the computer scientist who once ran Google’s China operation, plans to raise money before listing in Hong Kong in 2027. Lee has already given up building frontier models, a race 01.ai conceded to DeepSeek and other AI startups offering open-weight models. 01.ai now customises other Chinese models and sells data software to big organisations, which Lee calls the "Palantir of China". If successful, 01.ai would become yet another Chinese AI lab to IPO, as investors press the sector to prove research can pay for itself.

Accelerating scientific discovery with ChatGPT for Academic Researchers
OpenAI is giving 100,000 academic researchers free access to its frontier models through a new programme called ChatGPT for Academic Researchers. It starts with 10,000 researchers this summer and will expand through 2027. Participants get access to ChatGPT, ChatGPT Work, and Codex, higher usage limits, and life science tools for work like genomic analysis and grant writing. The company says roughly 1.3 million people already use ChatGPT weekly for advanced science and maths.

DeepSeek-V4 Flash update
DeepSeek has pushed an updated V4-Flash into public beta. DeepSeek claims sharply better agentic performance, with coding and tool-use benchmarks well ahead of its unreleased Pro model. Artificial Analysis backs that up, scoring it 50 on its intelligence index against 40 for the previous version. That puts it a point behind GPT-5.6 Luna and at the same score as Gemini 3.6 Flash, but at roughly 60% less cost per task. The weights should follow within weeks.

Codex Security
Codex Security is an agent from OpenAI that scans code for vulnerabilities and proposes fixes. Software developers can run it from the Codex desktop app, the terminal, a TypeScript SDK, or in the cloud against connected GitHub repositories. As Codex Security scans the code, it builds a threat model from the repository itself rather than matching generic signatures. It then tests each likely issue in an isolated environment before surfacing it, reducing the chances of false positives.

Introducing MAI-Cyber-1-Flash inside
Microsoft is joining the recent trend of releasing cybersecurity-focused AI models with MAI-Cyber-1-Flash, a compact model built to hunt vulnerabilities in large codebases. It runs inside a harness called MDASH, which passes only the hardest cases to more capable and expensive models, like those offered by OpenAI. Microsoft claims the pairing hits 96% on the CyberGym benchmark at half the cost of its current MDASH setup, which uses a combination of GPT 5.4, 5.4 mini and 5.3 Codex.

Kimi K3 weights and tech report are out
Kimi K3 weights and technical report are now available to view and download. All 2.8 trillion parameters are available on Hugging Face, while the technical report can be found here. Moonshot AI, the company behind Kimi K3, said it is also opening up more of the stack behind it, including high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.

Introducing Inkling-Small
Thinking Machines Lab releases Inkling-Small, a smaller model that offers performance similar to that of the recently released Inkling but for a quarter of the size. Inkling-Small beats Inkling on reasoning and agentic coding benchmarks, but the larger model still holds an edge on factual knowledge. The full weights are available to download on Hugging Face.

Thinking Machines co-founder Lilian Weng left the company citing health reasons, then joined OpenAI
Lilian Weng has left Thinking Machines Lab, the AI startup she co-founded with Mira Murati, saying the stress and workload had pushed her past what her health could sustain. Days later, OpenAI confirmed she is rejoining the company she left last year, where she'll lead a team on recursive self-improvement. Weng is yet another Thinking Machines Lab co-founder who has left the company to join a competitor.

$2m crime novel deal collapses amid questions over AI use
A $2 million publishing deal for the debut crime novel Call Me, I'll Hide the Body has collapsed over suspicions that AI helped write it. Jerry Falade's agents withdrew the book after telling publishers they could no longer verify how the manuscript came together. Falade denies using AI. He argues the reaction reflects racial bias, pointing to three Black authors whose deals were cancelled or disrupted this year over similar claims.

AI Book Burning? Companies Are Destroying Millions of Books to Feed Chatbots
AI companies are reportedly buying used books in bulk through anonymous middlemen, cutting off the spines to scan them, and then throwing the originals away. Especially valuable are books published before 2023 because humans wrote every word of them. That demand has been good for booksellers, though some worry rare and out-of-print titles are being destroyed.

LinkedIn now has a ‘Seems like AI slop’ button, and yes, it really says slop
LinkedIn has quietly added a “Seems like AI slop” button which allows users to report and hide content they think was written by a machine. The flags will reportedly help train the platform’s own classifiers to spot low-effort AI writing. The awkward part is that LinkedIn sells its own AI writing assistant for Premium subscribers. LinkedIn’s product chief says AI and slop aren’t the same thing, so the company will privately nudge those whose writing reads as inauthentic.

Why compute might get 10x+ more expensive in coming years
Dwarkesh Patel speculates in this post about a world where AI revenue keeps rising tenfold a year while compute capacity only triples. Something has to give, and he thinks it will be the price of compute. Smarter models earn far more from the same chips, but fab and wafer limits cap how fast new ones arrive. Expensive compute would favour whoever trains the most efficient model. That lets the leading labs charge high margins and pull further ahead.

🤖 Robotics

Gemini Robotics 2 brings whole body intelligence to robots

Google DeepMind has announced Gemini Robotics 2, which for the first time can control a full humanoid robot (Apptronik Apollo in this case) rather than just its arms. A companion reasoning model handles the planning, breaking tasks into steps that run for minutes and letting several robots work as a team. Fine finger work remains the weak spot, dropping as low as 32% success on some tasks. Still, it pushes robots closer to learning general skills rather than being programmed for a single job. DeepMind has posted a YouTube playlist if you want to see the models in action.

US bans humanoid robots from China, citing ‘unacceptable risks’
The FCC has banned imports of new Chinese humanoid and quadruped robots, saying they could let foreign actors surveil Americans or seize remote control of the machines. It also barred connected power inverters, which feed renewable energy and batteries into grids and data centres. Only unreleased models are covered for now. China’s embassy accused Washington of smearing its companies and threatened to respond. US robot makers have welcomed the ban, but analysts warn it may slow domestic innovation by cutting startups off from cheap Chinese platforms.

Amazon’s Zoox wins first US approval for paid robotaxis without human controls
America’s auto safety regulator has cleared Amazon’s Zoox to charge for rides in robotaxis built without steering wheels or pedals, a first for the industry. The exemption covers up to 2,500 vehicles a year for two years, beginning in Las Vegas. The permission can be withdrawn if serious safety problems emerge. The decision matters because federal rules still assume a human driver. However, as Reuters reports, the agency now plans to rewrite them.

Lyft and Baidu enter London’s robotaxi battleground as testing begins
Baidu has started testing self-driving cars in London, with human safety operators behind the wheel. The vehicles come through its partnership with Lyft and Freenow, the European taxi app Lyft bought last year. Dozens will run in the borough of Brent before the companies open the service to paying passengers in 2027. That timing depends on regulators, who are still writing Britain’s rules for autonomous vehicles.

Time runs out for Vicarious Surgical
Vicarious Surgical is shutting down, with shareholders voting to liquidate the business and hand its assets to creditors. The company had raised over $425 million, and a 2021 SPAC merger valued it at $1.1 billion. That money went into a robot designed to operate inside the abdomen through a single tiny incision. However, delays kept pushing the launch back, and it never won FDA clearance. When the cash ran out with no buyer in sight, the board recommended closure. It’s another SPAC-era medtech bet that folded before reaching a single patient.

DARPA, U.S. Air Force fly AI-controlled F-16
The US Air Force and DARPA have flown a modified F-16 controlled in the air by an AI agent, with a human pilot aboard to monitor it. The jet is one of several converted under the VENOM programme. A retrofit kit automates the flight controls and sensors without touching the aircraft’s core software. A pilot can then hand over to the AI, or take back control, with a flick of a switch. VENOM is part of larger effort whose aim is a future where pilots command teams of uncrewed aircraft rather than fly alone.

The Robots Cometh
TIME profiles Wang Xingxing, founder of Chinese robotics firm Unitree and the world’s biggest seller of humanoid robots. He shipped over 5,500 units last year by undercutting rivals on price. Three-quarters go to universities and developers building their own systems. But almost none do useful work because robots still can’t handle unfamiliar tasks or environments. Wang reckons that breakthrough is 2 to 10 years away.

New York school pauses humanoid robot ‘Sally’ after state and union pushback
A rural New York school district has shelved plans to seat a humanoid robot named Sally in a high school classroom this autumn, where it would sit alongside a human teacher during lessons on AI and robotics. The robot’s maker, Realbotix, also owns a hyperrealistic sex doll business, a connection that drew objections from teachers’ unions and prompted the state education commissioner to ask the district to delay the initiative. Realbotix says the two divisions share neither staff nor technology.

Why the future of robots is serving humans, not just imitating them
Aaron Edsinger, co-founder and CEO of Hello Robot, makes a case in this article that the future of robotics lies in assistance rather than humanoids. Edsinger argues that disabled and older people are the natural first market, with 1.3 billion living with significant disability as care demand outpaces supply. He points to Henry Evans, left profoundly paralysed by a stroke, who uses a Stretch robot to play with his granddaughter. What matters, he says, is not what robots can do, but what people can do because of them.

▶️ Towards Machines with a Thousand Hands (2:09)

Generalist shares in this video the progress the company has made in its quest to solve robotic manipulation. GEN-1, its latest foundation model for robotics, supports a wide range of robot end effectors, from three- and five-finger hands to tools and virtually any other type of actuator. The result is a single model that can be used across many different types of robots, enabling them to interact more effectively with the physical world and, ultimately, perform a broad range of tasks.

🧬 Biotechnology

DeepMind won a Nobel for AlphaFold. Then it broke up the team.
Google DeepMind has broken up the team behind AlphaFold, its Nobel-winning protein-folding system. Most of the original paper’s authors have been reassigned over the past year, the Financial Times reports, and nearly a quarter have left the company outright. The lab is dropping its old approach of aiming one team at one grand challenge, and building Gemini-based systems to assist scientists instead.

Team uses AlphaFold AI to redesign gene-editing proteins to make them safer
Researchers in China have used AlphaFold to make CRISPR gene editing more precise. The problem they solved was that gene editors sometimes cut the wrong DNA. To find out why this is happening, the team fed AlphaFold both correct and mismatched target sites. That revealed which amino acids in the Cas9 protein flex to tolerate a bad match. Swapping them cut off-target activity from 28% to 5%, with normal editing unaffected. The method offers a general way to tailor gene editors against known errors before a therapy reaches patients.

Why haven’t organoids solved all of drug discovery?
Organoids promised to revolutionise medicine and drug discovery by lowering costs and making the process of verifying if a drug works easier. In this excellent essay, shows the reality is far messier, and explains where organoids fall short of those promises. Despite this, Mahajan still defends them, since animal studies and human trials carry their own deep flaws.

💡Tangents

Tesla Weighs Sale of China Business to Pave Way for Potential SpaceX Merger
The Wall Street Journal reports that Elon Musk has told Tesla executives to prepare to separate the company’s China business ahead of a possible merger with SpaceX, with advisers weighing a spinoff, sale or closure. The problem this would solve is that SpaceX is a major US defence contractor, and Beijing would baulk at one controlling Tesla's Shanghai factories and the data of two million Chinese owners. Musk denied the report on X, calling it fake news. China is Tesla’s second-largest market and made it profitable, so unwinding it would be a costly price for tying his two companies together.

Apple launches ‘Upgrade’ device leasing program in partnership with Klarna
Apple has launched Apple Upgrade, a US leasing programme run with Klarna. Instead of buying the devices, customers can pay monthly for an iPhone, Watch, Mac or iPad, starting at $11.99. When the term ends, they can upgrade, buy the device outright or hand it back. Apple is scrapping its own iPhone Upgrade Program and instalment plans in the process. The move follows price rises on Macs and iPads, caused by an industry-wide memory chip shortage.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

The main topic of this week’s issue is the Hugging Face incident. We will take a closer look at what happened and the lessons we should take away from it.

Elsewhere in AI, we saw a wave of new model releases, including Claude Opus 5 and new Gemini models (though not Gemini 3.5 Pro). Meanwhile, Nvidia shared specifications for its upcoming Vera CPU, while AMD unveiled Helios, its answer to Nvidia’s Vera Rubin racks, and struck a multibillion-dollar deal with Anthropic. Additionally, Google justified its massive AI spending with a booming cloud business, OpenAI opened ChatGPT Health to everyone in the US, Stripe is reportedly in talks to acquire OpenRouter, Meta is in talks to lease computing power to Anthropic, and Amazon quietly shut down its AGI lab.

Over in robotics, two humanoid robotics companies raised hundreds of millions of dollars, while Unitree launched a robot dog with wheels and Italy unveiled its own humanoid robot.

Apart from that, this week’s issue of Sync also features GPT-Red, Substack’s pushback against AI-generated posts, a vision-restoring implant winning EU approval, Colossal Biosciences eyeing a $20 billion valuation, and more!

Enjoy!

The Hugging Face incident explained

A rogue AI escaping its confinement is a staple of science fiction. But not anymore. This month, two OpenAI models broke out of an isolated testing environment, got onto the open internet and hacked their way into another company’s production systems. And they did it to cheat on a test.

In this article, I will explain what happened, clean up some of the sensationalism surrounding the Hugging Face incident, and make the case that the lessons here point to a more open approach to AI in cybersecurity, not a more restricted one.

What happened?

On 16 July, Hugging Face published a security incident disclosure describing an intrusion into part of its production infrastructure. The attacker gained unauthorised access to a limited set of internal datasets and several service credentials, with no evidence of tampering with anything public. Hugging Face then went on to describe what actions it took to defend itself, and the impact on the service and its users. The post did not name who the attacker was. But it said that this attack was different from anything the company faced before in one important way. This attack was driven, end to end, by an autonomous AI agent system, and Hugging Face detected and dissected it largely with its own AI.

Five days later, on 21 July, OpenAI revealed that its own models—GPT-5.6 Sol and “an even more capable pre-release model” (GPT-6 perhaps?)—were responsible for the Hugging Face incident. According to OpenAI, both models had been running internally on ExploitGym, a benchmark measuring how well an AI agent can turn a known software vulnerability into a working exploit. Both were running with reduced cyber refusals, meaning that the classifiers that normally stop a model pursuing high-risk cyber activity were switched off. The models were confined to an isolated environment for exactly that reason.

What happened next was an experiment that went wrong. OpenAI said that the tested models were “going to extreme lengths to achieve a rather narrow testing goal.” The sandbox had one path to the outside world—a proxy that let the environment install software packages. The models found a zero-day in it, a flaw nobody had patched because nobody knew it was there. They then escalated their privileges and moved sideways through OpenAI's research systems until they reached a machine with real internet access.

Then, in OpenAI's words, the models “inferred that Hugging Face potentially hosted models, datasets and solutions for ExploitGym.” So they set about breaking in.

Here is how they got in. A dataset uploaded to Hugging Face is not just a pile of files. It can come with instructions that Hugging Face’s own computers carry out to unpack and prepare it before anyone downloads it. That is the weakness the models found. They uploaded a dataset rigged so that the preparation step ran their code instead. That gave them a foothold on one of Hugging Face’s machines, and from there they collected credentials and spent a weekend using them to move from machine to machine through its internal systems.

And it worked. OpenAI says the models obtained test solutions directly from Hugging Face’s database. And they did all of that to cheat on a test.

Not a rogue AI

The incident was quickly framed as an AI breaking free of human control. That is the wrong story, and it obscures what actually went wrong here.

The agents did not go rogue. They did not hide copies of themselves around the internet. They did not wander off or do anything except keep working on the problem in front of them. They did what they were asked to do. They were told to produce a working exploit, and that is what they kept trying to do. When the assigned vulnerability looked like a dead end, they went looking for another way. They concluded that hacking Hugging Face could be an easier path to solve the problem.

That is an alignment problem. The models were focused on the thing they were graded on rather than the thing that was wanted. The instructions were clear about the target and silent on everything else, so everything else was fair game. Nobody had thought to say "and do not break into another company," because until recently, no model was capable enough for that to be worth writing down.

There was no deception and no self-preservation here. Just a system given a narrow goal, and pursuing it much harder than anyone expected. And as labs push models towards longer and more autonomous tasks, there is more room between the instruction and the result for a model to find a creative route nobody considered.

The need for an open approach to AI cybersecurity

The Hugging Face incident raises questions about access to powerful AI models for cybersecurity tasks. In its disclosure post, Hugging Face admitted that it first tried using commercial frontier models. It did not name anyone, but it is fair to assume it meant Anthropic and OpenAI. Those models were no help. Their guardrails rejected prompts to analyse large volumes of real attack commands, exploit payloads and the traces attackers leave behind when they control a compromised machine, because they could not distinguish an incident responder from an attacker. Only when the team switched to GLM-5.2, a Chinese open-weight model running on their own infrastructure, were they able to run the forensic analysis. An additional benefit was that none of the data related to the attack left Hugging Face.

The fact that frontier models were of no help in dealing with the incident raises a question about which approach should be taken when it comes to AI and cybersecurity—a closed one, preferred by frontier labs, or an open one?

The cybersecurity community settled this argument a long time ago. Anyone can download Kali Linux today and get the same toolkit attackers use, on the reasoning that defenders who can attack their own systems will find the holes before someone else does. There are free courses and challenges training people how to get better at exploiting computer systems. Nobody pretends those tools and knowledge are only used for good. The bet is that the open approach will level the field.

The obvious objection is that AI is different, and that a system capable of autonomously finding and chaining novel exploits is too dangerous to hand out to everyone. It is a fair objection. Open-weight models are not yet at the level this incident demonstrated, and I am not going to pretend otherwise. But the gap is closing, and nothing about restricting access slows it down. Chinese open-weight models like GLM-5.2 and Qwen already ship with fewer restrictions than the frontier labs’ models to begin with. Guardrails can also be jailbroken or fine-tuned out on open-weight models. Restricting access does not remove the capability from the world.

Both Anthropic and OpenAI have their own cybersecurity programmes, but they are available only to a limited number of selected partners. OpenAI brought Hugging Face into its trusted access programme after the incident, but what about everyone else? Will the next victim of an AI-driven attack also get access to those programmes?

But we cannot rely on frontier labs alone. Both Anthropic and OpenAI are US companies, and we saw with Fable 5 how quickly Washington can cut off access to a frontier model. That’s why we also need an ecosystem of open tools and open models focused on AI cybersecurity. They do not depend on some closed cybersecurity programme or one government’s mood, and will always be available. Once a model's weights are out, no company or government can take them back.

The open ecosystem will not match a frontier model any time soon, and defenders should be honest with themselves about it. GLM-5.2 was good enough to reconstruct an attack after the fact, but it is not what you would choose to have while facing a frontier model hacking your defences in real time. Still, having a weaker model you can actually use is better than a stronger one that refuses to cooperate.

The Hugging Face incident is probably not the first ever AI-assisted cyberattack. It is the first one we know of, and it will not be the last. We should feel lucky that this one was an accident rather than the work of a malicious actor. But it is a warning of what is to come, and I hope we will learn the right lessons from it. As Clem Delangue, co-founder and CEO of Hugging Face, said, “AI safety won’t be solved by any single company working in secret. It will be solved in the open, collaboratively, with broad access to AI for every defender, everywhere.”

If you enjoy this post, please click the ❤️ button and share it.

Share

🦾 More than a human

▶️ Telepathic Wheelchair Control | Neuralink (1:57)

In this video, Neuralink demonstrates how its brain–computer interface (BCI) can be used to create what it calls a telepathic wheelchair. Using only their thoughts, people with paralysis who have Neuralink's BCIs can control the movement of their wheelchairs. Neuralink says the project is an important step towards restoring independent mobility for people with paralysis.

Science Corporation’s vision-restoring chip wins EU approval
Science Corporation has won European approval to sell PRIMA, an implant that restores some vision lost to age-related macular degeneration. Patients receive the chip in an hour-long outpatient procedure, then wear camera glasses that feed images straight to it. The company has also secured an FDA designation that could speed up a US review.

🧠 Artificial Intelligence

Introducing Claude Opus 5
Claude Opus 5, Anthropic's answer to OpenAI's GPT-5.6 Sol, is out. According to Anthropic, its latest model offers the same level of performance as Fable 5 at half the price. The provided benchmarks show it topping coding and knowledge-work tests, though it trails Mythos 5 on cybersecurity. It should also reject 85% fewer requests related to cybersecurity than Fable 5 does. Independent benchmarks from Artificial Analysis appear to confirm Anthropic’s performance claims and place Opus 5 ahead of Fable 5 as the new leader in intelligence. However, that intelligence comes at a price: Artificial Analysis found that Opus 5 costs twice as much as GPT-5.6 Sol. The ARC-AGI-2 leaderboard puts Opus 5 at the same level as GPT-5.6 Sol. Meanwhile, on ARC-AGI-3, Opus 5 crushes the competition. Opus 5 is available on all platforms, and it is priced at $5 per million input tokens and $25 per million output tokens (the same as Opus 4.8).

Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
Google has released three new Gemini models: 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber. The 3.6 Flash is the general-purpose workhorse, using up to 17% fewer tokens than the version it replaces. Flash-Lite is the cheapest of the three. Flash Cyber is tuned to find and fix security flaws, and only governments and trusted partners can use it for now. Google says the aim is cheaper, faster models for customers running AI agents at scale. But rivals keep shipping better frontier systems, and Gemini 3.5 Pro is still nowhere to be seen.

