Anthropic vs Pentagon - Sync #559

Plus: OpenClaw joins OpenAI; Claude Sonnet 4.6; Gemini 3.1 Pro; AI Impact Summit in Delhi; AI researchers raise billions; Chinese martial arts robots; new CRISPR methods; scent, in silico; and more!

Hello and welcome to Sync #559!

This week’s main story covers the dispute between Anthropic and the Pentagon over the use of Claude in military operations, and what happens when Anthropic’s principles are tested by the US government.

Elsewhere in AI, Anthropic released Claude Sonnet 4.6, Google released Gemini 3.1 Pro and Lyria 3, and Qwen3.5 topped the open model charts. We also saw two AI researchers raise billions of dollars for their start-ups, the AI Impact Summit in Delhi, and the creator of OpenClaw joining OpenAI.

Over in robotics, the world was amazed by a group of Unitree humanoid robots performing synchronised martial arts routines alongside children, Agility Robotics secured another commercial customer, and a Hungarian startup is weaving its robots into existence.

In addition, this week’s issue