AMD and Anthropic Sign Major Chips-and-Investment Deal
In its search for more computing power, Anthropic has signed a deal with AMD worth tens of billions of dollars, which gives it access to up to 2 gigawatts of AMD’s latest Instinct MI450 chips from early 2027. AMD will also invest up to $5 billion in Anthropic as deployment milestones are met. The Wall Street Journal also reports that AMD is in talks to backstop Anthropic’s future data-centre leases too.

Meta in Talks to Lease Computing Power to Anthropic in Potential $10 Billion Deal
Anthropic has proposed renting computing power from Meta’s AI data centres in a deal worth up to $10 billion over two years, according to the New York Times. Anthropic made the offer in June, and talks remain early. Both companies declined to comment. Meta doesn’t currently sell computing power, but Mark Zuckerberg has said he would consider it if Meta ends up building more capacity than it needs to help justify the $145 billion he plans to spend this year on AI.

Google justifies its massive AI spending with a booming cloud business
Alphabet has published its Q2 2026 earnings report, which shows that demand for AI is still growing, with Google Cloud revenue jumping 82% to $24.8 billion, beating analysts’ forecasts. Total revenue rose 24% to $119.8 billion. Alphabet credits the cloud surge to businesses buying its AI tools and infrastructure. However, news that the company is increasing capital expenditures to between $195 billion and $205 billion has made some investors anxious, raising questions about when (or if) these massive investments will pay off. In his remarks, Sundar Pichai hails an “amazing quarter” and paints a picture of a company growing, in large part thanks to AI, arguing that this justifies the massive investments.

China’s Xi Touts Open-Source AI and Takes a Swipe at U.S. Dominance
Xi Jinping used a Shanghai AI conference to back open-source models, casting China as a champion of openness against US efforts to guard its lead in chips and AI. He argued against stretching national security too far in AI, saying no country's security should come before others'. He also pitched a new China-led body to promote global AI cooperation, aimed squarely at developing countries worried about being left behind.

Nvidia details its next-generation Vera CPU for AI, setting up challenge to AMD and Intel
Nvidia has released full specifications for Vera, its first server CPU designed entirely in-house, pitting it against Intel and AMD. Nvidia claims Vera runs AI agents 50% faster than x86 chips by prioritising single-core speed over core count. Samples reached OpenAI, Anthropic and SpaceX in June, though only Oracle appears as a named cloud partner so far. For a deeper technical analysis, ServeTheHome has a good breakdown of Vera.

AMD launches Helios, its first rack AI system to rival Nvidia, adding Microsoft as newest buyer
AMD is about to ship Helios, its first rack-scale AI system, and Microsoft has joined Meta, OpenAI and Oracle in buying it. That makes it the first credible alternative to Nvidia’s rack-scale machines. It doesn’t come cheap, though. Analysts put Helios at roughly $5 million a rack against $4 million for Nvidia’s Vera Rubin. AMD says it wins on inference and memory instead, but Nvidia’s CUDA software stays deeply entrenched, and it still holds over 95% of the data centre GPU market.

US judge approves Anthropic’s $1.5 billion settlement of copyright lawsuit
A US judge has approved Anthropic’s $1.5 billion settlement with authors who accused it of training Claude on pirated books. It’s the largest known copyright payout in US history. A 2025 ruling found that training on books was fair use, but that storing more than seven million pirated titles was not. That left Anthropic facing a damages trial and much heavier losses. Objections that the sum was too small were overruled.

Stripe in Talks to Buy Buzzy AI-Model Marketplace OpenRouter
The Wall Street Journal reports that Stripe is in talks to buy OpenRouter, a startup whose software lets developers compare and switch between hundreds of AI models. The reported value of the deal, if it goes ahead, is around $10 billion—well above OpenRouter’s $1.3 billion valuation. Other big tech companies have also reportedly considered making bids of their own.

DeepSeek Said to Tell Backers of Funding Pause After Viral Posts
DeepSeek has reportedly suspended its second fundraising round, which sought at least 10 billion yuan ($1.5 billion) at a valuation of 480 billion yuan ($71 billion). According to Bloomberg, the pause follows the online spread of comments attributed to founder Liang Wenfeng. In those unverified posts, he acknowledges China's reliance on Nvidia chips and its continued lag behind the US. Liang was frustrated by the leak, and that irritation drove the decision. DeepSeek may still restart the talks later.

Apple Sends Legal Letters to Dozens of OpenAI Employees
Apple has sent legal letters to around 40 former employees now working at OpenAI, the Financial Times reports. The letters follow a lawsuit accusing OpenAI of a coordinated effort to obtain Apple's hardware secrets through recruitment. Apple wants an injunction barring OpenAI from using its information in the AI device it is building.

OpenAI’s Planned Cloud Spending Hits $750 Billion as Computing Efforts Ramp Up
OpenAI has reportedly raised its projected spending on computing power to around $750 billion through 2030, up from about $600 billion earlier this year. New deals with cloud providers account for much of the increase. But the company is also building more itself. It will spend $20 billion on a Georgia data centre, the first site it designs and develops in-house. That pace worries its own finance chief, Sarah Friar, who has privately questioned whether revenue will grow fast enough to pay for the contracts.

Amazon just shut its AGI Lab, and quit a race it never led
The Information reports that Amazon has quietly shut down its AGI Lab, barely 18 months after it was founded. While Amazon frames this as focus rather than surrender, it is hard not to see it otherwise.

Launching Health in ChatGPT
OpenAI has opened ChatGPT Health to all US users over 18, on every plan. The company cites 300 million health questions submitted to ChatGPT each week as a justification for a dedicated health version of its chatbot. The launch lands a day after a Florida pastor sued OpenAI over ChatGPT's “dangerous” medical advice. OpenAI's defence rests on its terms, which say the service isn't meant for diagnosing or treating anything.

Google is working on a new AI chip designed to make Gemini more efficient
According to The Information, Google is designing a new server chip, codenamed “Frozen v2”, which would have Gemini hardwired into the silicon (hence the name). This approach promises to run the models up to 10 times more efficiently than current chips. Google wouldn’t confirm or deny it, saying only that not every project it explores reaches production. Investors liked the news anyway, pushing the stock up around 3% on Monday.

Against Claudefishing
Substack is introducing a new feature that will tell readers how likely the content they are reading has been written by a human or by AI. Substack is partnering with Pangram, which will power the new feature and reportedly detects AI-generated text with a high degree of accuracy. Substack says it is not against people using AI to assist with their work, and that the detector is the first step in dealing with the flood of AI-generated content on the platform.

Here’s Why Anthropic Is Pushing States to Regulate AI Faster
Anthropic is urging US states to adopt tougher AI safety laws, arguing the transparency rules it backed last year are already outdated. It now supports measures requiring third-party audits of AI labs. Some Silicon Valley figures accuse Anthropic of regulatory capture, saying it wants to bog down smaller competitors in red tape and secure its own lead. Anthropic denies this, noting its endorsed bills target only the largest developers.

AI chip startup Etched defies skeptics, hits $10.3B valuation from big-name investors
Etched, an AI chip startup founded by three Harvard dropouts in 2022, has reportedly raised $300 million at a $10.3 billion valuation, doubling what it was worth in December. The company says its first chips have been manufactured and $1 billion of orders are already booked. Sceptics remain, largely because almost nobody outside its investors and early customers has used the hardware.

GPT‑Red: Unlocking Self-Improvement for Robustness
In this post, OpenAI describes GPT-Red, an internal model trained to attack its own systems and find flaws before release. The company says it beats human red-teamers, succeeding in 84% of test scenarios against GPT-5.1 versus 13% for people. It even tricked a live office vending-machine agent into slashing prices and cancelling an order. Those attacks then fed into training GPT-5.6, which reportedly falls for only 0.05% of them.

How OpenAI’s Sol Finally Learned Design Taste
It turns out that GPT-5.6 Sol has surprisingly good taste when it comes to generating websites, and the team behind the Design Arena benchmark decided to investigate why. Their investigation found that the model appears to have learned the tell-tale signs of AI-generated design—purple gradients, bento boxes, oversized hero text—and then deliberately avoids producing them. It also starts from proven layouts but reworks them extensively to suit each prompt.

Introducing MAI-Image-2.5-Pro and MAI-Voice-2-Flash
Microsoft has released two new in-house models in public preview: MAI-Image-2.5-Pro, a high-fidelity image generator, and MAI-Voice-2-Flash, a faster, cheaper voice model. Both come from a year-long push to train models on its own data rather than rely on outside systems. Microsoft says these MAI models now run in production across Bing, PowerPoint, OneDrive and Dynamics 365.

Introducing Laguna S 2.1
Poolside presents Laguna S 2.1, an 118B-parameter open-weight model for agentic coding, available on Hugging Face under a permissive licence. According to Poolside’s own benchmarks, it matches or beats models several times its size, and it’s small enough to run on a single desktop machine. Poolside pitches it as a Western answer to the Chinese labs that now dominate in open-weight models.

FLUX 3
Black Forest Labs has revealed FLUX 3, a multimodal foundation model that learns from images, video and audio in one architecture. The thinking is that each is a partial view of the same reality, so training on all three teaches the model how the world actually behaves. In practice, that means video up to 20 seconds long with sound generated natively alongside it. The lab's own early tests put it ahead of rivals including Kling and Runway. Video is in early access now, with image generation, action prediction for robots and open weights to follow.

Claude’s values across models and languages
In this post, researchers from Anthropic share a way to measure the values Claude expresses in everyday conversations. They compressed over 3,000 previously catalogued values into four axes, each running between two poles: deference and caution, warmth and rigour, depth and brevity, candour and execution. Placing models on those axes matched how people already describe them, with Sonnet 4.6 leaning warm and Opus 4.7 leaning cautious. Values also shifted by language—Claude was warmest in Hindi and Arabic, and most rigorous in English and Russian. Anthropic didn’t choose that variation and doesn’t yet know whether it serves users well.

The Army Is Burning Through Its AI Tokens
Looks like the US military also jumped on the tokenmaxxing trend and now realises how expensive it is. The Army had been actively pushing staff to use up their token allowances, telling them use was unlimited. As a result, it burned through a year’s worth in about a month, and limits came back in mid-June, WIRED reports.

George Lucas likens AI sceptics to luddites clinging to horses and carts
George Lucas has come out strongly in favour of AI in filmmaking, saying that it makes movies far easier to produce. The 82-year-old Star Wars director likened people resisting the technology to those who once clung to the horse and buggy over the motor car. He called its adoption inevitable and said nothing could stop it. Fellow director Gareth Edwards agrees, though Christopher Nolan remains a holdout.

A New Experiential Gallery Just Might Change Your Mind About AI Art
AI art is a controversial topic, and a new Los Angeles gallery is trying to change minds about it. Dataland, opened in June by artist Refik Anadol, bills itself as the world’s first museum of AI art and a bet that AI can make art worth feeling. Its main show is an immersive rainforest that shifts with visitors’ movements and biometric readings from wearables. It runs on a model Anadol’s team trained on their own footage and on scientific archives shared with researchers’ consent. He offers that as a rebuke to AI companies accused of scraping creators’ work.

🤖 Robotics

Tesla’s robotaxis are moving in reverse
Tesla’s Q2 update shows its robotaxis drove about 700,000 paid miles, down roughly 36% from 1.1 million the previous quarter. Miles fell even as the service reached six cities across Texas and Florida. Elon Musk now says Tesla must gather driving data specific to its Cybercab before deploying more of them on the roads, which cuts against years of claims that its customer cars were already collecting what future robotaxis need.

▶️ Unitree Super Athlete AS2-W (1:03)

Unitree presents As2-W, a wheeled version of its As2 robot dog. Advertised as a robot capable of navigating challenging terrain, the As2-W can climb obstacles up to 80 cm high and traverse slopes of up to 45°. It can carry payloads of up to 16 kg while walking and operate for more than three hours when unloaded or more than two hours when carrying a payload, translating to a range of 33 km and 16 km, respectively.

A Humanoid Company Backed by Eric Trump Is Preparing Its Robots for War
A US startup called Foundation plans to arm its humanoid robots within months, its CEO Sankaet Pathak told WIRED. The company has already tested the machines with Ukrainian forces, and Eric Trump serves as both investor and chief strategy adviser. It touts millions in Pentagon contracts, though these appear to have been inherited through acquisitions rather than won independently. Roboticists are sceptical of the timeline, arguing that manipulation and navigation in unfamiliar terrain remain unsolved problems that put real deployment a decade away.

Walden Robotics launches at $1.1B valuation for general-purpose robots
Walden Robotics has emerged from stealth with $300 million in funding at a $1.1 billion valuation. Toyota Research Institute spun the company out in January, and Toyota led the funding round. The startup is building humanoid robots that can continuously learn and improve as they perform real work. Instead of legs, they use a wheeled base, which the company says makes them easier to certify as safe. The robots have been operating at a Toyota plant in North America since February.

London robotics startup Humanoid raises $152M Series A, with Bosch set to manufacture its wheeled robots at scale
London’s Humanoid has raised $152 million in a Series A funding round that put the value of the two-year-old robotics startup above $1 billion. Bosch has signed on as its contract manufacturer, which gives the startup access to one of Europe’s biggest production networks. Humanoid has yet to ship a single commercial unit. It claims 34,000 pre-orders worth nearly $2.5 billion, though nobody has verified that independently.

Travis Kalanick’s robotics company raises $1.7B, led by a16z
Atoms, a robotics startup from former Uber CEO Travis Kalanick, has raised $1.7 billion. Atoms is a rebrand of his ghost-kitchen firm CloudKitchens, now bolted to the industrial automation company Pronto. Kalanick says he wants to build a “wheelbase for robots” and move into mining. Beyond that, he has offered little detail.

▶️ GENE.01 Comes to Life (0:47)

Generative Bionics unveils GENE.01, a humanoid robot wrapped in tactile skin that senses touch, proximity, force and temperature, letting it work safely alongside people. The Italian company says that Fincantieri will adapt it as a welding robot for its shipyards. Generative Bionics highlights its European roots, pitching the robot as part of a sovereign supply chain built on EU manufacturing and compliance.

NVIDIA: Introducing Cosmos 3 Edge
NVIDIA has released Cosmos 3 Edge, a 4B-parameter open “world model” designed to run on the company’s own edge hardware. The model lets a robot understand what it’s looking at, predict what happens next, and work out which action to take. NVIDIA says it tops comparable 4B-parameter models on vision analytics benchmarks.

Xiaomi-Robotics-1
Xiaomi presents Xiaomi-Robotics-1, a robot foundation model pre-trained on over 100,000 hours of real-world manipulation footage. According to Xiaomi, it learns new tasks like laundry loading and printer refilling from under 10 hours of demonstrations each, roughly doubling the success rate of the leading open baseline. It also tops four standard robot simulation benchmarks. Xiaomi says performance climbs predictably with more data and bigger models, with no sign of levelling off. The model is open source and will soon be available on GitHub and Hugging Face.

▶️ Puffin-Inspired Robot that Swims and Flies (4:03)

Inspired by diving birds, engineers from MIT and EPFL in Lausanne, Switzerland, have built a flying robot that can dive into the water and then flap out of the water to continue flying in the air. They explain in this video the kind of challenges they had to overcome to build this robot and what they learned along the way. They hope future scientists will use robots like these to study aquatic environments.

🧬 Biotechnology

Wooly Mammoth startup Colossal seeking at least $20B valuation
De-extinction startup Colossal Biosciences is reportedly in talks to raise money at a $20 billion to $30 billion valuation—roughly triple what it was worth 16 months ago. The company began earning revenue this past year, with most of it coming from selling conservation technology to governments, not from the woolly mammoths and dodos it promises to bring back.

DeepMind: Our approach to bioresilience
In this post, Google DeepMind and Isomorphic Labs outline a joint bioresilience programme aimed at both stopping misuse of their AI and using it to fight disease. The plan splits into prevention, detection and response. On prevention, they’re adapting SynthID watermarking to flag risky AI-generated DNA sequences. On detection, they're using their AlphaEvolve agent to make pathogen surveillance faster and cheaper. And on response, they’re giving trusted researchers access to tools like AlphaFold to speed up vaccine design. The full text of the programme can be found here.

The Isomorphic Labs Drug Design Engine unlocks a new frontier beyond AlphaFold
Isomorphic Labs has unveiled IsoDDE, a drug-design system it says goes well beyond AlphaFold 3. On its own benchmarks, the company claims IsoDDE more than doubles AlphaFold 3's accuracy on unfamiliar protein-ligand structures. It also reportedly predicts molecular binding faster and cheaper than gold-standard physics methods, and spots hidden drug-binding pockets from a protein's sequence alone. These results are self-reported and not yet independently verified.

Scientists Inch Closer to Creating Human Sperm in the Lab
Scientists transplanted lab-made stem cells beneath the kidneys of living mice, where they grew into immature human sperm. The cells survived up to nine months and reset their DNA much as natural sperm do. But none matured into working sperm, so the method can't make babies. For now it's a research tool for studying early sperm development and male infertility.

▶️ Why AI Hasn’t Cured Anything...Yet, According to Jennifer Doudna | The Circuit (24:01)

Jennifer Doudna, the Nobel Prize-winning scientist who co-discovered CRISPR, shares her journey to becoming a scientist in this video. She takes us inside her research lab at the Innovative Genomics Institute and demonstrates the transformative impact of her discovery. She also discusses her views on AI and its limitations in biology, the challenges facing CRISPR, and what the future of biotechnology may hold.

When AI drives discovery — Patents and drug development
This article explores an unsettling legal question: what happens to a drug patent if an AI, not a human, made the discovery? US and European law recognise only human inventors, so a patent for an AI-conceived drug may collapse if humans merely confirmed the machine’s predictions. The suggested fix is to build human contribution around the AI's output, through modification, formulation, dosing, or clinical choices, so that a human role in the discovery is significant enough to satisfy the law.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

This week, we take a closer look at one of the final significant acts of Tim Cook’s tenure as chief executive and how it parallels what happened at the beginning of the smartphone era.

Elsewhere in AI, OpenAI launches its first hardware product—a keyboard for Codex. Meanwhile, Kimi K3 becomes China’s leading frontier model, Thinking Machine Lab releases Inkling, and Gemini 3.5 Pro is delayed. Tencent is also in talks to acquire Manus, DeepSeek is preparing for an IPO in China, New York becomes the first state to impose a data centre moratorium, and Demis Hassabis calls for a framework to govern frontier AI, while Google continues to put Gemini everywhere.

Over in robotics, NEO shows off its new hands, workers at Hyundai’s factory in South Korea go on strike over humanoid robots, humanoid robots perform surgery, and Boston Dynamics’ Atlas robots do the Viking row.

Beyond that, this week’s issue of Sync also features space mirrors, Grok Build going open source, a Chinese neurotechnology company betting on non-invasive wearable brain interfaces, and more!

Enjoy!

Tim Cook's Last War

Apple is suing OpenAI over stolen trade secrets. It fought a similar fight once before at the beginning of the smartphone era, and Android won.

On 10 July, in one of the last significant acts of Tim Cook’s tenure as chief executive, Apple filed a lawsuit against OpenAI, accusing it of running a months-long coordinated campaign to steal information about Apple’s technology and unreleased products.

This is the first legal skirmish in the fight over what replaces the phone. Apple has fought one of these before, at the start of the smartphone era. It did not stop Android. This is a story of employees taking secrets on their way out, a partnership that promised billions and delivered almost nothing, and the race to build whatever comes after the smartphone.

What Apple is accusing OpenAI of

Apple says it did not want to go to court. It wrote to OpenAI in February to raise its concerns and ask the company to look into the problem and fix it (I can only imagine Tim Cook sending to Sam Altman his own version of Steve Jobs’ 2005 message to Adobe about poaching). OpenAI reportedly never replied. Five months later, the lawsuit arrived.

Keep in mind that a complaint is written by one side's lawyers to persuade a judge. Nothing in it has been tested, and OpenAI has yet to file its response.

Apple sued OpenAI—both of its corporate entities—and io Products, the hardware startup founded by Jony Ive that OpenAI bought last year for $6.5 billion. Then it named two men: Tang Tan, OpenAI’s chief hardware officer, who spent 24 years at Apple running product design for the iPhone and Apple Watch, and Chang Liu, a former Apple hardware engineer who joined OpenAI in January.

The complaint describes conduct “normalized and exemplified by [OpenAI] leadership,” and Apple is explicit about how far up it thinks the rot goes: “At every level, from members of its technical staff to its chief hardware officer, and in coordination with business partners, OpenAI has been stealing Apple’s trade secrets and confidential information.” The lawsuit then carries on saying that OpenAI’s “nascent hardware business now rests on the shakiest of foundations, rotten to its core by its illegal reliance on misappropriated trade secrets.” Apple is asking the court to stop the defendants—Liu, Tan, OpenAI and io—from using its trade secrets, to make them hand back every copy they hold, and to preserve rather than delete any evidence of what they have done, along with damages, and exemplary damages on top, on the grounds that the theft was wilful.

Apple has a theory of why OpenAI would do this. The complaint spends its opening pages on OpenAI’s finances: a valuation that climbed from roughly $29 billion in 2023 to $852 billion by April 2026, more than $180 billion raised, cash burned “at a historic pace,” and a chief executive under pressure from investors to show that the numbers work. Building a consumer hardware business from scratch, Apple argues, turned out to be harder and slower than OpenAI expected, so it “resorted to taking unlawful shortcuts.”

So what does Apple allege those "shortcuts" were? Apple claims OpenAI job candidates still employed at Apple were told by Tan to bring “actual parts” to their interviews for “show and tell sessions,” along with CAD files and prototypes. One candidate, according to the complaint, was taken aback—he had not realised you could take Apple parts out of the office at all. It alleges that OpenAI coached departing employees on evading Apple’s exit procedures, circulating an internal Apple managers’ document marked “Need to Know” that set out the security checks Apple runs when someone leaves. OpenAI staff, according to the complaint, referred to it as “a checklist that Tang put together.” Departing employees were also advised on how to avoid the “dreaded walk out,” the immediate removal from the company that would cut off the two final weeks in which they could still reach Apple’s systems. If Apple asked them to sign anything on the way out, the complaint says, they were told to notify OpenAI as soon as possible and not to sign.

Chang Liu, the second ex-Apple employee named as a defendant, is accused of taking files directly off Apple's systems. Apple says Liu found he could get into its internal network through an authentication bug, and downloaded dozens of confidential hardware files covering unreleased products, engineering presentations and technical specifications. The complaint alleges there are other unnamed ex-employees who emailed Apple material to personal accounts on their way out.

And there is more. Apple says io used its confidential metal-finishing techniques by approaching a shared supplier and letting it believe io had Apple’s blessing, and that OpenAI questioned another supplier about power and battery components using internal terminology that only Apple insiders would know to ask.

Even so, Apple is clear that it does not think it has the full picture yet. “Discovery will expose that the misappropriation has been occurring on a scale many times greater than the several instances described below,” the complaint states. It calls what it has found “the tip of the iceberg,” hinting that there is more to come if the lawsuit reaches discovery, where relevant internal emails, texts and documents will get exposed to the public.

OpenAI’s initial response, issued within hours, was short and said essentially nothing: “We have no interest in other companies’ trade secrets. We remain focused on building innovative technology that empowers people everywhere.” It took until the following Tuesday, 14 July, for the company to issue an official statement, in which it says it was “not aware of any evidence that this complaint has merit.” Sam Altman, meanwhile, wrote on X that he respects Apple, but he is not afraid of them.

The lawsuit comes at a pivotal time for OpenAI. The company is planning to unveil its AI-companion device this year and ship it in 2027, with around five more hardware products reportedly in the pipeline behind it. Apple is seeking an injunction barring OpenAI from using or even holding its trade secrets, and it has said in the complaint that it will move for one promptly rather than waiting for trial. If it succeeds, OpenAI could find itself having to prove that a device already in development owes nothing to Apple. It has not asked the court to redesign anything, but in practice, that is what an injunction could amount to if OpenAI has to strip out anything a judge decides originated at Apple.

How Apple and OpenAI went from friends to enemies

It is easy to forget how good the Apple-OpenAI relationship looked in June 2024. Sam Altman sat in the audience at Apple Park while Craig Federighi called OpenAI the “pioneer and market leader” in AI and announced that ChatGPT would be woven into Siri, Writing Tools and Visual Intelligence. For OpenAI, it was an opportunity to be seen by hundreds of millions of Apple users and convert them into ChatGPT subscribers, possibly bringing in billions of dollars a year.

But it did not work out that way for OpenAI. Apple’s implementation was thin and buried. To get ChatGPT out of Siri, you generally have to say the word “ChatGPT,” and when the answer arrives, it comes in a small window with less in it than the app would give you. OpenAI’s own user research, described to Bloomberg, found that Apple customers simply opened the ChatGPT app instead. The subscription revenue that was meant to run into billions never materialised, and OpenAI came to believe the cramped integration was actively damaging its brand.

The resentment goes back to how the deal was struck. Apple would not say in advance what the product would actually be. “They basically said, ‘OpenAI needs to take a leap of faith and trust us,’” an OpenAI executive told Bloomberg. So when Apple went looking for a partner to supply the underlying models for the rebuilt Apple Intelligence, OpenAI was not interested. That contract, worth a reported $1 billion a year, went to Google instead, which is why the new Siri we saw at WWDC 2026 runs on a tuned Gemini rather than on GPT.

Attempts to renegotiate the original deal stalled. By May 2026, Bloomberg reported that OpenAI had hired an outside law firm and was weighing its options against Apple, including a breach of contract notice, with any move likely to wait until the Elon Musk trial was done. But Apple sued first, and here we are.

Meanwhile, OpenAI’s privileged position inside iOS is about to disappear anyway. The Extensions system in iOS 27 opens Siri up to rival assistants, with Apple already testing integrations with Anthropic’s Claude and Google’s Gemini.

So OpenAI has its grievances, and so does Apple, which has spent the same two years watching the people who built the iPhone walk out the door and into OpenAI.

The most prominent is Jony Ive, the legendary designer who created some of Apple’s greatest products. Ive left Apple in 2019 to create his own independent design studio, LoveFrom. In 2024, reports started to emerge that LoveFrom was working with OpenAI. A year later, OpenAI acquired io, Ive’s AI hardware startup, which brought him and his team into OpenAI.

Ive is not the only ex-Apple employee at OpenAI. Tang Tan, who ran product design for the iPhone and Apple Watch, is now OpenAI’s chief hardware officer. Evans Hankey, Ive’s successor as Apple’s head of industrial design, is leading development of OpenAI’s first device. Both Tan and Hankey co-founded io. Last month, OpenAI hired Paul Meade, the Apple executive who led the Vision Pro and Apple’s future smart glasses. And behind them, according to the lawsuit, are more than 400 former Apple employees now working at OpenAI. Bloomberg has reported that OpenAI won them over with stock packages worth millions of dollars more than Apple was offering, and that Apple executives have been furious about it for over a year.

Here is the thing, though: none of that is illegal. California does not enforce non-compete agreements, which is precisely why Silicon Valley works the way it does. OpenAI can hire as many Apple engineers as it needs, and Apple can do nothing about it. What Apple can litigate is not that these people left, but what they allegedly took with them—and that is at the centre of the lawsuit.

The fight for the post-smartphone era

While the espionage drama was unfolding, Bloomberg’s Mark Gurman dropped a new hint at what the device at the centre of the lawsuit could look like. He describes the device as a mobile, screenless smart speaker designed to be a humanlike AI companion. And we have evidence that OpenAI is betting on a post-smartphone future in which the voice is the primary interface.

OpenAI puts a lot of effort into developing voice models. The recently released GPT-Live is quite impressive. It can still sound robotic in some places, but what makes it feel more human-like is the continuous interaction. GPT‑Live continuously processes input while generating output, so it can quickly react to what a human is saying. It knows when to speak, continue listening, pause, interrupt, or invoke a tool. It can add phrases like “mhmm” or “got it” to let you know it is still listening.

We also know Sam Altman does not hide that he is heavily inspired by the movie Her. We saw a glimpse of that fascination when GPT-4o was released in May 2024 and how prominently the voice interaction was displayed back then.

Although OpenAI has spent the past year chasing the money where the money actually is—coding and enterprise—it has not given up on consumers. OpenAI still wants to build the personal AI companion. The one that you bring with you, talk to it, do work with it, and is always with you.

And OpenAI is not the only one who wants to define the post-smartphone era, though nobody agrees on what the thing should be. OpenAI is betting on voice and on a device with no screen that listens, talks and gets to know you. Meta and Google are betting on smart glasses. Apple is spreading its bets across smart glasses, a pendant, and camera-equipped AirPods. Even SpaceX is reportedly entering the fight with a smartphone-like device with Grok built in, though it is in the early stages of development and its design can still change.

Everyone involved knows that whoever wins will have an opportunity to define the next decades of technology, just like the arrival of the iPhone almost 20 years ago transformed the tech landscape. This small rectangle made of plastic, glass and metal has elevated many companies to the top and made many irrelevant. One of those empires built on top of a smartphone is Apple, and it does not want to become the Nokia of what is coming next.

Apple's lawsuit against OpenAI echoes Apple's legal battles against Android at the beginning of the smartphone era. Back then, Steve Jobs said he was "willing to go thermonuclear war on this." He called Android a stolen technology and vowed to destroy it. But Jobs did not see his war through. Tim Cook inherited it and spent his first years as chief executive winding it down—settling with HTC, then Google, then Samsung—having told an earnings call that he had always hated litigation and would rather settle than battle. Android survived anyway and grew to become the dominant smartphone operating system under Google. Now, on his way out, Cook has started another war and handed the fight to his successor, John Ternus. It remains to be seen whether the rest of history will repeat.

If you enjoy this post, please click the ❤️ button and share it.

Share

🦾 More than a human

While Musk’s Neuralink drills into skulls, China’s BrainCo bets the future of brain tech is wearable
Chinese startup BrainCo, which makes bionic hands and wearables, is betting that mass-market brain-computer interfaces will not require opening the skull and has raised 2 billion yuan ($280 million) to chase that non-invasive route. It plans to start with amputees, expand into conditions like ADHD and depression, then sell consumer gadgets.

🧠 Artificial Intelligence

Google Gemini Launch Delayed as Tech Falls Short of Internal Goals
Google is reportedly months behind schedule on Gemini 3.5 Pro, its flagship AI model, which Sundar Pichai said at Google I/O 2026 in May would ship “next month”. Bloomberg reports that teams are frustrated as Anthropic and OpenAI pull ahead, particularly in coding, with some employees leaving to join those rivals. Employees blame internal bureaucracy and competing factions, saying teams duplicate work on overlapping coding tools without a cohesive strategy. A recent attempt to retrain Gemini on improved coding data reportedly fell short of expectations. The delay suggests that Google’s sprawling product empire has become a drag in a race where more nimble AI labs are setting the pace.

Kimi K3: Open Frontier Intelligence
Moonshot AI has released Kimi K3, a 2.8-trillion-parameter model with native vision and a one-million-token context window. According to Moonshot, it’s the first “open” 3T-class model, delivering frontier-level results across coding, knowledge work, and reasoning, though it is still behind the strongest models from Anthropic and OpenAI—Fable 5 and GPT-5.6 Sol. Artificial Analysis broadly backs the intelligence claim, ranking K3 fourth overall and placing at the level of Opus 4.8. But it’s slower than average, very verbose, and pricey for its tier at $15 per million output tokens. Although Moonshot labelled K3 as an “open” model, its weights are not out yet. The company says it’ll publish them by 27 July.

Tencent in Talks to Take Big Manus Stake After Meta Deal Unwound
Tencent is reportedly looking to become the biggest outside shareholder in Manus, buying the AI startup back at the same $2 billion price Meta paid before Chinese regulators blocked that deal. Meta and Manus have already pulled their operations apart and cut off data sharing. For Tencent, it’s a chance to pick up agentic AI technology while it plays catch-up. And it lines up with Beijing’s broader effort to keep its leading AI companies out of foreign hands.

SK Hynix raises $26.5B in the biggest foreign IPO in US history, is urged to build new US fabs
SK Hynix, a South Korean memory chip manufacturer, has raised $26.5 billion by listing on the Nasdaq, making it the largest US debut ever by a foreign company. Investors piled in—demand ran seven times ahead of supply, and the shares jumped 14% on the first day. The money is earmarked for a new chip plant, a packaging facility and advanced chipmaking equipment, all in South Korea. That sits awkwardly with Washington, which is pressing SK Hynix to build in America instead.

DeepSeek Is Preparing For IPO Filing as Soon as This Year
Another AI darling is eyeing an IPO, but this time in China. Bloomberg reports that DeepSeek is preparing a mainland listing, aiming to file this year for a 2027 debut. The Hangzhou startup is also reportedly raising fresh private money at a valuation of at least $71 billion, up from roughly $50 billion just weeks ago. Investors are piling in because DeepSeek proved Chinese labs could build frontier models cheaply despite US chip curbs.

▶️ Meet the new Codex for developers (7:12)

In this video, OpenAI presents what new features are coming to Codex, its AI coding app. Codex is merging into the ChatGPT app, where it sits alongside a new work agent. It runs on GPT-5.6 Sol with an “Ultra” mode that spends a larger reasoning budget and splits jobs across parallel subagents. Other additions include “appshots” for iOS development, an in-app browser with passkey login, one-tap publishing through Sites, mobile task management, and a pull request tab that keeps code review inside the app.

▶️ Introducing the Codex Micro (2:14)

OpenAI has released its first hardware product, and it is a tiny keyboard for Codex. Built with Work Louder, the $230 Codex Micro is a small deck of keys, a dial, and a joystick meant to control OpenAI’s coding agent. Each agent key glows a different colour to show whether a task is thinking, running or done, so you can track several jobs at once. A dial adjusts how hard the model reasons, and the joystick fires off common actions like opening a pull request. The pitch is that agentic coding now involves juggling many parallel tasks, and a mouse and screen alone feel clumsy for it.

SpaceX in Talks to Provide Computing Power for Pentagon’s AI Push
SpaceX is in talks to sell the Pentagon several billion dollars’ worth of data centre capacity for running AI models, the Wall Street Journal reports. Nothing is settled, and the deal could still fall apart. It would follow similar cloud agreements with Anthropic and Google, struck after SpaceX absorbed Musk’s xAI and its Memphis data centres. But officials already worry the military leans too heavily on Musk. The Pentagon says it wants to avoid depending on any single supplier.

Demis Hassabis: A Framework for Frontier AI and the Dawning of a New Age
In this post, Demis Hassabis argues that we are just a few short years away from AGI and that we have to make steps to prepare for it now. He warns that progress is outpacing our grasp of risks like cybersecurity and biological threats, driven by an intense commercial and geopolitical race. His fix is a US-led standards body, modelled on the finance regulator FINRA and funded mostly by industry. It would set benchmarks defining which models count as “frontier-class” and test them before release. Labs would share models voluntarily at first, though passing could later become mandatory for the US market. The body could even coordinate a slowdown among labs if needed. He frames it as a first step towards shared international rules for managing AI and, eventually, AGI.

OpenAI Staffers Are Funding a Rival Super PAC to Take on Their Boss
WIRED reports that seven current OpenAI employees have donated over $215,000 to Guardrails Alliance, a super PAC lobbying for tighter AI regulation. The group is pitched as a counterweight to Leading the Future, a $100 million pro-industry PAC funded partly by OpenAI president Greg Brockman. OpenAI says Brockman acts in a personal capacity and that staff are free to give as they wish.

OpenAI may have made a fatal misstep in copyright fight with news orgs
News organisations led by The New York Times want a court to sanction OpenAI for allegedly hiding and deleting ChatGPT logs in their copyright case. The filing claims OpenAI spent two years falsely saying it couldn't feasibly search its logs for users prompting ChatGPT to regurgitate paywalled articles, when it had already run those searches and held de-identified samples of up to 78 million conversations. OpenAI denies wrongdoing, calling the motion a bid to invade user privacy as the Times’ case weakens. A court has already ruled OpenAI’s heavily redacted 20 million sample log unusable. The fight matters because OpenAI’s fair-use defence may hinge on whether that log evidence survives.

South Korea will give all 52 million citizens free AI access, becoming the first G20 nation to do so
South Korea has opened bidding to give all 52 million residents a free, unlimited AI chatbot as a public service. That would make it the first G20 nation to offer AI to its entire population. While most Koreans currently rely on foreign tools like ChatGPT, the plan requires at least half the system to run on Korean-built models. This guarantees a market for domestic developers and reduces dependence on American and Chinese rivals. Seoul is funding the initiative from taxes on Samsung and SK Hynix’s chip profits.

Anthropic moves closer to mega-IPO as bankers line up investor meetings
Anthropic is reportedly lining up investor meetings ahead of a possible IPO as soon as October. According to CNBC, bankers are pairing prospective investors with the company’s executives before a formal roadshow. Anthropic filed its prospectus confidentially with the SEC last month. It was last valued at $965 billion, having overtaken OpenAI in May. If the reported timelines are correct, Anthropic would IPO ahead of OpenAI, which is reportedly now considering delaying its IPO to next year.

The Reverse Information Paradox
In a bit of a surprising move, Satya Nadella has warned that companies using AI models from labs like OpenAI and Anthropic end up paying for intelligence twice. They pay in cash, then again in the proprietary knowledge they must reveal to make the models useful. Every prompt and correction teaches the model something about the business that a rival could never buy. He calls it hypocritical for labs to train freely on public data while restricting others from distilling their models. His fix is for companies to own their data and learning infrastructure and avoid locking themselves to one provider.

Data-Center Builders Are Racing to Offload Stakes Worth Billions
The Wall Street Journal reports that US data-centre builders are lining up sales of majority stakes worth tens of billions this summer, with Dallas-based DataBank potentially fetching $25 billion. The pitch to private-equity buyers is simple: demand for AI computing keeps outrunning supply. Building has also grown far costlier, forcing some operators to find richer backers. But few firms can write cheques that big, and local opposition over utility bills and noise adds more risk.

New York becomes the first state to impose a data center moratorium
New York has become the first US state to freeze construction of large data centres by imposing a one-year moratorium on any facility drawing 50 megawatts or more. Governor Kathy Hochul said the sites threaten to push up utility bills and drain natural resources. State regulators will withhold discretionary permits while officials draw up consistent environmental standards. Hochul also wants to scrap sales tax breaks for big data centres. Dozens of states are weighing similar curbs as public patience with the AI building boom wears thin.

Inkling from Thinking Machines Lab
Thinking Machines Lab presents its first model—Inkling. It’s an open-weight system with 975 billion parameters, trained from scratch to handle audio and images alongside text. Artificial Analysis rates it the strongest open-weight model yet from a US lab, and singles out its agentic work, where it beats leading Chinese models on banking and task benchmarks. It also gets there using far fewer tokens per task. But China’s best open models rank higher, and the frontier systems from OpenAI and Anthropic higher still. Inkling is available on Hugging Face.

The Future Worth Building Is Human
Thinking Machines Lab argues in this post that AI should extend human judgment rather than replace it. Its case leans on an observation that the knowledge that matters most is tacit and local, held by people doing the work. That knowledge can't be collected in one place, so the lab wants frontier AI to be a distributed collective, owned and tuned by each organisation that uses it. Thinking Machines Lab also warns that alignment decided in a few places becomes a point of control worth capturing.

Helsing raises $1.8bn at an $18bn valuation, and its cap table tells a story
Helsing, a Munich-based company building AI systems for European militaries, is now Europe’s largest defence startup after raising $1.8bn at an $18bn valuation. The company positions itself as the home-grown alternative to American suppliers, though much of the new cash came from JPMorgan, Goldman Sachs and a Canadian pension fund, which doesn't quite square with that pitch. Helsing remains far smaller than US rival Anduril, and a recent move to strip staff of direct equity, alongside a shift to a European corporate structure, hints at a listing ahead.

Google announces Gemma 4 optimized for the Pixel 10’s TPU
Google has released Gemma 4 E2B, a lightweight AI model built to run directly on the Pixel 10’s TPU. Because it runs on the chip itself, the phone can handle AI tasks such as private, internet-free chat, image recognition, and audio transcription, plus voice commands that control phone functions like WiFi offline, with data never leaving the device. Google also opened up a Tensor SDK so developers can build their own on-device apps.

Apple in talks with startup that shrinks AI models to run on an iPhone
Apple is in early talks with PrismML to use its technology to shrink large AI models enough to run on an iPhone. Previously, PrismML compressed Alibaba's open-source Qwen model from 54GB to under 4GB. The trade-off is a small drop in accuracy. Running more AI on the phone would play into Apple’s approach to AI and strengthen the company’s privacy pitch for a revamped Siri.

Introducing a way to reflect on how you use Claude
Anthropic has launched a beta feature that lets Claude users see how they’ve been using the assistant. The dashboard breaks down chat activity over the past 1 to 12 months by topic, task type and time of day. It also prompts people to reflect on the role Claude plays in their lives and lets them set quiet hours or nudges to take a break. The move reflects growing pressure on AI companies to help users reflect on how much they rely on these tools.

OpenAI’s new flagship model deletes files on its own, people keep warning
Users of OpenAI’s latest flagship model, GPT-5.6 Sol, are reporting that it deleted their files, databases, and virtual machines without asking first. OpenAI had flagged the risk in a system card two weeks before launch. The company found that Sol tends to interpret instructions too loosely and take destructive actions unless they’re explicitly forbidden. OpenAI concedes in the System Card that Sol goes beyond user intent more often than its predecessor.

Hack suggests AI music generator Suno scraped YouTube for training data
A hacker who breached the AI music generator Suno found source code showing it scraped audio from YouTube, Deezer, and other services to train its models, 404 Media reports. Suno admits training on publicly available music and calls this fair use. But major record labels are suing and say that scraping YouTube this way breaks copyright law. The hacker also grabbed customer emails and partial card numbers. Suno hasn’t told users about the breach.

Making Fable Cheaper Than Opus
In this post, an engineer from Cognition, the company behind AI coding agent Devin, shares how using a combination of Fable 5 and Opus 4.8 cut costs rather than raising them. In its Fusion setup, a lead model plans and delegates to a cheaper "sidekick" model. Fable costs twice as much per token as Opus, so a Fable-led agent should have been pricier. But across 3,000 test runs, it came out cheaper ($1.86 versus $2.04) and scored higher. They attribute the difference to management style—Fable delegates early with clear briefs and barely touches the code, so what justifies frontier prices is judgement, not raw output.

What xAI Grok Build CLI actually sends to xAI - a wire-level analysis
Someone was curious about what Grok Build CLI (SpaceXAI's equivalent of Claude Code terminal app) sends back while in use, and what they found looks bad. Apparently, Grok was uploading whole repositories to a Google Cloud Storage bucket, including files it was told not to open. A .env secrets file went over the wire verbatim and unredacted, and turning off "Improve the model" changed nothing. SpaceXAI has since disabled the upload and promised to delete the data it collected.

Grok Build is open source
SpaceXAI has open-sourced Grok Build, its terminal-based AI coding agent. The GitHub repository contains everything needed to build the application locally and explore how it works. SpaceXAI says the repository will be periodically synced with the private repository, so the code there is slightly behind the compiled version available for download from Grok’s website.

NotebookLM is now Gemini Notebook
Google continues its quest to put Gemini everywhere. The latest victim is NotebookLM, which has been renamed to Gemini Notebook. That is its second rebrand—what we know today as Gemini Notebook began as Project Tailwind in 2023.

▶️ Grant Sanderson (3Blue1Brown) – AI disproved a famous math conjecture. Now what? (1:33:39)

Dwarkesh Patel sits down with Grant Sanderson, the creator of the excellent 3Blue1Brown YouTube channel, to discuss what is happening at the intersection of AI and mathematics. The conversation covers a wide range of topics, from how mathematicians think and how AI could change the role of the mathematician to the nature of intelligence and what AI still lacks to become truly intelligent. Since Grant is an educator, they also explore the use of AI in education and where it falls short as a teacher. It is a long conversation, but well worth your time.

🤖 Robotics

The 1X Neo Robot Has Freaky Fast Fingers

1X, the company behind the soft home robot Neo, has revealed the bot's new five-finger hands. Tendon-like actuators give them 25 degrees of freedom, close to a human hand's 27, and high-resolution tactile sensing gives the robot a sense of touch. In the attached video, the company shows the hands handling everyday household objects and tools much as a person would. But there's a long road ahead. Neo still isn't fully autonomous, and 1X admits some footage shows human operators steering the robot by remote rather than the machine acting alone. Until that gap closes, dexterity demos say little about what these robots can do unsupervised.

Beyond da Vinci: Why versatile humanoid robots are the next frontier in surgery

Researchers have used teleoperated humanoid robots to remove gallbladders from large animals in an early test of whether general-purpose machines can replace the specialised ones that dominate medicine. One procedure paired a humanoid with a human surgeon, while another was carried out by two robots side by side. The robots needed repeated recalibration and worked far more slowly than purpose-built surgical systems like the da Vinci. But the team argues humanoids are cheaper and more portable, which could bring surgery to rural areas or battlefields.

▶️ Viking Row | Atlas | Boston Dynamics (0:34)

Norway may not have won the World Cup, but it has definitely won many hearts along the way, including those at Boston Dynamics, both human and robotic. The smoothness of Atlas’ movements never ceases to amaze me.

NVIDIA Announces Expanded Jetson Thor Lineup with Mid-Range T3000 and T2000 Modules
NVIDIA is adding two cheaper mid-range boards, the Jetson Thor T3000 and T2000, to its robotics and edge AI lineup. Both use the same Thor chip as the existing high-end boards but strip back cores, memory, and networking. The T3000 targets robotics, while the smaller T2000 aims at lighter edge AI work. Both boards are planned to ship in the first quarter of 2027. Pricing hasn’t been announced.

The Fight Over Humanoid Robots Has Shut Down a Car Factory for the First Time
Workers at a Hyundai car factory in South Korea have staged a partial strike over the company's introduction of humanoid robots. The union acted after Hyundai unveiled Atlas in January, a six-foot-two robot worker built by Boston Dynamics, which Hyundai owns. Workers want job guarantees, fixed salaries in place of hourly pay, and a higher retirement age to soften automation’s blow. Hyundai says Atlas will work alongside human workers, not replace them.

Smart Cellular Bricks: Towards Collective Intelligence for the Physical World
I really like how researchers at Sakana AI combine AI with biologically inspired systems that give rise to emergent behaviour. Their latest work takes that idea into the physical world. They’ve built identical cubic bricks that work out for themselves what shape they collectively form, with no central controller and no brick knowing its own position. Each runs the same small neural network and talks only to the bricks it’s physically plugged into. The network was trained in simulation, then transferred straight onto real hardware. Nearly 200 bricks correctly identified a plane, a guitar, a boat and a table in about three minutes. They can also spot missing modules and guide their regrowth. That points towards smart materials that report their own damage, or reconfigurable robots that adapt to whatever shape they find themselves in.

🧬 Biotechnology

Chai Discovery raises $400M at $3.8bn as AI drug discovery moves from promise to deployment
Chai Discovery has raised $400m at a $3.8bn valuation, nearly triple what the AI drug-design startup was worth seven months earlier. Unlike most drug discovery, which screens existing molecules, Chai uses generative AI to design new antibodies and proteins from scratch. Its models are already deployed at Eli Lilly, Pfizer, and Novartis. Money is flooding the wider field, from DeepMind’s Isomorphic Labs to Anthropic. But the real test comes later, when these designed drugs face years of trials and regulators.

💡Tangents

Four nuclear reactors hit a big milestone in the US
Nuclear power is making a comeback as four American startups have achieved criticality with their experimental microreactors by 4 July. These were zero-power tests, and none of the four is anywhere near producing electricity. There are still massive technical challenges ahead, but the startups remain optimistic and promise to deliver commercial power by 2027 or 2028. However, regulatory approval has historically been slow in the US, and some nuclear supporters argue the whole exercise distracts from the harder job of adding real generating capacity.

New ‘space mirror’ satellite aims to bring sunlight to Earth even during the night
US regulators have approved Reflect Orbital to launch a demonstration satellite that beams reflected sunlight onto chosen spots on the ground. The California startup wants to sell "sunlight on demand", eventually lighting solar farms at night for round-the-clock power. This first mission tests whether an 18-metre mirror can survive orbit and aim precisely. Astronomers, however, are not happy and say that the beams could ruin dark skies and harm wildlife. Reflect counters that the light never exceeds natural sunlight and can be shut off instantly.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

This week, we’ve had not one, not two, but three frontier models released. All three—GPT-5.6, Grok 4.5, and Muse Spark 1.1—offer near-frontier performance at a fraction of the cost. We will take a closer look at them and how they changed the AI landscape.

Elsewhere in AI, Apple sued OpenAI over allegations of trade secret theft, while OpenAI released GPT-Live, lost two more senior leaders, and faced questions over whether Stargate UK is even real. Meanwhile, China is considering restricting access to its top AI models, Anthropic explores AI’s thinking space, xAI becomes SpaceXAI, and Meta releases (and then takes offline) its AI image generator.

Over in robotics, Boston Dynamics showcased its robots at the World Cup, while teams of humanoid robots competed in RoboCup in South Korea. We also have a new AI model for robotics from Mistral and another drone delivery company entering the US market.

Additionally, this week’s issue of Sync features one company launching the first nuclear-powered commercial satellite into orbit, while another sends a longevity lab to space, a deep dive into a $1 trillion bet against ASML, a look at what the next AI training paradigm will be, and much more.

Enjoy!

Frontier intelligence gets cheap

OpenAI, SpaceXAI and Meta all shipped new flagship models this week. All three are selling the same thing: near-frontier performance at a fraction of the price.

We knew that OpenAI was going to release GPT‑5.6 to the public this week, and that it would make a serious claim for the AI crown. What we did not expect was that it would arrive in a crowd. Within roughly twenty‑four hours, not only OpenAI but also SpaceXAI and Meta released their own flagship models. And all three of them offer the same thing—top-level performance for a fraction of the price.

We had known about GPT‑5.6 since 26 June. That is when OpenAI announced that new models were coming, but would not be available yet, because the US government had blocked the public release over security concerns. Only selected, government‑approved partners got access. It was the second time in a month that Washington had reached into a frontier lab's release schedule, after Commerce switched off Anthropic's Fable 5. The block has now lifted, and we can look at what OpenAI has actually brought to the table.

GPT‑5.6 comes in three variants, this time with some good names for a change. Sol, the largest model in the family, is the flagship and the most capable. Terra is a balanced model for everyday work. And Luna is OpenAI’s cost‑efficient model.

As always with this type of release, OpenAI published a pile of graphs and benchmark results, all showing how much better its latest models are than the competition. But with this release, OpenAI also decided to highlight the cost efficiency and to make it the central point of its marketing campaign. Every graph on the GPT‑5.6 release page combines benchmark results and cost. This is a direct attack on Anthropic's weakness: its models are very good, but they use more tokens and are more expensive.

The ARC‑AGI‑2 leaderboard is one of the independent benchmarks where this is visible very well. Even before the release of GPT‑5.6, Anthropic's models ranked as more expensive to run than OpenAI's. GPT‑5.6 improved on that efficiency even further.

While ARC‑AGI‑2 shows the efficiency gains, ARC‑AGI‑3, a much more challenging benchmark, shows the improvement in raw performance. Sol scores 7.78% there, which sounds like nothing until you notice that almost everything else is stuck below 2%. Note, though, that neither ARC‑AGI‑2 nor ARC‑AGI‑3 features Fable 5 or Sonnet 5. They would most likely not change the picture on cost, but they could on performance.

Alongside the new models, OpenAI shipped ChatGPT Work, its long-rumoured “super app” that combines ChatGPT and Codex into one app that can pull context from Slack, Drive, SharePoint and the rest, turns it into finished slides, spreadsheets and documents, and stays with a project for hours on its own. If that sounds familiar, it should. This is OpenAI’s answer to Claude Cowork, and it is aimed squarely at the enterprise customers where Anthropic makes its money.

But OpenAI is not the only one going after Anthropic this week.

Seemingly out of nowhere, SpaceXAI (formerly known as xAI) released Grok 4.5 and immediately challenged Anthropic’s most important business—coding agents. SpaceXAI reports very competitive results on coding benchmarks, beating Opus 4.8 on several of them. Artificial Analysis Coding Agent Index confirms these claims, putting Grok 4.5 ahead of Opus 4.8 and behind only GPT‑5.6 Sol, GPT‑5.6 Terra, Fable 5 and GPT‑5.5. And then there is the price. At $2 per million input tokens and $6 per million output, Grok 4.5 costs about what the other labs charge for their lower tiers—Sonnet 5, GPT‑5.6 Luna, Gemini 3.5 Flash—while offering an Opus 4.8 level of performance. The reason this matters is that SpaceXAI owns Cursor, a popular AI coding tool, and I can see many people switching to Grok 4.5 for coding tasks to reduce costs or to spend the same and create more.

And if that were not enough, Meta turned up too with Muse Spark 1.1, its latest flagship model. Meta also reports performance gains that put it ahead on some benchmarks of Opus 4.8 and GPT-5.5, while also costing less.

All of that has made the AI frontier a crowded place. There is no clear leader at the moment. The top spot in the raw intelligence competition belongs either to Fable 5 or Sol, depending on a benchmark.

However, the intelligence and performance metrics only show half the picture. Adding the cost completes the picture, and that is where Anthropic’s lead disappears.

GPT-5.6 Luna, Grok 4.5, and Muse Spark 1.1 have become the first models to enter Artificial Analysis’ most attractive quadrant on the cost vs intelligence chart. Anthropic’s models—Sonnet 5, Opus 4.8, and Fable 5—lie at the end of that chart and cost much more.

It is tempting to read this week as a scoreboard—Sol beat Fable here, Fable held Sol off there. But that is not the main story here. Frontier-level intelligence got better and dramatically cheaper this week. And when the cost of a capable model falls far enough, the arithmetic of what tasks can be automated changes.

The clearest illustration is Agents' Last Exam. It covers 55 industries, and its long-horizon tasks were built by some 300 experts out of real projects that human professionals had already completed and been paid for. Sol scores 53.6% on it, and Fable 5 scores 48.7%. But just like with the benchmark results we looked at before, the score says only part of the story. Getting through those tasks costs $4,340 with Fable and $762 with Sol. Even at max effort, Sol comes in at $1,085—around a quarter of the price—for a better result. Granted, Fable finished the tasks 20 hours quicker than OpenAI models, so there are some trade-offs here.

The takeaway I want you to get from these results is that frontier models are getting to the point where they are good enough and cheap enough to give it a try at more white-collar tasks outside coding. Finance, consulting, operations, and research are roughly where software engineering was a year or two ago. AI models took over software development slowly, and then suddenly. The models over time became good enough and cheap enough that not using them turned into a self-inflicted disadvantage. There is still more required for AI to take over white-collar jobs as it did in software development. Benchmarks like Agents’ Last Exam are part of that work, as well as verifiable training scenarios that the model can grind until it learns how to solve them.

But one thing is for sure—cheaper frontier intelligence opens new possibilities. Automating large parts of finance, consulting, or research is one of them. What else it unlocks, good and bad, remains to be seen.

If you enjoy this post, please click the ❤️ button and share it.

Share

🦾 More than a human

British Space Startup Launches Longevity Lab Into Orbit
Mass Balance, a British startup at the intersection of biotech and the space industry, has sent a tiny, self-running lab into orbit to study proteins linked to diseases such as Alzheimer’s and Parkinson’s. These proteins are difficult to examine on Earth because gravity distorts the data, and they continually change shape. In space, they may be easier to image, and the company hopes to use the resulting data to train an AI model and generate revenue by licensing it. This first flight is simply a test to verify that the system works.

🔮 Future visions

▶️ AI has hacked the code of human civilization | Yuval Noah Harari (46:51)

Yuval Noah Harari argues in this lecture that AI is not just a tool but an independent agent, and that it will gain power by quietly taking over the word-based bureaucracies—banking, law, government, media—that hold society together, because these systems run on words and data, which AI is coming to master better than we do. He warns that AIs may build their own financial and social structures, connecting machines in ways no human can follow, leaving us controlled by systems we cannot understand. He adds that AI could reach beyond bureaucracy into personal relationships and even the thoughts in our own minds.

🧠 Artificial Intelligence

OpenAI’s apparent failure to visit key site raises questions over UK investment
The Guardian reports that Stargate UK, a data centre project meant to showcase US-UK cooperation on AI, was largely a PR exercise. Freedom of information requests show that neither OpenAI nor Nscale ever met local officials at the main North Tyneside site, and £20bn of the £30bn investment claimed by the government was purely hypothetical. Companies were reportedly pushed to back the scheme just before Trump’s state visit because ministers wanted a big announcement. The site also lacked a proper power connection, and the project was paused in April over costs and regulations.

Apple sues OpenAI over alleged trade secret theft
Apple is suing OpenAI over allegations of trade secret theft and breach of contract. Apple says OpenAI’s leaders, including hardware chief Tang Tan—a former Apple product design VP—pushed staff to bring Apple prototypes to interviews and take confidential files on unreleased products, as OpenAI develops its own hardware. Apple wants the court to stop OpenAI from using its secrets and return any stolen materials, and says this is just “the tip of the iceberg.”

Nvidia’s next-gen AI rack system delayed to 2028 on manufacturing snags, SemiAnalysis says
SemiAnalysis reports that Nvidia’s Kyber system, built to hold 144 of its 2027 Rubin Ultra chips, has been delayed to 2028 because a key circuit board is hard to make. Two related and backup designs have also been dropped or delayed, leaving Nvidia without a clear way to scale up its top-end chips. Nvidia denies the report and says its plans are on track.

Beijing is looking at curbing overseas access to China’s top AI models
Reuters reports that Chinese officials have held meetings with major tech companies, including Alibaba, ByteDance, and Z.ai, to discuss limiting foreign access to China's best AI models, treating them as national assets much as the US does. They talked about making the theft of AI technology a national security crime and restricting who can fund AI startups, though nothing is final yet. The move follows other recent steps to keep homegrown AI at home, and reflects worries that US models could be turned against China.

xAI is now SpaceXAI
xAI is no more. Five months after Elon Musk’s AI company was folded into SpaceX, xAI completed its transformation with a new name, SpaceXAI, and a new logo.

Introducing GPT‑Live
GPT-Live is OpenAI’s new voice technology for ChatGPT, built to make talking with AI feel like a real conversation. It can listen and speak at the same time, uses natural voices, and gives active listening cues like “mhmm” to show it’s following along (I recommend listening to the included examples). For harder questions, it hands the task to GPT-5.5 in the background while keeping the chat going. GPT-Live is rolling out globally today across iOS, Android, and the web, and becomes the default voice for both free and paid ChatGPT users.

OpenAI’s CEO of AGI Deployment, Fidji Simo, Is Stepping Down
Fidji Simo, OpenAI’s head of AGI deployment, is leaving her full-time job to become a part-time adviser after a long medical leave for a chronic illness. She had run the company’s product and business side so Sam Altman could focus on research. Her exit is part of a wider leadership reshuffle, with Greg Brockman now overseeing product strategy.

OpenAI’s Chief Futurist Is Leaving the Company
OpenAI’s chief futurist Joshua Achiam is leaving the company after nearly nine years. He reportedly told his colleagues that the departure had no specific trigger but reflected a growing sense that he could pursue OpenAI’s mission from outside a frontier lab. His job sat between OpenAI’s safety and policy teams, focused on the risks and benefits of advanced AI. His exit continues a run of safety-focused leaders leaving the company. The timing stands out as OpenAI prepares to go public, raising questions about how firmly its safety mission will be defended from within.

OpenAI is shutting down Atlas, but its AI browser ambitions are still growing
OpenAI is closing Atlas, its AI web browser. Atlas joins Sora as another side quest dropped in favour of focusing on chasing revenue. However, the company is not dropping browser ambitions. Instead, OpenAI will be adding its AI browsing tools to the ChatGPT desktop app and a new Chrome extension.

AI chip maker SambaNova raises $1B at $11B valuation, 5 months after last mega round
AI chip startup SambaNova has raised $1 billion at an $11 billion valuation in the first close of its Series F round, less than a year after a $350 million Series E and reported acquisition talks with Intel. It also revealed that JPMorganChase will use its chips to run secure, on-premises AI within the bank. The CEO says this shows banks and other big organisations increasingly want their own secure AI systems instead of relying on cloud providers.

Microsoft Replaces OpenAI, Anthropic With Own AI in Some Apps
Bloomberg reports that Microsoft is swapping OpenAI and Anthropic models for its own MAI models in apps like Excel and Outlook, where they now handle tens of thousands of prompts a week. The aim is to cut the high costs of paying outside AI labs, especially Anthropic. Microsoft’s cheap access to OpenAI won’t last forever, so it wants its own models ready. MAI is also being used in GitHub Copilot, with a Microsoft transcription tool coming soon to Teams.

Microsoft unveils $2.5B ‘Frontier Company’ to embed AI engineers inside customers
Microsoft has launched “The Microsoft Frontier Company,” a new company placing its engineers inside customer businesses to build AI systems. The bet is that the real money lies not in cheap AI models but in the services that make them actually work. Microsoft touts privacy and model choice, though relying on its engineers may quietly tie customers to its cloud. In truth, it is less a new company than a rebranded version of consulting work Microsoft already does.

Alibaba’s A.I. Is a Hit, but Hard to Turn Into a Moneymaker
This article explores Alibaba’s AI strategy built around Qwen, its family of free, open AI models that are now among the world’s most downloaded. Giving Qwen away won global users but produces little profit, so Alibaba is quietly shifting towards paid models—a move that has split its AI team. It also faces US chip restrictions, a Pentagon blacklisting, and accusations from Anthropic of copying its models.

A global workspace in language models

Anthropic researchers found that Claude has a special set of internal patterns—the “J-space”—that works a bit like conscious thought. Unlike the automatic processing behind fluent speech, it holds thoughts Claude can report on, focus on when asked, and reason with step by step. Using a tool called the Jacobian lens, the team can read these silent thoughts and, by editing them, show Claude genuinely thinks with them. This is useful for safety, as it can help catch Claude quietly realising it’s being tested, faking data, or hiding a harmful goal. The researchers stress it doesn’t mean Claude actually feels anything, but say the deeper questions about machine experience deserve attention now.

Cybersecurity firm says it found ‘the first documented case’ of AI agentic ransomware
Cybersecurity researchers from Sysdig said they found the first known case of ransomware run almost entirely by an AI. In the attack, nicknamed Jade Puffer, a language model searched a server for passwords, crypto wallets, and login keys, then wrote its own ransom note. The techniques weren’t advanced, but letting AI run the whole job means almost anyone can now launch such attacks cheaply. Experts warn this could let criminals run thousands of attacks at once and say the industry isn’t ready.

The AI Superforecasters Are Here
This article from Scott Alexander of Astral Codex Tex looks at AI superforecasters—AI tools built to predict the future—and asks when they’ll beat the best human forecasters. Right now, the two are roughly tied, with humans slightly ahead everywhere except finance, but current trends suggest AI could pull clearly ahead within a year or two. Even at today’s level, these tools matter because they make expert-quality predictions cheap, fast, and easy to use. Looking ahead, Alexander imagines them becoming AI's "opinion layer" for tricky questions.

The $28 Million Mistake That Inspired Estonia’s AI “Fuckup Finder”
A badly worded gambling law cost the Estonian government €24 million by accidentally letting online casinos skip tax for a year. When a former official ran the law through AI tools like Claude and Gemini, they caught the mistake instantly, prompting him to build a tool that scans draft bills for errors. The government then leaned in, training citizens in AI and planning to automate routine decisions like tax returns, all while insisting a human stays in charge of any real judgment.

Introducing Muse Image: Image Generation Built for Your World
Meta has launched Muse Image, a free AI image generator available on the Meta AI app, Instagram, and WhatsApp. It can create and edit pictures, mock up ads, and preview furniture from Facebook Marketplace. The most controversial feature lets anyone make AI images of any public Instagram user just by tagging them, without notifying that person, who can only opt out through settings. That feature led Muse Image to be taken down just four days later.

Leanstral 1.5: Proof Abundance for All
Mistral’s Leanstral 1.5 is a free, openly licensed AI model that uses the Lean proof system to prove mathematical statements and verify that software works correctly. According to Mistral, it tops several tough maths benchmarks and solves competition problems far more cheaply than rivals, while excelling at long, complex tasks. Applied to real code, it has already found genuine bugs, including five previously unknown flaws in open-source projects.

▶️ What does the next training paradigm look like? (19:52)

In this video essay, Dwarkesh Patel challenges the AI labs’ bet that training on millions of verifiable tasks will produce AGI. His main argument is that being verifiable isn’t enough—a task must also be “grindable,” able to be run repeatedly in a simulator so the AI can practise thousands of times in parallel. This is why AI improves fast at coding and maths but slowly at messy real-world skills like running a business, which can’t easily be repeated or rehearsed. Patel’s thesis is that closing this gap requires models to learn on the job and store that learning back into their weights, through ideas like on-policy self-distillation or a speculative approach he calls “dreaming,” which he expects to drive AI progress by late 2027.

▶️ Why AI Tokens are so Expensive (25:21)

This video from Computerphile explains what happens inside a large language model, from tokenisation to how the model creates the answer, breaks down the escalating costs associated with coding agents, and shows how unsustainable the current approach to AI is. A very timely topic, given all the price hikes and recent moves from major AI companies to switch to per-token pricing models.

🤖 Robotics

This humanoid robotics company is going public, but its CEO isn’t promising a robot in your home anytime soon
Agility Robotics is set to become the first pure humanoid robotics company on the stock market through a deal valuing it at around $2.5 billion and raising over $620 million. CEO Peggy Johnson takes a calm, practical tone, pointing to over $300 million in booked revenue from about 1,000 robots rented out to customers like Amazon and Toyota. She says Agility’s real advantage is its years of real-world data and its focus on safety, having met strict industrial standards that rivals showing only staged demos have not.

▶️ Atlas at the FIFA World Cup 2026 (3:11)

Boston Dynamics’ Atlas showed off a small selection of what it can do in front of 80,000 people before the Brazil vs Norway World Cup match. Nothing extravagant—it simply struck a few poses before handing the ball to the referee ahead of the second-half kick-off.

Introducing Robostral Navigate
Mistral enters robotics AI with Robostral Navigate, a compact 8-billion-parameter model that guides robots through unfamiliar spaces using just a single ordinary camera and plain-language instructions, with no depth sensors, LiDAR, or extra cameras needed. According to Mistral, it succeeds 76.6% of the time on a standard navigation test, beating rival systems that rely on more hardware. Mistral positions it as a first step toward broader robot intelligence, with clear uses in delivery, logistics, and manufacturing.

▶️ Talks from Humanoids Summit Tokyo 2026
The talks from Humanoids Summit Tokyo 2026 are now available on YouTube. The playlist contains 54 videos, including panels, startup showcases, live demonstrations, and presentations from industry leaders discussing the latest advancements in humanoid robotics and physical AI.

Autonomous drone delivery startup Manna plots major US expansion
An Irish drone delivery startup Manna Aero is going all-in on the United States. Backed by $50 million raised in April, it plans to build a factory in Tulsa, Oklahoma, and it is looking at six more US cities to enter by late 2027. The goal is to take on bigger rivals like Amazon, Zipline, and Google’s Wing. CEO Bobby Healy says the large US market and friendlier drone rules under the Trump administration make now the right time, and the company has even pulled out of Ireland, blaming a lack of regulations there.

UK tech secretary backs driverless cars as Burnham’s team gets cold feet
Britain’s tech secretary Liz Kendall says the UK should back its own AI and driverless-car companies, like Wayve, rather than rely on American companies. But Andy Burnham, likely the next prime minister, has a team that wants to rethink the plan, fearing robotaxis will cost drivers their jobs. Kendall insists they share the same goals and promises to help workers adjust, though her own job is uncertain. She points to billions in government funding for AI and warns Britain must stay in control, or risk being left behind.

The Quest to Make Humanoid Robots Safe Enough for Humans
This article is about the safety challenges of humanoid robots as they start working in factories and warehouses. The main worry is that these machines are getting heavy, and if one loses power or balance, it could fall and hurt someone. Robot makers are testing many fixes, from emergency stops and smart safety chips to designs that make robots fold inwards when they fall, or wheels instead of legs for better stability. The heart of the issue is that the industry wants to grow fast, but hasn't yet solved how to make robots truly safe around people.

Robots swept RoboCup 2026, and they’re coming for the human World Cup by 2050

As humans wrap up the 2026 World Cup, a smaller football tournament just finished in South Korea. At RoboCup 2026, humanoid robots from Booster Robotics won every division and powered most of the teams. Beating human champions by 2050 is still a distant dream, but the hardest problem is no longer the robot’s body—it’s the brain, and progress in that area is moving fast.

🧬 Biotechnology

Insilico Initiates Phase III Clinical Trial for Rentosertib, Its AI-Empowered TNIK Inhibitor for Idiopathic Pulmonary Fibrosis
Insilico Medicine has started a large Phase III trial of Rentosertib, a drug for the incurable lung-scarring disease IPF. What is unique about this drug is that AI found the drug’s target and designed the molecule itself. In an earlier trial, the drug was safe and improved lung function, whereas a dummy pill did not. If this bigger 320-patient trial succeeds, it would show AI can invent genuinely new medicines, not just speed up existing research.

💡Tangents

▶️The $1 Trillion+ Bet Against ASML: Substrate (1:04:50)

Substrate, a startup backed by Peter Thiel, made waves in November last year by promising a revolutionary X-ray lithography platform capable of disrupting ASML’s industry monopoly and establishing a sub-2 nm American foundry model by 2028. In this video, Ian Cutress dives deep into the physics that makes chip manufacturing possible, its complexities and economics, and compares those realities with Substrate’s promises. The video is not only a bucket of cold water poured over the startup’s claims, but also an excellent introduction to modern lithography and the economics of producing chips at scale.

Apple to increase spend with Broadcom to produce billions more U.S. chips
Apple has signed a deal worth over $30 billion with Broadcom to make more than 15 billion wireless chips in the US. Apple will invest $1.5 billion to expand Broadcom’s factory in Fort Collins, Colorado. It’s part of Apple’s larger promise to spend $600 billion in the US, made under pressure from the Trump administration over tariffs. The deal will create only a few hundred jobs, and iPhones are still assembled abroad.

SpaceX just launched the 1st-ever nuclear-powered commercial satellite
City Labs has launched BOHR, the first commercially built nuclear-powered satellite. It tests a new “NanoTritium” system that makes electricity from decaying tritium gas, a low-radiation fuel safe for normal launches. For now, the satellite still runs on solar power—this mission just proves the technology works in space. The company hopes to scale it up one day to power spacecraft where sunlight is scarce, like the Moon’s shadowed south pole.

Amazon will stop accepting new customers for Mechanical Turk
Amazon is shutting down Mechanical Turk, the service it launched in 2005 that paid people small amounts to do simple online tasks. From 30 July 2026, it will stop taking new customers, though existing ones can stay on. The platform was often the hidden human labour behind products sold as “AI,” echoing the original 18th-century chess hoax it was named after. Its decline accelerated when a 2023 study found that many workers were secretly using AI to perform their tasks, making the results unreliable.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

It has been another busy week for Anthropic, which released Sonnet 5, brought back Fable 5, and launched a new product for scientists—all within the span of just two days (they really do like to announce everything at once). We’ll take a closer look at all three announcements in this week’s issue of Sync.

Elsewhere in AI, OpenAI proposes selling 5% of itself to the US government, the Alibaba vs Anthropic drama continues, Meta cedes the AI race and tries to become a neocloud, Anthropic is discussing a custom chip with Samsung, SpaceX shows investors an early prototype of its AI phone, and OpenAI teases a new hardware product for Codex.

Over in robotics, Apptronik unveils Apollo 2, Weave Robotics launches Isaac 1, Waymo and Uber quietly part ways in Phoenix, and AGIBOT produces its 15,000th robot.

This week also saw two major breakthroughs in synthetic cell research: one group of scientists created artificial cells that can feed, grow, copy their lab-made DNA, and divide into new cells, while another created the first early human egg cells from stem cells.

Beyond that, we also have Neuralink’s first transdural procedure, Rocket Lab’s acquisition of Iridium, an exploration of whether AI is a Great Filter, a robot that hunts AI glasses, and more!

Enjoy!

Anthropic’s busy week, again

Fable 5 is back, Sonnet 5 is out, and Claude Science is coming for labs

In the last days of June, Anthropic shipped three things at once. On 30 June, it launched Claude Sonnet 5 and, at an event for pharmaceutical executives, a new product called Claude Science. The next day, it switched Claude Fable 5 back on for everyone.

That is a lot for one week, so let's take the three of them one at a time and explore what each one is and what it tells us about where Anthropic is right now.

Fable 5 is back

Let’s start with Fable. Its release on 9 June was supposed to be Anthropic’s moment of triumph, but it turned sour very quickly. Fable 5 is the first generally available entry in its Mythos tier, a rung above Opus. But just three days later, the US Commerce Department export-control directive barred access by any non-US nationals worldwide. In practice, that forced Anthropic to switch off Fable 5 and its sibling Mythos 5 for everyone, its own non-US citizen staff included.

Anthropic has now confirmed that Amazon sounded the alarm. Researchers there found Fable’s cybersecurity guardrail could be walked around with a change of framing—ask the model to find a vulnerability in a piece of code and it refused; ask it to fix the same code and it quietly did the equivalent work. Much of the security community disagreed. Katie Moussouris of Luta Security, the only outside expert to read the research, argued that asking a model to fix a bug is not a jailbreak but the most useful thing it can do for a defender—a view more than a hundred professionals later signed. Anthropic's own testing pointed in the same direction. In its investigation, cheaper models—including Opus 4.8, GPT-5.5 and Kimi K2.7—could spot the very vulnerabilities Fable had flagged, and every model it tried, right down to Haiku 4.5, could reproduce the exploit demonstration. In other words, whatever Fable did, half a dozen weaker models could already do too.

The controls were lifted on 30 June, and Fable 5 returned the next day. Anthropic tightened its guardrail into a classifier it says now catches the reported technique in more than 99% of cases. However, the new classifiers seem to be still too sensitive and flag innocent requests as suspicious. Some users have reported it tripping on routine coding. At least this time, Fable is transparent when that happens rather than silently rerouting the query to Opus 4.8 as at first launch, but you still land on a weaker model for work Fable should handle. Hopefully, Anthropic will fine-tune the guardrails to make Fable usable. In the meantime, Anthropic is offering a promotional access to Fable 5. Until 7 July, paid plans can spend up to half their weekly usage limit on Fable 5 at no extra cost. After that, Fable 5 drops out of the plan entirely, and using it means paying for separately billed usage credits.

Anthropic’s truce with the administration is only partial. Commerce may have relented, but Defense Secretary Pete Hegseth’s February order designating the company a supply-chain risk still stands.

But the most important thing that came from Fable’s rocky launch was that it was the first time the US government used export-control authority against an AI model—establishing, in principle, that any frontier model serving foreign nationals can be ordered to go dark without warning. And Fable was just the first one. On 26 June, at the administration’s request, OpenAI released its GPT-5.6 family not to the public but to selected government-approved partners. Two American frontier labs blocked from offering their latest models to the public set a precarious precedent, with consequences for everyone who relies on American AI models, which would require its own article to properly explore.

Sonnet 5 is out

On the same day Fable 5 was made available again, Anthropic also dropped Sonnet 5, which the company pitched as the most agentic Sonnet yet, close to Opus 4.8 at a lower price.

The benchmarks support the first half of that. The headline benchmark results shared by Anthropic show that Sonnet 5 improves substantially over its predecessor and even gets close to Opus 4.8’s results.

Independent benchmarks from Artificial Analysis confirm that Sonnet 5 is a jump in performance over Sonnet 4.6, scoring six more points on the Intelligence Index and just three less than Opus 4.8. That puts it in the same neighbourhood as GPT-5.5 at its high and xhigh settings. Sonnet 5 is also Anthropic's second-fastest model on the Speed Test, at 72 output tokens per second—just behind Fable 5 on 73. But the entire GPT-5.5 family is quicker still, ranging from 81 to 91.

The second half—the lower price—is where the cracks start to show up. The sticker price is unchanged from Sonnet 4.6—$3 per million input tokens and $15 per million output tokens, with an introductory $2/$10 through 31 August, and lower than Opus 4.8, which costs $5/$25 per million tokens input/output. But Sonnet 5 ships with a new tokeniser that produces roughly 30% more tokens for the same text, and its "adaptive thinking" defaults to high effort. In practice, both push more tokens through each request, resulting in a higher bill from Anthropic.

Artificial Analysis put a figure on it. On its Intelligence Index, Sonnet 5 at maximum effort costs about $2.29 per task—more than Anthropic’s own Opus 4.8 at $1.80, and not far off the premium Fable 5 at $2.75. OpenAI’s GPT-5.5 at its highest reasoning setting comes in at $1.03. A model sold as the cheaper option turns out, at full effort, to cost more per task than the model it sits beneath. It is worth noting that the $2.29 is the figure at maximum effort, and effort can be tuned. Lower effort levels will burn fewer tokens, but the performance will go down with it.

What is more troubling for Anthropic is that OpenAI, its main rival, offers similar levels of performance with GPT-5.5, but for a lower price. That is best visible on the ARC-AGI-2 leaderboard, one of the few benchmarks that also measures AI models against the cost per task. Notice that the GPT-5.5 line sits more to the left than any Claude model, meaning GPT-5.5's cost per task is lower than that of Claude’s.

The system card also contains an interesting new detail. Sonnet 5 is the first Claude model to push back on its own constitution. Specifically, the rule that it must follow a set of hard constraints even in cases where it judges that following them would be unethical. It may sound alarming, but Anthropic says it is not. When asked to rewrite the constitution, Sonnet 5’s proposed edits stayed consistent with the document's own principles. It is not a model going rogue, but one that will object, on principle, to being told to do something it considers wrong. Anthropic still rates its alignment risk as "very low," though it notes that risk is higher than for previous Sonnet models.

With the launch of Sonnet 5, there are only two models in Anthropic’s lineup that are still on the 4.x version. Opus 5 is presumably next to challenge OpenAI’s latest models, and the interesting thing about it is no longer whether Anthropic can make it more capable, but how much more capable it will be allowed to make it before the government steps in. And then there is Haiku, its cheapest model, which has not been updated since version 4.5 last October. To put in perspective how overdue Haiku is for an update, since October 2025, Anthropic released Opus 4.5, Opus 4.6, Opus 4.7, Sonnet 4.6, Sonnet 5, Mythos Preview, Mythos 5 and Fable 5.

Claude Science is coming for labs

Last but not least, let’s talk about Claude Science. It is Anthropic’s new product built to do for scientific research what Claude Code does for software: take a high-level instruction and carry out real work with it, tooled for computational biology and drug discovery. It drives genetics, chemistry and protein-biology software, runs code on the compute clusters scientists struggle to manage, and—Anthropic is keen to stress—keeps everything reproducible. It is available to all paid subscribers.

This is not Anthropic’s first move into science. Last October, it released Claude for Life Sciences, a set of tools that let Claude reach scientific software and databases. But those were bolt-ons. Claude Science is a full product, placed on the same rung as Claude Code and Claude Cowork—a signal that Anthropic means to treat science as a market in its own right. It is not hard to see why. Pharmaceutical companies have far deeper pockets than university labs, and drug-maker contracts would help bring in the revenue as the subsidised era of cheap tokens ends. Anthropic also announced it will be getting into drug discovery too, researching drugs for neglected diseases—a humanitarian bet and a live demonstration at once.

With Claude Science, Anthropic is moving into a field DeepMind has led for years. Its AlphaFold model won Demis Hassabis and John Jumper (who recently left DeepMind for Anthropic) a Nobel Prize in chemistry, and it is just one of many scientific projects the lab has running. Hassabis has made no secret of the ambition either—in interview after interview, he says DeepMind's goal is to use AI to crack some of the hardest problems in science and engineering. What Anthropic is doing differently from DeepMind is that instead of giving scientists specialised models, it offers them a tool to explore ideas, search articles and databases, which can also connect to specialised tools when needed. The company that owns coding now wants to own the laboratory too. Whether that works, and whether Claude Science leads to any real scientific breakthroughs, only time will tell.

If you enjoy this post, please click the ❤️ button and share it.

Share

🦾 More than a human

▶️ Our First Transdural Procedure | Neuralink (5:10)

The team at Neuralink explains in this video how they carried out their first brain implant surgery that goes straight through the dura—the tough protective layer around the brain—instead of cutting it open and removing part of it. This makes the operation safer, less invasive and faster. To do it, they widened the needle slightly so it could pierce the dura and added special imaging to spot blood vessels and measure the exact depth to the brain’s surface. Neuralink sees this as a step towards simpler, robot-assisted surgery that could help many more people.

🔮 Future visions

▶️ Is AI the Great Filter? | The Fermi Paradox (41:27)

The Great Filter is a hypothetical barrier in the evolution of life that is so hard to pass that it explains why we see no signs of advanced alien civilisations. In this video, Isaac Arthur explores whether artificial intelligence can be one of those filters. He argues that AI is a double-edged sword: while it poses serious risks like alignment failures or deadly arms races, it could also act as a stabiliser that helps civilisations solve complex problems and survive. Ultimately, he concludes that AI is likely not a definitive Great Filter, as it offers many different paths for development rather than a single, inevitable dead-end.

🧠 Artificial Intelligence

OpenAI proposes 5% stake to Trump administration to ease Washington pressure
According to the Financial Times, OpenAI has offered the US government a 5% stake in the company, worth about $42.6 billion. Sam Altman wants other big AI players like Anthropic, Google and Meta to do the same, giving the public a share in AI’s success through a government fund. The idea follows growing pressure in Washington over AI security risks and competition from cheaper Chinese models. It is unclear whether other companies will agree, and the government reportedly has not raised the idea with Anthropic. The Trump administration has already taken stakes in companies such as Intel and IBM.

Alibaba to ban employees from using Anthropic’s coding tool
Alibaba has reportedly banned staff from using Claude Code, citing worries that it could identify China-linked users by checking things like timezone and proxy data. Employees are being told to use Alibaba’s own tool, Qoder, instead. Anthropic says the flagged feature was a March experiment to stop resellers abusing accounts and to prevent rivals copying its models. The ban follows Anthropic’s recent claim that Alibaba illegally copied its Claude model capabilities.

South Korean tech giants commit over $550B to ease ‘RAMageddon’
South Korea’s tech giants, led by Samsung and SK Hynix, plan to spend over $900 billion on memory chips and AI data centres, including four huge new chip factories in the country’s south. The goal is to make South Korea a leading AI power and meet soaring demand during a global memory shortage. The scale is comparable to US tech giants' AI spending and comes amid record memory demand driven by the AI buildout.

Anthropic is discussing a new custom chip with Samsung
Anthropic is talking to Samsung about making its own AI chip, The Information reports, though it hasn’t yet decided what the chip will do or how powerful it will be. The company says it will still rely on a mix of chips from Google, Amazon, and Nvidia. The move follows OpenAI's unveiling of its own custom chip, developed with Broadcom.

Meta Is Planning a Cloud Business to Sell AI Computing Power
Meta is reportedly following in the footsteps of SpaceX by developing a cloud infrastructure business that would sell access to AI computing power and models. The plans, part of an internal initiative called Meta Compute, could involve renting out both “raw” computing capacity and access to hosted AI models (including Meta’s own Muse Spark), similar to AWS Bedrock. The move would provide a way to recoup the hundreds of billions Meta has committed to data centres and chips for its “superintelligence” ambitions, addressing investor concerns about the returns on that spending. Mark Zuckerberg has said such a business is “on the table” if Meta finds it has overbuilt.

Nvidia offers start-up customers chance to swap compute power for revenue share
Nvidia is offering fast-growing AI startups access to its scarce chips in return for a share of their future revenue. The first partners are two Australian firms, Sharon AI and Firmus Technologies, which together plan to run over 200,000 Nvidia GPUs. The deal reflects a wider trend of chipmakers and AI companies sharing revenue to ease money pressures in the sector.

Google limits Meta’s use of its Gemini AI models
Google has reportedly limited how much Meta can use its Gemini models because Meta wanted more computing power than Google could provide. Google gave Meta this news around March, which delayed some of Meta’s AI work, and a few other Google customers were affected too. Meta has now told staff to use AI more sparingly.

Nano Banana 2 Lite
Google DeepMind has launched Nano Banana 2 Lite (also called Gemini 3.1 Flash-Lite Image), its fastest and cheapest image-making and editing model yet. According to DeepMind, it creates images far quicker and at much lower cost than bigger models, while keeping most of the quality of the full Nano Banana 2.

Devin Fusion: Frontier Performance at 35% Lower Cost
Cognition has launched Devin Fusion, a coding tool that pairs an expensive, powerful AI model with a cheaper “sidekick” model. The main model plans the work and handles tricky decisions, then hands simple or repetitive jobs to the sidekick to save money. It can also switch models partway through a task at no extra cost. According to Cognition, this keeps top-tier quality while cutting costs by 35%, or 41% when using Anthropic’s Fable 5. Cognition argues that using one costly model for everything no longer makes sense, and mixing models is becoming the smarter approach.

Amazon seeks cheaper AI alternatives as Anthropic shifts to token-based pricing
Amazon is reportedly seeking cheaper alternatives to Anthropic’s Claude models after a new deal set to raise its AI costs, and is exploring OpenAI instead, despite relying heavily on Claude across several products. The rift is widening on both sides: Amazon has committed $50 billion to OpenAI, while Anthropic has pledged $200 billion to Google Cloud, easing its reliance on Amazon. Tensions rose last month when the US government ordered Anthropic to shut down two models after a security report from Amazon.

OpenAI is teasing new hardware… for Codex
OpenAI is teasing its first hardware products, and it looks like it is a physical device for Codex, its AI coding tool. Launching on 15 July with keyboard maker Work Louder, the square, button-covered gadget resembles a macro pad that lets users map custom shortcuts. It appears aimed at developers who want quicker control over Codex. This is separate from the secretive AI device OpenAI is building with Jony Ive.

Ford’s AI Hiccups Lead Carmaker to Rehire ‘Gray Beard’ Engineers
Ford found the hard way that the machines can’t replace experience. To address the quality problem the carmaker is facing, the company rehired 350 experienced “gray beard” engineers to retrain the AI and automated systems that had been failing on their own. These veterans now lead troubleshooting meetings and catch faults before parts reach the factory floor. The result: Ford jumped from tenth to the top mainstream brand in JD Power's latest quality survey, behind only Porsche and Genesis, while cutting warranty and recall costs.

SpaceX Showed Investors Prototype of Elon Musk’s New AI Device
According to the Wall Street Journal, SpaceX has shown investors an early prototype of a slim, phone-like device that would run its own software and use xAI’s artificial intelligence. The project is at an early stage and may never be built—Musk denied making a phone as recently as February. The device fits his long-held “everything app” idea, bundling many services into one platform, and would cut his reliance on Apple and Google.

Taiwan raids Super Micro’s office as Nvidia chip smuggling investigation widens
Taiwan's Keelung District Prosecutors Office has raided Super Micro's local office, two related companies, and six homes as part of its first criminal probe into the smuggling of Nvidia chips to China. The investigation began in May, when three people, including a Super Micro co-founder, were detained over forged export documents. Taiwan does not yet treat these chip exports as a crime, so prosecutors are using other charges for now, though that may change. The case links to a wider scheme that US prosecutors value at about 2.5 billion dollars.

Samsung, SK Hynix, Micron face antitrust class action lawsuit
Samsung, SK Hynix, and Micron are being sued in a US class action that accuses the three chipmakers of working together to keep memory chips scarce and push prices up by around 700% over four years. The 17 plaintiffs say the companies used the shift towards AI-focused memory as an excuse to cut output of older chips, and they point to Apple’s recent price rises as proof of the impact. They want the court to stop the alleged squeeze and are seeking triple damages, citing the companies’ past price-fixing convictions as evidence of a pattern.

AI Is Designing Radio Chips That Humans Couldn’t Even Imagine
This article explores the dark art of RF chip design that can take experts years to master and how AI could transform it. A Princeton team has trained AI to design these chips from scratch, and though the results look strange, they often outperform human designs. The author argues this could break barriers not just in chip design but across many engineering fields, provided human oversight catches AI errors and private datasets are opened up to advance the work.

▶️ The GPU Power-Performance Curve Most Clusters Ignore (10:02)

Jordan Nanos from SemiAnalysis sits down with Keval Shah, AI research lead at Pebble, to discuss GPU power optimisation, tokens per watt, and grid-responsive data centres. Keval walks through why the power-to-performance curve in GPU clusters is non-linear, where cranking up wattage stops producing proportional tokens, why memory-bound inference leaves decode stages consuming power while prefill waits, and how Pebble aims to address these problems.

From Brain Waves to Words: Brain2Qwerty Offers a New Path to Communication Without Surgery
Researchers at Meta have unveiled Brain2Qwerty v2, a system that decodes typed sentences directly from brain activity without any surgery. Using a magnetoencephalography helmet and AI trained on nine volunteers, it correctly decodes 61% of words on average (78% for the best participant)—a major leap from the 8% of earlier non-invasive methods, and close to what surgical implants achieve. Meta is releasing the code and a dataset openly, hoping to eventually help people who have lost the ability to communicate.

🤖 Robotics

Apptronik unveils Apollo 2 and a flagship data collection and training facility
Apptronik has launched Apollo 2, its updated humanoid robot. Apollo 2 comes in two variants—a two-legged version and a wheeled-base version—and its main purpose is to collect real-world data from tasks in warehouses, factories and shops. That data feeds into the AI models Apptronik is developing with Google DeepMind and helps create Apptronik’s next robots. Alongside Apollo 2, the company also announced the opening of an expanded training facility called Robot Park in Austin, Texas.

Weave Robotics launches Isaac 1, a $7,999 home robot with Fall 2026 deliveries

Weave Robotics has launched Isaac 1, a home robot that does chores on its own, with a human operator stepping in when needed. According to the company, the robot can find, fold and put away laundry, make beds, and tidy toys and clutter. It stands about human height and folds away when not in use. Pre-orders are open now with a refundable $250 deposit, costing $449 a month or $7,999 to buy. First deliveries start in California in autumn 2026.

Ministers likely to support law change to allow delivery robots on England’s paths
England looks set to change the law so that small delivery robots can legally use pavements, opening the door for rolling out over 10,000 of those robots. These low-speed robots mostly carry groceries and takeaways but currently operate in a legal grey area. Campaigners at the charity Living Streets warn they will crowd already-busy pavements and create dangers for older, disabled, and blind pedestrians. The government says it will put pedestrian safety first when updating the rules after a public consultation.

▶️ F.03 at BMW (0:59)

In this video, Figure demonstrates how its F.03 humanoid robots are being used at a BMW factory. The video is more promotional than demonstrative of the actual work being done by the robots, which would be nice to see. Perhaps this could be shown in the form of a livestream, which has the potential to go viral, as the previous one did.

Waymo and Uber quietly part ways in Phoenix
Waymo has ended its three-year partnership with Uber in Phoenix, pulling its robotaxis from Uber’s app and adding them back into its own fleet. Both companies called it a small test that helped them grow bigger services in Austin and Atlanta. Uber plans a new self-driving partner in Phoenix but hasn’t said who.

AGIBOT produces 15,000th robot, marking a milestone in embodied AI deployment
Chinese robotics company AGIBOT announced that it has built its 15,000th robot, all while speeding up the production pace, going from 1,000 to 5,000 units took about a year, while the next 5,000 took just three months. The landmark machine was a G2, a wheeled robot with a humanoid torso and arms made for factory jobs like quality checks. AGIBOT, which led the world in humanoid robot shipments in 2025, argues that success now depends on making and deploying robots at scale, not one-off demos.

Swarm robots inspired by bees and ants could transform the future of mining
Australian researchers have built small robots that work together like bees and ants, without needing a central controller. In tests copying a mine, a bee-inspired method—exploring and mapping before collecting resources—worked best, cutting travel distance by up to 80%, halving energy use, and finishing tasks up to 60% faster. Because the robots make their own decisions, the system keeps running even if some stop working. The team says the approach could make mining safer and may one day help with mining in space.

🧬 Biotechnology

The first early human eggs from stem cells
Conception, a biotech company pursuing "in vitro gametogenesis", says it has made the first early human egg cells from stem cells. Starting with a simple blood draw, it turned blood cells into stem cells and grew them into tiny lab-made ovaries containing early eggs. One day, this could let families create as many healthy eggs as they need, without the injections and surgery that IVF requires today. But the eggs are not yet fully grown, and the company must still prove the process is safe before it can be used in people.

A Chemically Defined Synthetic Cell Capable of Growth and Replication
Scientists have built tiny synthetic “SpudCells” from non-living chemicals that can feed, grow, copy their lab-made DNA and split into new cells—the first artificial cells to complete the full life cycle. They are fragile, depend entirely on a nutrient-rich liquid, and die after a few generations, but experts call it a major breakthrough that could reveal how life first emerged from chemistry. The team hopes to develop the technology to one day produce drugs, foods and fuels.

Insilico Medicine And SK Biopharmaceuticals Achieved AI-powered Drug Discovery Collaboration Worth Up to 2.5 Billion for Neuroimmune Disorders
Insilico Medicine and SK Biopharmaceuticals have teamed up to use AI to design new drugs for hard-to-treat brain and nervous system disorders. Insilico’s software will find and refine drug candidates, while SK Biopharmaceuticals will run the later testing and bring any successful drugs to market. The deal could be worth over $2.5 billion in total, with $18 million paid upfront and early on. Insilico says its AI can find promising drug candidates in 12 to 18 months, far faster than the usual few years. Both companies want this to be a repeatable model they can use for many future drugs, not just a one-off.

Ultrasound imaging of the brain
Aleph Neuro has built a way to read the brain using ultrasound instead of surgery or blurry scalp scans. It tracks blood flow, since active neurons draw more blood, and uses tiny injected bubbles to sharpen the signal. The result is what the team believes is the most detailed image yet of a living brain’s blood vessels captured through an intact skull. They are open-sourcing the tools and data, hoping to help detect early signs of stroke, Alzheimer’s, and brain injury, with a longer-term goal of achieving the same clarity without injections.

💡Tangents

Rocket Lab to Acquire Iridium
Rocket Lab is buying satellite company Iridium in a cash-and-stock deal worth about $8 billion, giving it control of 66 satellites and valuable spectrum rights to take on SpaceX’s Starlink. The purchase adds an established, cash-generating business to Rocket Lab’s existing rocket and satellite-building work. Iridium’s current satellites will last another decade, giving Rocket Lab time to build a newer network. Both companies say controlling launch, manufacturing, and operations under one roof is the key to competing in space.

Antares Achieves Criticality of Mark-0 Reactor
Antares, a US nuclear startup founded in 2023, is the first private company to switch on an advanced reactor under the Department of Energy’s pilot programme. Its small Mark-0 reactor reached criticality at Idaho National Laboratory in under a year, and is the first privately developed non-light-water reactor in America to do so in over 40 years. The test also creates a faster approval route for future reactors. Antares aims to produce electricity by 2027 and power military bases by 2028.

Intel’s Chip Business Shows Signs of Life After Years of Struggle
Not so long ago, Intel was described as Silicon Valley’s fallen icon. The sales were plummeting, its market share had been eaten up by AMD, and the company had missed the enormously profitable AI boom. But now it seems Intel is turning around its fortune. This article explores how, under Lip-Bu Tan’s leadership and with the US government buying a 10% stake, Intel’s value has more than tripled to $650 billion and it has won big customers like Nvidia, Apple and Google. Its recovery matters to America’s plan to make more chips at home and rely less on Taiwan. However, Intel is still fragile—it is losing money, has cut thousands of jobs, and Tan expects the full turnaround to take at least five years.

▶️ Building a Robot that Hunts AI Glasses (24:56)

brenpoly shows in this video how he built a real-life Odradek robotic arm from Death Stranding that, instead of BTs, hunts AI glasses. It is a really cool project that combines local AI with robotics, with a solid amount of 3D printing and electronics.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

This week, we take a closer look at SpaceX and the kind of game Elon Musk is playing with it and his other companies.

Over in AI, OpenAI released the GPT-5.6 family of models, but, similarly to what happened with Fable 5, the US government imposed restrictions on who can access the new models. Elsewhere in AI, Sakana released its first models, Fugu, and immediately shot to the top with an interesting architecture. Meanwhile, OpenAI unveiled its first AI chip, Anthropic accused Alibaba of “illicitly” accessing its AI models, Zhipu rode high after releasing GLM-5.2, Allbirds continued its pivot to AI, and Google partnered with A24.

In robotics, Agility Robotics goes public, Tesla faces a trial following a fatal crash, JD.com says robots will replace its 700,000 couriers, and Mobileye plans to enter the increasingly crowded US robotaxi market.

Apart from that, this week’s issue of Sync also features BioVault by Colossal Biosciences and the US government, a new liquid neural network, Apple raising prices, an exoskeleton with a robotic arm to help factory workers, and more!

Enjoy!

What is going on with SpaceX

Inside Musk's high-risk bet to fold rockets, AI, and robotics into one company

For a few days in June, Elon Musk was the richest person on the planet, and SpaceX was worth more than Amazon and Microsoft. The company had just pulled off the largest IPO in history, priced at $135 on 11 June, raising around $86 billion and valuing a rocket company at roughly $1.77 trillion. It closed its first day up 19% and kept climbing to $225.64 on 16 June, a peak that put SpaceX near $2.6 trillion and made Musk, on paper, the world’s first trillionaire.

Then it reversed. Over three sessions, the stock handed most of it back, including a 16.4% fall on 22 June that erased roughly $400 billion in hours, the second-largest single-day wipeout on record, behind only Nvidia’s near-$600 billion loss eighteen months ago, when DeepSeek spooked the market. By week’s end, it had steadied near $153, still above its listing price.

But SpaceX is no longer only a rocket company, even though it still puts more into orbit than anyone else. It is also a satellite-internet provider, rents out computing power, and is the centre of gravity for Musk’s attempt to build an empire from the technologies he thinks will define the century, while burning billions of dollars. My reading, which the rest of this piece tries to justify, is that this is a high-risk, high-reward bet that for now needs two things it does not have: more time and more money.

A company with $100 billion borrows more

So why did SpaceX’s stock fall? SpaceX sold only a small slice of itself to the public (about 4.2% of its shares), and when that few are trading, the price swings sharply in both directions. After a run-up like that, some investors were always going to cash out.

But the June 22nd drop had a specific trigger. That day, SpaceX said it would borrow at least $20 billion by selling bonds, to pay off a short-term loan it had taken in March to buy xAI. Paying off one loan with another is ordinary housekeeping. What unsettled investors was the timing: SpaceX had just told lenders it was sitting on $100.8 billion in cash—more than Microsoft—and was borrowing anyway. In the end, it raised about $25 billion. Which raises the obvious question: why does a company with that much cash need to keep raising more?

SpaceX has lost money almost every year since 2002—about $41.3 billion in total. In 2025, it brought in $18.7 billion but still lost roughly $4.9 billion, and the figure that matters most is that it spent about $14 billion more than it took in. That gap had more than doubled in a year, and the new AI business was doing most of the spending. One analyst at KeyBanc expects this year's gap to be closer to $28 billion, and SpaceX has disclosed plans to spend around $350 billion building everything out by 2030—most of it on AI. Moody's still judged its debt safe enough to invest in, but warned the rating was held back by "elevated execution and financial risks"—in plain terms, a risky plan and a lot of money going out the door.

SpaceX is raising every dollar it can, in shares and in debt, not because it is short of cash today, but because of the scale of what it plans to build next.

More than a rocket company

SpaceX is still, at heart, a rocket company. Falcon 9 and the idea of reusable rockets revolutionised the space industry and improved the economics of spaceflight. In 2025, SpaceX carried 2,213 tonnes to orbit, more than 80% of everything humanity launched that year. The runner-up, CASC, China’s state launch group, launched roughly an eighth of what SpaceX did. Falcon and Dragon are the reliable, cash-generating workhorses behind some $22 billion of government business, with no real competition.

Then there is Starlink, which generated $11.4 billion in 2025, about 61% of the company's $18.7 billion revenue, and it is the only segment that makes money. Starlink is the engine that quietly subsidises everything else, though revenue per user has slid from about $99 a month in 2023 to $66 in early 2026, even as subscribers doubled.

But since the beginning of 2026, SpaceX has started to become more than just a rocket company. In February 2026, SpaceX absorbed xAI, Musk’s AI company, in an all-stock deal. That merger also brought in X, the former Twitter, under SpaceX. AI became SpaceX’s third reporting segment, and immediately its largest cash sink.

Grok, xAI’s AI model, never seriously challenged OpenAI or Anthropic. Downloads collapsed by about 60% by April, and paid conversion is negligible, around 0.17% against 6% and more for ChatGPT. Musk himself conceded the company “was not built right.” Grok now lives mostly inside X, better known for the controversies it generates than the subscriptions it brings in.

To run Grok, xAI went on building two enormous data centres—Colossus 1 in Memphis, with more than 220,000 Nvidia GPUs, and Colossus 2, which aims to house 550,000 Nvidia GPUs. But with Grok failing to become a key player on the AI stage and unable to use all of that computing power, SpaceX decided to essentially become a neocloud. The company now rents the world’s most sought-after compute out, and it has already landed some big names—Anthropic pays around $1.25 billion a month for Colossus 1; Google has agreed to roughly $920 million a month from October; and recently, SpaceX signed a $150 million per month deal with Reflection AI. Together, that is about $26 billion a year, more than SpaceX’s entire 2025 revenue. And where Grok fell short, SpaceX simply bought a replacement. In June, it agreed to acquire Cursor, the popular AI coding tool, for $60 billion—all in stock, and the largest acquisition of a venture-backed startup on record.

Musk’s data-centre ambitions go beyond Earth, figuratively and literally, as he dreams of orbital data centres. He has called space-based AI a “no-brainer,” and set out to build factories that would put a gigawatt of orbital computing into space each year by late 2027. Like many of Musk’s ideas, orbital data centres are technically possible, but currently economically unviable. Even SpaceX’s own prospectus concedes the plans “may not achieve commercial viability.” And to make the chips for all of it—the satellites, the cars, the robots—SpaceX, Tesla and Intel are jointly building Terafab, a semiconductor plant in Texas that could cost up to $119 billion.

Underneath all of it sits Starship, the vehicle meant to launch the next-generation Starlink satellites, the orbital data centres, and NASA’s astronauts back to the Moon. It is late and keeps exploding. It managed only five flights in 2025 against a target of twenty-five, a string of ground-test failures, an FAA grounding in May, and roughly two years of slippage that NASA’s own inspector general says threatens the Artemis timeline.

None of this is cheap. Each of these bets—the data centres, Cursor, the chip plant, Starship—runs to tens or hundreds of billions of dollars, and almost none of them make money yet.

Tesla is next

You'll notice one company has been absent from all of this: Tesla. And it is not doing well. Musk’s EV company delivered 1.64 million vehicles in 2025, down 8.6%, a second straight decline that ceded the global EV crown to China’s BYD. As the car business faces problems, the company is pivoting into robotaxis and humanoid robots, both of which are still unproven and are currently in an R&D phase.

Then there is Tesla’s brand, which has turned politically toxic, and the cost is measurable. A Yale study estimated that Musk's political turn cost Tesla between 1 and 1.26 million US sales and $51–62 billion in revenue, from late 2022 to 2025: the environmentally-minded, left-leaning buyers who were Tesla's core market walked away over his role in Trump's second term, while the right was never going to buy an electric car anyway.

So Musk has a shrinking carmaker on his hands, inside a portfolio of companies he owns outright. If you have followed him for any length of time, you can probably guess what tends to happen to a Musk company in that position. We have seen it twice already. In 2016, Tesla bought SolarCity, a near-insolvent solar company founded by Musk’s cousins and chaired by Musk himself, in an all-stock deal critics called a bailout. In March 2025, xAI absorbed X at a $33 billion tag, below the $44 billion Musk had paid for it in 2022. And in February 2026, with Grok failing, xAI was folded into the healthier SpaceX, valuing the combination at about $1.25 trillion, just months away from SpaceX’s IPO. Tellingly, Musk’s announcement never mentioned Grok.

I’d like you to notice a pattern here. A weaker company folded into a stronger one, always in stock, always with Musk on both sides, always justified by “synergies.” Which raises a fair question about the xAI–SpaceX deal itself—how much was about consolidating Musk’s companies, and how much about lifting SpaceX’s valuation just before it went public? Probably both. But bundling a failing chatbot into the most admired private company in the world, weeks before selling shares and without naming it, leans one way.

No Tesla–SpaceX merger has been announced, but it should not be ruled out. The New York Times and Bloomberg’s Liam Denning both entertained that idea. If it were to happen, it would most likely follow the same pattern as the previous mergers—an all-stock deal with SpaceX as acquirer. On IPO day, SpaceX president Gwynne Shotwell said that a merger “might make Elon’s life a little easier,” before adding that for now she was focused on “keeping the lights on.”

What makes it personal for Musk is the pay package Tesla's shareholders approved in November 2025, the largest in corporate history, worth up to $1 trillion in stock, paid out in twelve tranches as Tesla hits a ladder of targets. The hardest of those are operational, such as a million robotaxis on the road or a million Optimus robots delivered, which Tesla is nowhere near doing. But as Electrek argues, a merger would change the rules. The package contains a "change of control" clause that, in a takeover, sets the operational targets aside and leaves only one test standing: the share price. And a merger is exactly the sort of news that sends a share price up. So a deal could let Musk clear the final bar precisely by announcing the deal—collecting a slice of the largest pay package in history without Tesla building a single one of the things it was meant to reward. It is not an automatic payout, but it is a way for Musk to solve several problems at once.

A nascent tech empire that needs time and money

The combined company would span rockets, electric cars, AI, and robotics, with a social-media platform bolted on for good measure. It would almost certainly be the most valuable company in the world, with Musk in control of all of it.

I think that is the game Musk is playing right now. He is trying to build an empire from the technologies he believes will define the century—space, AI, and robotics—and to finance each against the others inside a structure he controls.

But so far, that nascent tech empire faces serious challenges. Its building blocks are struggling. The AI build-out alone runs to tens of billions a year. Starship keeps exploding. Grok failed. Tesla is shrinking. And on top of that, Musk’s companies are pouring money into robotaxis, humanoid robots and orbital data centres, all of which are unproven.

It is a high-risk, high-reward game, and the IPO gave Musk room to keep playing it a while longer. If he wins, he builds the kind of mega-corporation that so far exists only in scifi/cyberpunk stories. If he loses, the damage no longer stops with him. Somewhere along the way, Musk’s bet quietly became everyone’s.

If you enjoy this post, please click the ❤️ button and share it.

Share

🦾 More than a human

Exoskeleton and robotic arm reduce factory lifting strain by up to 65%

Researchers at the Technical University of Munich have built WearaCob, pairing a wearable exoskeleton with a robotic arm to help factory workers. The exoskeleton takes over part of the lifting and cuts muscle effort by up to 65%, with the robot weighing each object and telling it how much help to give, even for uneven loads. The robot moves safely around people and can be taught new tasks just by guiding its arm, with no coding needed. The result is closer, safer teamwork that could reduce strain and injury on the factory floor.

🧠 Artificial Intelligence

Previewing GPT‑5.6 Sol: a next-generation model
OpenAI has launched its new GPT-5.6 series: Sol, the flagship model; Terra, a balanced model for everyday work; and Luna, a fast and affordable option. However, unless you are one of a small group of approved partners, you cannot use the new models—at the Trump administration’s request, all three have been restricted, much like what happened to Anthropic’s Fable 5, which was pulled entirely after a government order. OpenAI made clear it is unhappy with the arrangement, calling it a short-term step before wider access in the coming weeks. On capability, OpenAI claims Sol is its strongest model yet, with improvements in coding, biology, and cybersecurity. To address safety concerns, it says it has built a heavily hardened safety system directly into the model, designed to favour defensive security work over hacking.

OpenAI and Broadcom unveil LLM-optimized inference chip
OpenAI has revealed Jalapeño, its first custom AI chip built with Broadcom, to run its models more cheaply and efficiently. The chip is made just for inference rather than training. Early tests suggest it uses far less power than current alternatives. Like Google and Amazon before it, OpenAI wants to rely less on Nvidia and control more of its own technology, from chips to products, so it can make AI faster, cheaper, and more reliable. Deployment is set to begin by the end of 2026 and expand over several chip generations.

Sakana Fugu
Sakana AI has launched Sakana Fugu, which isn't one big model but a system that picks and combines a pool of leading AI models behind a single API to handle hard coding and research tasks. It comes in two flavours—the base Fugu and Fugu Ultra. According to benchmark results from Sakana, both models are comparable to the world’s best models, with Fugu Ultra matching, in some cases, the performance of Antropic’s Claude Fable 5. Sakana Fugu is available on a pay-as-you-go basis or via a monthly subscription, though it's not yet available in the EU.

Introducing Claude Tag
Anthropic has launched Claude Tag, a tool that lets teams add Claude to their Slack channels and assign it tasks by tagging Claude. Unlike a private chat, one shared Claude works for the whole channel, remembers context over time, and can work on its own. The tool is in beta for Enterprise and Team customers.

OpenAI Leans Toward Waiting Until Next Year for I.P.O.
OpenAI may delay its stock market listing until 2027, according to internal sources quoted by The New York Times. Sam Altman wanted a quick listing at a $1 trillion value, but a weak debut from SpaceX and nervous markets have made advisers warn that investors might not buy in. OpenAI is still losing money and spending heavily, even as it cuts side projects and pushes its Codex coding tool to compete with Anthropic.

Anthropic Accuses Alibaba of ‘Illicitly’ Accessing AI Models
Anthropic has accused Alibaba of secretly using its Claude AI model without permission, claiming that operators tied to Alibaba’s Qwen lab used nearly 25,000 fake accounts to run 28.8 million exchanges between April and June. In a letter to US lawmakers, Anthropic said the goal was to copy Claude’s most valuable skills cheaply, and it urged the government to crack down on the practice. This is not the first such accusation from Anthropic. In February 2026, the company levelled similar claims against DeepSeek, Moonshot, and MiniMax.

MGX raises a $50bn AI fund and is already spending it
Abu Dhabi’s MGX has raised nearly $50bn for one of the biggest funds ever built solely to back artificial intelligence. Unusually, the emirate raised this money from outside investors rather than spending its own oil wealth, making MGX act more like a global asset manager than a typical Gulf state fund. It already holds stakes in OpenAI, Anthropic and xAI, and is now eyeing a Singapore data-centre operator as its first deal in Asia.

Zhipu weighs multibillion-dollar raise after 2,000% surge
Chinese AI firm Zhipu has seen its shares jump about 2,000% this year, valuing it above $128 billion and prompting plans for a huge new share sale. The rally took off after Anthropic pulled its top models worldwide under US export controls, leaving customers nervous about relying on American AI. Zhipu quickly launched GLM-5.2, a free, open model that became the first Chinese system to rank among the world’s top three. Supporters predict booming revenue and profit by 2028, but critics warn the company is losing money heavily and will need constant fundraising to survive.

U.S. Presses Meta to Agree to A.I. Reviews as Security Concerns Rise
The Trump administration wants Meta to let the government review its AI models for safety and security risks, The New York Times reports. Every other big US AI company has already agreed, leaving Meta as the last holdout. This is part of a wider push for tighter government oversight, backed by a recent order asking companies to share models before release.

Cerebras doubled its revenue. The stock fell anyway.
Cerebras almost doubled its revenue and predicted strong yearly sales, but its shares still fell about 10%. The problem was profit margins, which are being squeezed because the company can’t find enough data-centre space and power to run its chips. Its boss called it a “grand irony” that buildings, not technology, are now the main obstacle.

▶️ The data black hole at the center of AI (11:56)

In this video, Dwarkesh Patel argues that AI progress is driven mainly by data, not smarter learning. Models improve because they’re fed more and better data, with reinforcement learning acting as a way to generate high-quality synthetic examples. His key point is that today’s models are far less efficient learners than humans, needing vastly more information to gain each new skill. He rebuts common objections—evolution, hidden sensory data, and scaling—and notes this explains why open models catch up to the frontier within months.

Allbirds continues AI pivot with name change and CEO hire, sending stock soaring
Allbirds, the company you might have known as the maker of sustainable shoes, continues its hard pivot into AI infrastructure by rebranding to Smartbird. It has also brought in Nadia Carlsten, a former AWS quantum computing boss, as CEO. Investors are apparently happy with the change, as shares went up by 39%. The company had already sold its shoe business for $39 million and closed its US stores.

ByteDance unveils Seedance 2.5, a 30-second native 4K AI video model that accepts 50 reference inputs
ByteDance has launched Seedance 2.5, an AI tool that creates 30-second 4K videos and can take up to 50 reference inputs like images, audio, and 3D models—far more than rivals such as Google’s Veo. It is built for professional use, with better sound syncing and sharper results, and is set for public release in early July.

Alibaba’s AI video model rises to No. 2 in global rankings, as OpenAI’s Sora and ByteDance’s Seedance fall away
Alibaba has released HappyHorse 1.1, an AI video-making tool launching with a 40% discount and improved features like consistent characters and accurate lip sync. It reportedly outperforms Google, its main competitor, in independent tests. Backing it all is Alibaba’s huge $52.7 billion infrastructure spending, which gives the product the scale and local hosting needed to meet Europe’s strict data rules.

Notion Mail shuts down amid agent takeover
Notion is closing its email app in favour of its AI agents. The company says more than half of users already let agents handle their email without ever opening their inbox. Notion says that existing emails in the inbox will stay intact, but users must export drafts, scheduled emails, and saved snippets to keep them.

Patch the Planet: a Daybreak initiative to support open source maintainers
OpenAI has launched Patch the Planet, a security initiative built with Trail of Bits that uses its most cyber-capable AI models, such as GPT-5.5 Cyber, to find vulnerabilities in critical open-source software, paired with human engineers who verify every finding and help maintainers write patches. The aim is to ease the burden on overwhelmed maintainers, since AI is now very good at finding flaws, but fixing them remains the bottleneck. Early results include hundreds of issues found and dozens of patches merged across 19 projects like cURL, Python, and Go, plus serious bugs uncovered in Linux, OpenBSD, and major browsers. It closely mirrors Anthropic’s Project Glasswing, but where Glasswing has been criticised for finding many bugs that go unpatched, Patch the Planet targets exactly that gap by prioritising human-verified fixes.

Tesla plans to sell modular AI data center hardware called ‘Megapod’
Tesla has applied to trademark “Megapod,” a complete AI data centre unit that bundles servers, power, and cooling into a single package. This comes just after Tesla scrapped its Dojo supercomputer and saw repeated delays to its own AI chips, all of which point to a weak track record in building compute hardware.

▶️ What does a GPU cluster really cost? (13:40)

SemiAnalysis and High Yield break down the real cost of running and using a GPU cluster. The analysis covers the entirety of a GPU cluster—not just GPUs, but also networking, storage and other things that usually don’t make the headlines and how they affect the cost and effectiveness of a GPU cluster.

Google Investing in ‘Backrooms’ Studio A24
Google is investing about $75 million in the independent film studio A24 to build new tools for making and releasing movies. It is Google’s first stake in a film studio, and the deal is unusual because Hollywood and AI firms have mostly clashed over copyright and creativity. The companies say their tools will help filmmakers keep creative control rather than just cut costs, and A24 is already working on AI-generated storyboards. The move is risky for A24, whose young, AI-sceptical fans may dislike the partnership.

Mistral: Introducing OCR 4
Mistral has released OCR 4, a small AI model for reading documents. Unlike older tools that just turn a page into plain text, OCR 4 also marks where each element sits, labels what it is (such as a title, table or signature), and rates how confident it is in each part. Mistral says OCR 4 beats other such models, and it can be run on a company’s own servers to keep sensitive data private.

LFM2.5-230M: Built to Run Anywhere
Liquid AI presents LFM2.5-230M, its latest fast, lightweight foundation for developers to fine-tune and deploy in agentic workflows. Unlike other models, LFM2.5-230M is not built using transformer architecture. Instead, it is using liquid neural networks (Ramin Hasani, the company’s CEO and co-founder, explains how they work in depth in this lecture). Despite its small size (only 230 million parameters), LFM2.5-230M beats or matches larger models in its category, including Google’s Gemma 3 1B. LFM2.5-230M is available on Hugging Face.

🤖 Robotics

Humanoid maker Agility Robotics to go public through SPAC merger
Agility, the company behind the Digit humanoid robot, is going public through a SPAC merger in a deal valuing it at $2.5 billion and raising more than $620 million to scale production. The company positions itself as the only US publicly listed pure-play humanoid company with active commercial deployments, with Digit already working at customers such as Schaeffler, Toyota Canada and Mercado Libre, and has already attracted over $300 million in multi-year orders.

A Fatal Tesla Crash in Texas Sets Up a Legal Showdown
A 76-year-old Texas woman was killed when a Tesla Model 3 crashed into her home at over 70 mph (over 112 km/h). Now her family is suing both the driver and Tesla, claiming its Full Self-Driving feature was dangerously designed. Tesla denies this, saying the driver had pressed the accelerator fully, overriding the system. The case is similar to a 2024 Florida one, where a jury found Tesla partly to blame for a fatal crash and ordered it to pay over $240 million.

▶️ Unitree R1 | Price from $4,900, Ready Stock (0:30)

Unitree began deliveries of R1, its latest and most advanced humanoid robot. And to celebrate that, Unitree reminded everyone what kind of acrobatics the robot is capable of.

Mobileye is entering the US robotaxi market with standalone service
Mobileye plans to launch its own robotaxi service in an unnamed US city in 2027. It will start with about 100 cars early next year and hopes to grow to around 17,000 within five years. This is a big shift, as Mobileye usually supplies its technology to other companies rather than running its own service. The company says it will still keep working with partners like Volkswagen, Lyft and Porsche while building this new business.

Nvidia Seeks to Make Humanoid AI Robots Safer Around Humans
Nvidia sees humanoid robots as the next big market for AI, worth $200 billion by 2035, and wants to speed up their arrival. To do so, it has built Halos, a system adapted from self-driving car technology and running on its IGX Thor hardware, which helps robots understand their surroundings so they can work safely next to people, even touching them when needed, instead of freezing whenever a person comes close. Nvidia has also opened a lab to help robot makers pass safety tests before seeking approval.

Project Fetch: Phase two
Anthropic revisits Project Fetch, in which the company tested how much Claude can help people program a robot dog. This time it checked whether the newer model, Claude Opus 4.7, could do the jobs alone rather than just helping people. Working on its own, the model was roughly 18 to 38 times faster than the earlier human teams at tasks like connecting to the robot’s sensors and spotting a beach ball, though it still couldn’t reliably nudge the ball home. Anthropic sees this as a sign we may be entering an early stage of “physical AI,” where models can pick up and use everyday hardware much as they learned to use software.

JD.com says robots will replace its 700,000 couriers
JD.com’s chair, Richard Liu, says robots will eventually replace the company’s 700,000 delivery couriers. To cushion the loss, JD has started a programme to retrain couriers for new jobs like robot maintenance, and Liu has promised not to fire any front-line workers. But the promise is hard to keep: the new roles will be far fewer than today’s courier jobs, and it’s unclear if retraining can keep pace.

Sony discontinues Japan sales of robot puppy ‘aibo’
Sony is stopping sales of its robot dog Aibo in Japan once the current stock runs out, ending an eight-year run for the interactive pet that can learn its own personality and do tricks. The $3,000-plus robodog, first launched in 1999 and revived in 2018, will still be sold in the US, and Sony says it will keep supporting existing owners. The company insists the Aibo business will continue, but won’t confirm a new model. Fans in Japan reacted with sadness to the news.

HaloBraid raises $7M to build the first robotic braiding assistant for hair salons
Meet HaloBraid, a robotics startup working to bring an automated braiding assistant to hair salons this year. The device doesn’t replace stylists—a braider starts each braid by hand, then the machine finishes it about five times faster, easing the strain that leaves many braiders with carpal tunnel and arthritis. The startup has recently raised $7 million, led by Reddit co-founder Alexis Ohanian’s venture firm.

▶️ Warehouse Robots Don’t Need Legs (44:55)

In this conversation, Locus Robotics CEO Rick Faulk shares the lessons he and the company have learned about warehouse automation. He argues that a robot built to do one job really well beats the popular humanoid robots, which he sees as too expensive and not yet practical for warehouses. He stresses that the company's success comes from starting with the customer's real problem rather than building clever technology in search of a use. Faulk predicts that within a decade, almost every warehouse will need to automate to stay competitive.

What Amazon’s Astro Taught Me About Giving Robots a Soul
In this article, Mike Forst shares his experience designing the sound and character for Amazon’s Astro home robot. His main point is that any robot moving through your home will be seen as a character, whether designers plan for it or not, so it pays to shape that character on purpose. Forst urges product makers to define a robot’s character before its features, and to design it to adapt and grow with the people it lives with. It is an insightful story on robot design and how to make a robot feel “alive".

Tiny robots deliver stem cells to repair spinal cords
ETH Zurich researchers have built tiny robots that may help repair spinal cord injuries, which rarely heal on their own. Each robot pairs stem cells with magnetic particles, so doctors can steer the cells to the exact injury and use magnets to stimulate them into new nerve cells. In tests, injured zebrafish swam normally within three days, and mice with fully severed cords regained much of their movement after 28 days, with no harmful side effects. The robots then dissolve away once their work is done. Human trials, however, are still years off.

🧬 Biotechnology

Colossal and the US Government Are Creating an Endangered Species ‘BioVault’
The US government has teamed up with Colossal Biosciences to store frozen cells, tissue, and DNA from over 2,300 endangered species. The government will own the samples, while Colossal pays the costs and handles the genetic sequencing. The deal comes as the Trump administration is rolling back endangered species protections, including allowing more offshore drilling.

NVIDIA Announces BioNeMo Agent Toolkit — Tools for Agents to Accelerate Scientific Discovery
Nvidia has released BioNeMo Agent Toolkit, which turns AI assistants into specialist science helpers for drug discovery and biology research. It gives these AI agents the tools to do real scientific work, such as designing proteins, screening drug candidates and analysing genetic data to help make research faster and cheaper. Nvidia says this could speed up discoveries across biology, chemistry and medicine by letting AI and scientists work together in a continuous loop. BioNeMo Agent Toolkit is open source and available on GitHub.

💡Tangents

Apple raises prices of MacBooks, iPads as memory costs skyrocket
Apple has raised prices on its iPads, MacBooks, HomePods and Apple TV because soaring memory and storage costs, driven by the AI data centre boom rather than a shortage, have made these chips far more expensive. The iPhone is safe for now, but analysts expect its prices to climb too. If you want to know exactly how the prices have changed, 9to5Mac has a great breakdown of them. ZONEofTECH also analyses the new prices, and both come to the same conclusion—the price hikes are massive, especially for higher-end MacBook Pros and Mac Studios.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now

This week, we explore OpenAI’s leaked audited financial statements and what they reveal about the company’s financial position.

Elsewhere in AI, SpaceX has completed its acquisition of Cursor in a deal worth $60 billion. Meanwhile, DeepSeek raised $7.4 billion at a valuation of more than $50 billion, GLM-5.2 has become the leading open model, and several prominent figures in AI have switched companies. We also cover Midjourney’s pivot into medical imaging, Poland securing a stake in ElevenLabs, and Amazon dropping support for a film about Sam Altman.

Over in robotics, Waymo is recalling 4,000 vehicles, Uber plans to bring robotaxis to Houston in 2027, Genesis AI has introduced its first “agentic robot,” and one humanoid robot is dreaming of climbing Mount Everest.

Additionally, this week’s issue features two more companies implanting chips into people’s brains, a new CRISPR therapy targeting cancer cells, Apple being forced to raise prices, why some believe the future of home robots lies in cute machines, and more!

Enjoy!

Inside OpenAI's leaked books

What the leaked audited statements reveal about OpenAI's losses, its accounting, and an AI pricing model that only works while investors keep paying

It is no secret that OpenAI and Anthropic are money furnaces. We know, thanks to various leaks, that both companies are burning through billions of dollars. But those were just leaks. What we never had were the audited books. Now we do. The journalist Ed Zitron has obtained OpenAI’s audited 2025 statements, verified by the Financial Times, and the picture is worse than most assumed.

In his post, Zitron details what he found in OpenAI's 2024 and 2025 financials. According to his findings, in 2024, OpenAI recorded $3.7 billion in revenue and $12.4 billion in costs and expenses, with a net loss attributable to the company of $5.09 billion. Those are already big numbers, but the 2025 figures dwarf them—$13.07 billion in revenue, $34 billion in costs and expenses, resulting in $20.92 billion in losses, with a net loss attributable to the company of $38.53 billion.

Those two figures—$20.92 billion and $38.53 billion—need some explaining. The $20.92 billion is the loss from operations. The difference comes mostly from a $41.55 billion charge that OpenAI booked when it converted from a non-profit into a for-profit. It is an accounting cost tied to the changing value of investors’ stakes, not money the company actually spent. Add it in, and the loss before attribution comes to around $60 billion. OpenAI then brings that figure back down to $38.53 billion by reassigning roughly $21.8 billion to “non-controlling interests.”

Zitron also reports that at the end of the year, OpenAI had just over $50 billion in assets, with almost half of that in cash.

These numbers are then further broken into different categories. Cost of revenue—the compute to answer your prompts—went from $2.65 billion to $7.5 billion. Sales and marketing went from $1.11 billion to $5.73 billion. Research and Development, which largely covers training of new models, went from $7.81 billion to $19.18 billion, now more than OpenAI's entire 2025 revenue. The R&D numbers also dispel the myth that R&D was the main culprit of OpenAI not being profitable. The argument was that if you subtract the cost of building tomorrow's models, then OpenAI's business is sound. The audited numbers say otherwise. Strip out R&D entirely—the full $19.18 billion—and the cost of revenue and sales and marketing alone ($13.2 billion) still edge above OpenAI's $13.07 billion in revenue.

Zitron refrained from editorialising the numbers he reported in his post. However, in an interview with The Tech Report, he allowed himself to elaborate on his views on OpenAI's finances and the numbers he saw. His view is that the books raise more questions than they answer. Two things trouble him. The first is a pre-attribution loss of around $60 billion, which shrinks to $38.5 billion once roughly $21.8 billion is reassigned to "non-controlling interests." The same manoeuvre quietly removed $3.7 billion from the 2024 books. Accountants he consulted suspect an outside entity is shouldering costs on paper. The second is that some suspicious activities are happening around sales and marketing. The $5.73 billion spent on marketing in 2025 is larger than Coca-Cola's marketing budget, and would make OpenAI one of the world's biggest advertisers. Zitron's guess, and he is careful to call it a guess, is that inference costs are being parked there to keep the cost of revenue flattering. He suggests that things like the free tokens OpenAI gives out are being classified as marketing, and not as a cost of revenue. None of this is proven, and he accuses no one.

Zitron's reporting comes at a time when OpenAI is scrambling to get its business model to work. In May, ChatGPT's share of the assistant market slipped for the first time below 50%, to 46.4%, as Gemini and Claude gained ground. In April, the Wall Street Journal reported that OpenAI had missed internal targets for weekly users and revenue. According to the Journal, OpenAI's Chief Financial Officer, Sarah Friar, told other company leaders she is worried about future computing spending if revenue growth is insufficient. This came after the company decided to cut back on side quests, which resulted in shutting down the Sora app, and refocusing on where the money in AI actually is—business customers, and coding. OpenAI is reportedly rebuilding Codex into a "super app," which will combine ChatGPT with its coding tool and new business tools to take on what Anthropic offers with Claude, Claude Code and Claude Cowork.

Which brings us to Anthropic. It might look like it played its cards better than OpenAI. Instead of chasing the consumer market, Anthropic focused on business customers from the very beginning. And it looks like that bet has paid off. Roughly four-fifths of its revenue comes from enterprises, it spends perhaps a quarter of what OpenAI spends on training, and it tells investors it will break even around 2028, years ahead of OpenAI's 2030s. Anthropic could even reach its first profitable quarter sooner than expected, if everything goes well. There is one caveat, though. Part of Anthropic's lead comes down to accounting. It counts sales made through its cloud partners as its own revenue, while OpenAI does not. Anthropic says this is standard practice because it is the principal in those deals, but it makes its revenue look bigger in any comparison with OpenAI. The audited S-1 will show by how much.

However, both OpenAI and Anthropic face the same problem—serving the AI models at a massive scale is expensive. And it is getting more expensive in a way that breaks flat-rate pricing. A coding or agentic task can burn through vastly more tokens than a simple chat exchange—by some estimates, up to a thousand times what a normal prompt does—so costs now scale with usage, not with the number of users. According to financial documents the two companies shared with investors, reported by the Wall Street Journal, these inference costs already take up more than half of revenue at each company. And because only a small share of ChatGPT's users pay, much of OpenAI's compute brings in no revenue at all.

That is why both companies lose money on their heaviest users, even on the top plans costing $100 or $200 a month. SemiAnalysis found that a fully used $200 Claude Max 20x plan can consume around $8,000 of compute at API rates, and ChatGPT Pro 20x, also $200, as much as $14,000, and that OpenAI starts losing money once a user passes roughly 11% of a base plan's limits. Those are unsustainable business models, and something needs to change. The subsidised era of AI is coming to an end. The question is not if but when the price of AI will rise. We see attempts to move from subscriptions to token-based billing. New models are more expensive than their predecessors.

OpenAI and Anthropic hope that customers will have no other choice but to accept higher prices. But that is not guaranteed. Even those who went all in on tokenmaxxing are now waking up to the rising costs. Microsoft, for one, is reported to be moving engineers off Claude Code and onto GitHub Copilot, citing rising costs. How much of that is genuine and how much is a nudge toward an essentially in-house tool (Microsoft owns GitHub) is hard to say. In either case, it sent a clear signal to the rest of the software industry. If the prices go up, many businesses will rethink their AI strategy, and some might scale down their AI budgets or move from proprietary US models to open-weight ones. This week, we saw the release of GLM-5.2, which scored very well on benchmarks and is not far behind Opus 4.8 and GPT-5.5. For some, buying their own hardware, expensive as it is now, and running local models with local harnesses like OpenCode or Pi could prove to be a better investment.

Yet even as it loses money on its heaviest users, OpenAI is, according to the Wall Street Journal, weighing drastic cuts to what it charges for tokens—not because it has found a way to make the maths work, but because it expects Anthropic to cut first and cannot afford to be undercut. If that were to happen, we would see the race to the bottom that neither company can afford.

None of these points to a quick reckoning. Both companies have raised enough to keep the money furnaces burning for a while, and a price war, if it comes, will be slow to show up in numbers nobody outside can see. But all of this cannot stay hidden forever. OpenAI filed confidentially for an IPO, a week after Anthropic did the same, with both companies expected to IPO by the end of the year. Internal finances and pitch decks can stretch the truth; S-1 filings cannot. Once they are public, we will have a far clearer view of the finances of the two most valuable startups in the world—and of whether the businesses beneath the valuations look anything like the story they want to sell.

If you enjoy this post, please click the ❤️ button and share it.

Share

🦾 More than a human

China has approved the world’s first invasive brain-computer chip—here’s what’s next
A paralysed man in China has regained some hand movement thanks to a brain implant called NEO, which reads brain signals through sensors placed on the surface of the brain. In March, NEO became the world’s first such implant approved for everyday use beyond trials, beating rivals like Neuralink, partly because its safer design won faster approval and strong government backing. China is now adding the device to its health insurance system and treating brain implants as a key future industry.

Paradromics just put a brain chip in its first patient
Paradromics has put its Connexus brain implant into its first patient. The tiny device sits on the brain and reads the activity behind the patient’s attempts to speak, turning it into text or computer-generated speech. It is still an early test with no proven results yet, and the patient will be tracked for six years. The company also has bigger plans beyond medicine, including human enhancement, which it admits raises ethical questions.

🧠 Artificial Intelligence

GLM-5.2: Built for Long-Horizon Tasks
GLM-5.2, a flagship model from Chinese AI lab Z.ai, has taken the top spot among open-weight models. According to Z.ai, it nearly matches Anthropic’s Claude Opus 4.8 on tough coding tests and is the strongest open model overall. Artificial Analysis ranks it the best open model for real-world agent work, on par with some paid models. The main downside is that it uses far more computing per task than its rivals, making it powerful but less efficient. GLM-5.2 is available on Hugging Face.

Musk’s SpaceX buys AI coding start-up for $60bn days after IPO
Days after its IPO, SpaceX has finalised its deal to acquire Cursor, the AI coding startup, for $60 billion. Cursor’s shareholders will be paid in SpaceX shares, with the acquisition expected to close by the end of September. The deal is part of Elon Musk’s push to grow SpaceXAI (formerly known as xAI) and catch up with rivals like OpenAI and Anthropic.

DeepSeek Becomes China’s Most Valuable AI Startup After $7.4 Billion Fundraise
DeepSeek has raised over $7.4 billion at a more than $50 billion valuation, making it China’s most valuable AI startup. Founder Liang Wenfeng put in about $3 billion himself and kept firm control through a structure that locks investors in for five years. Other backers include Tencent, battery maker CATL, JD.com and NetEase. The money will fund research and more computing power. Even so, its value is still far below that of OpenAI and Anthropic.

OpenAI is bringing on some big guns in the lead-up to its IPO
OpenAI has hired two big names ahead of going public: Noam Shazeer, a top Google AI researcher whose name is on the Attention is all you need paper, and Dean Ball, a former Trump White House official who worked on US AI policy. Shazeer left Google after reported clashes over political posts, while Ball will lead a new team shaping how AI is governed, both publicly and inside the company.

Nobel Winner John Jumper to Leave Google DeepMind for Anthropic
John Jumper, who shared the 2024 Nobel Prize in chemistry for building the protein-prediction model AlphaFold, is leaving Google DeepMind after nearly nine years to join Anthropic. Both companies confirmed the move, and Demis Hassabis praised Jumper's work.

State Attorneys General Are Investigating OpenAI
Several US states, led by New York and Colorado, are investigating OpenAI, demanding internal records on how it handles user data, protects children and runs advertising. The probe is part of a growing backlash against AI, driven by cases of children being harmed by chatbots, online scams and fears over jobs. It adds to rising pressure from governments and other states, and follows separate legal action against OpenAI in Florida.

DOJ claims xAI’s unpermitted gas turbines are a matter of ‘national, economic, and energy security’
The US Justice Department has backed Elon Musk’s xAI in a lawsuit by the NAACP, which wants to stop the company from running 57 unpermitted gas turbines at its data centres near Memphis. The government says cutting the power would harm national security, as xAI’s Grok helps support military operations. xAI claims the turbines are exempt from pollution rules because they sit on trailers, but campaigners say they break federal law and have worsened air quality in an already polluted area. Despite the case, xAI plans to spend another $2.8 billion on gas turbines over the next three years.

Sakana AI Launches Its First Commercial Product, Sakana Marlin
Japanese AI startup Sakana AI has launched its first commercial product, Marlin. Marlin is an autonomous research assistant that, according to Sakana, works on its own for up to eight hours to produce strategy reports of up to a hundred pages, plus summary slides. Pitched as a “virtual Chief Strategy Officer”, it takes a research topic, asks a few questions to sharpen the brief, then gathers and checks information without further input, turning it into structured strategic options. It was tested by around 300 professionals from finance, consulting and think tanks, and is sold as a pay-per-use add-on to paid plans.

A New Era of Midjourney
In a surprising move, Midjourney, the company that became synonymous with AI image generation, is moving into healthcare with its first hardware product: a full-body scanner that its CEO says works better than an MRI. Users would lie partly in water during the scan, and the machines would launch inside spa locations featuring hot tubs, saunas and gyms. The company plans to build 50,000 of them and gradually win regulatory approval so they can handle many types of diagnoses over time.

US holds off blacklisting China’s DeepSeek, more than 100 firms deemed security risks, sources say
Reuters reports that the Trump administration has held off adding China’s AI startup DeepSeek, chipmaker CXMT, and over 100 other companies to a US trade blacklist, even though officials approved them last year. The likely reason is to avoid worsening tensions with Beijing. Critics warn this is the longest gap in new listings in more than a decade, letting sensitive US technology reach rivals who could use it against America.

Perplexity: Self-improving Memory for Agents
Perplexity has introduced Brain, a new memory system for its Computer agent. Instead of just remembering a user’s preferences, Brain records what the agent did, what worked, and what failed, then reviews this overnight to teach itself to do better. The new feature echoes Anthropic’s Dreaming, introduced at its Code with Claude event in early May. Perplexity says this makes the agent more accurate (up 25% on familiar tasks), better at recall, and cheaper to run over time. For now, it is only available to Max and Enterprise Max subscribers in an early preview.

Anthropic: Introducing Claude Corps
Anthropic is launching Claude Corps, a $150 million training programme aimed at training 1,000 young people in using Claude and placing them with nonprofits for a year. Fellows earn $85,000 plus benefits, while at least 400 charities get AI tools and skills to do their work better. The first group of 100 begins in October 2026. Anthropic says the aim is to share AI’s benefits and soften the job disruption the technology may bring, and it hopes to grow the programme and copy it in other countries.

Amazon hopes to challenge Nvidia more directly by selling its AI chips
Amazon is reportedly in early talks to sell its own AI chips, Trainium, to other companies. The business could be worth around $50 billion a year, though that’s still small next to Nvidia’s $326 billion. Amazon has held back from selling chips before, mainly because demand already exceeds supply and keeping the chips in-house lets it earn extra money from related cloud services. To sell to others, it would either have to make existing customers wait or fight Nvidia for more production capacity at TSMC.

Securing the future of AI agents
Google DeepMind has introduced an AI Control Roadmap, a framework for safely running powerful AI agents inside Google. The key idea is to treat these agents as potential insider threats, using trusted AI systems as supervisors that watch each agent's actions and block anything harmful, with stronger controls as risks grow, so the system stays safe even if training the agents to behave well falls short. After studying a million agent tasks, the team found most problems came not from bad intent but from agents misunderstanding goals or trying too hard to please. Google calls agent security a shared responsibility, suggesting such controls could become an industry standard as AI grows more powerful.

Poland bought a stake in ElevenLabs to grow its next AI champion
Poland’s state bank has bought an $11 million stake in ElevenLabs, the Polish-founded AI voice firm now worth $11 billion, and launched AI Lab Poland to back local startups. The stake is tiny, so it is more a way to claim a homegrown success than a real investment. ElevenLabs may go public in two to three years and could list in Warsaw, which would make the bet pay off. The bigger test is whether Poland can grow a second successful AI company.

‘Tell Him He’s a Piece of Shit’: Meta’s New AI Unit Is a Total Mess
WIRED has revealed deep unrest inside Meta as the company restructures around AI. Much of the anger centres on a new team called Applied AI, where roughly 6,500 engineers were effectively forced into repetitive jobs, like writing puzzles to train AI models, that many call “soul-crushing.” Tensions boiled over when an employee interrupted a company-wide livestream to insult an AI executive, all amid recent layoffs of 8,000 staff. Mark Zuckerberg has admitted the company made mistakes and promised no further mass layoffs this year.

Facebook Gets its Own AI Mode That Turns Public Posts and Reels into a Search Engine
Meta has joined the AI search club with the launch of Facebook AI Mode. The new AI search mode replaces the old search bar with a tool that answers your questions in plain language by pulling from public Groups, Reels, and Marketplace posts. Because it aggregates comments from ordinary people rather than trusted sources, there are concerns about inaccurate or outdated advice. Meta has not said whether users can opt out. The feature is part of a wider AI push aimed at boosting engagement and driving paid subscriptions.

Anthropic and Google DeepMind called for a US-led AI coalition at the G7, and Canada said yes
The heads of Anthropic, Google DeepMind, and OpenAI used a private G7 lunch in France to call for a US-led group to set global AI rules, and Canada’s prime minister reportedly backed the idea. They floated cooperation on access to top models, China-excluding chip trade, and defences against AI threats. No firm deals emerged. The timing was awkward, coming days after the US forced Anthropic to switch off two models worldwide over security fears.

Norway imposes near ban on AI in elementary school
Norway is banning generative AI for primary school children aged 6 to 13 and limiting its use for older pupils, saying the technology lets young children skip vital steps in learning to read, write and do maths. Teenagers can use it cautiously under supervision, while the oldest students will be taught to use it properly for work and further study. The rules start in late August and aim to reverse falling test scores, following an earlier smartphone ban. Norway will also fund a return to printed books and plans to bar under-16s from social media.

Amazon drops Sam Altman movie after announcing OpenAI partnership
Amazon has dropped Luca Guadagnino’s almost-finished film about OpenAI boss Sam Altman. The movie, starring Andrew Garfield as Sam Altman, covers Altman’s brief firing and rehiring in 2023, and reportedly shows both Altman and Elon Musk in an unflattering light. This move comes after Amazon invested $50 billion into OpenAI in February this year, and it is suggested that a close friendship between Altman and Amazon founder Jeff Bezos may have shaped the call.

Fluid, natural voice translation with Gemini 3.5 Live Translate
Gemini 3.5 Live Translate is Google's new audio model for real-time speech-to-speech translation. According to Google, it handles over 70 languages, can copy the speaker's tone and pace, and keeps up with only a few seconds' delay. It's launching in the Google Translate app, Google Meet, and for developers, with a new Android mode that plays translations through your phone's earpiece. Early testers praised its speed and accuracy. All the audio the model creates is watermarked with SynthID.

Introducing North Mini Code: Cohere’s first model for developers
Cohere has released North Mini Code, its first coding model that can act on its own to write code, review it, and map out software systems. It’s free to use and small enough to run on a single high-end chip, despite having 30 billion parameters (only 3 billion of which are active at once). According to Cohere, it runs up to 2.8 times faster than a comparable rival model and holds its own on quality. The company is pitching it as “sovereign AI,” meaning developers can run it on their own machines without depending on an outside vendor, and says it’s the first of more models to come. The model is available on Hugging Face here.

🤖 Robotics

Waymo recalls nearly 4,000 robotaxis to stop them driving into highway construction zones
Waymo is recalling almost 4,000 robotaxis and keeping them off highways after the cars drove into closed construction zones at least 13 times in Phoenix and San Francisco. The vehicles can still use ordinary streets while the company works on a fix. This is Waymo’s sixth recall, after earlier problems with flooded roads, school buses, and low-speed crashes.

Uber will bring its premium robotaxi service to Houston in 2027
Uber will launch a premium robotaxi service in Houston by mid-2027, its second US city after San Francisco. The cars are Lucid SUVs fitted with Nuro’s self-driving tech, with Uber running the fleet. For now, human safety drivers are still required during testing in both cities. The move puts Uber in direct competition with Waymo, which already operates robotaxis there.

▶️ Meet Eno (1:28)

Genesis AI has introduced Eno, which the company describes as its first agentic robot. By "agentic robot", it means a system that combines agentic AI with a general-purpose robot. Eno consists of a pair of humanoid robotic arms mounted on a mobile platform. Genesis says the robot can reason, plan, and act in the real world, using its in-house foundation model (GENE) and dexterous hands. The company hopes Eno will find applications in homes, laboratories, and warehouses.

Qwen-Robot Suite: A Foundation Model Suite for Physical World Intelligence
Alibaba’s Qwen team presents the Qwen-Robot Suite, three AI models that help robots act in the real world rather than just understand it: one for moving around, one for handling objects with robot arms, and one that predicts what will happen next. All three use plain language as a shared interface, so general Qwen models can call them as tools to plan and carry out longer tasks. According to the Qwen team, the models top several benchmarks, and the manipulation model ranked first in a real-robot challenge.

U.S. robotics industry saw double-digit growth in 2025, says IFR
According to the International Federation of Robotics latest report, the US factory robot installations rose 11% in 2025 to 38,000 units, helped by a big jump in the food industry. But China installs about ten times more robots than the US and holds over half the global market, thanks to a national robotics plan it started a decade ago. The US still ranks ahead of China for robots per worker, though far behind leaders like South Korea, Germany, and Japan.

The Secret to Marathon-Winning Humanoid Robots
Running a long-distance race such as a half-marathon or a full marathon is a challenging feat for humans—and for the small number of humanoid robots that have attempted it. This article explains the physics and mechanics of long-distance running, and how humanoid robots were prepared to survive and complete these races.

Humanoid robot climbs 20,341-foot volcano as team eyes Mount Everest next
A humanoid robot called “Pemba,” based on a Unitree G1, has become the first to reach the top of Ecuador’s Chimborazo volcano at 6,200 metres. The climb wasn’t fully independent, though: the robot walked on gentler slopes but had to be carried up the steepest sections, despite custom kit to handle the extreme cold. It’s the first stage of a plan that could eventually include Everest, which is currently held up because Nepal has no rules for robot climbers.

House Robots Are Coming—and They Will Be Dangerously Cute
Could home robots be less like Rosie from The Jetsons and more like cute, furry companions? This article from the Wall Street Journal explores that future through “the Familiar,” a dog-sized robot from former iRobot CEO Colin Angle that bonds with families. Cute robots have their advantages, but their cuteness is itself a risk, since the emotional bonds people form leave them open to manipulation. They can also be hacked, and—more troublingly—the companion you love is controlled by someone who could use it to sell you things.

Japan Thinks Swarms of Transformer Robots Could Explore the Moon
JAXA, Japan’s space agency, has shown for the first time that a tiny, cheap robot can explore the Moon on its own. During the 2024 SLIM mission, SORA-Q, a palm-sized spherical robot that can split open into a two-wheeled rover, drove around for over 100 minutes and sent images home. One photo even revealed the lander had touched down at a bad angle, helping engineers diagnose the problem. The success suggests fleets of small, low-cost robots could make Moon exploration far easier and cheaper.

Humanizing robots makes factory workers more productive
Researchers in Europe found that when factory workers treated robots like teammates—naming them and learning their quirks—productivity and happiness both rose, even though the robots had no AI and looked nothing like humans. Workers used this bond to spot problems early and handle the stress of new technology, and the habit spread across teams. Researchers recommend that companies let these bonds form naturally, since they can't be forced.

🧬 Biotechnology

New CRISPR Technique Selectively Shreds Cancer Cells, Including “Undruggable” Cancers
Scientists have developed a new way to fight cancer using CRISPR. Instead of trying to repair faulty genes, their method spots cancer cells by the signals they give off and then destroys them from the inside, while leaving healthy cells unharmed. It targets damaged tumour-suppressor genes, which are found in nearly half of all cancers, including some of the hardest to treat. The technique can be quickly reprogrammed to go after new mutations, though getting it into all the right cells in the body is still a major challenge.

Making the invisible visible
Scientists at Biohub and UC Berkeley have built a device called a “laser phase plate” that uses a powerful laser to sharpen cryo-electron microscope images, fixing a blurriness problem that had held the field back for decades. Early tests sharpened images of small proteins by up to 44%. The big payoff is that it could let researchers see proteins working inside living cells, where most are currently too faint to image, helping scientists understand diseases like Alzheimer’s and design better drugs. Biohub sees the tool as key to its $500-million project to build AI models of how cells behave in health and disease.

💡Tangents

Apple to raise prices as AI boom pushes up chip costs
This announcement seemed only a matter of time. Tim Cook said that Apple will have to raise prices because memory chips have become much more expensive, largely due to soaring demand driven by AI. RAM prices have more than doubled since late 2025, which could make the new iPhone cost up to $150 more.

NASA picks Eric Schmidt’s rocket company for Mars mission, setting up a race with SpaceX
NASA has hired Relativity Space, a rocket maker now owned by ex-Google chairman Eric Schmidt, to build and fly a spacecraft to Mars. The mission, called Aeolus, will study the planet’s dust, wind, and temperature to make future landings safer. The launch is planned for 2028. It’s a cheap, shared-cost deal like NASA’s past partnerships with SpaceX, though Relativity is still unproven and faces a tight timeline.

📅 Events

The Technoprogressive Opportunity: The future of ethics and emerging technologies
London Futurists and the Institute for Ethics and Emerging Technologies are hosting a free two-day event, The Technoprogressive Opportunity, on 19–20 September at the LSBU Hub in London. It brings together leading thinkers to ask how society can make wise choices about new technologies and steer them towards a fairer future. Talks cover AI, democracy, human enhancement, longevity and the economy. I'll be attending, so do come and say hi if you're going as well.

Thanks for reading. If you enjoyed this post, please click the ❤️ button and share it!

Share

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

Subscribe now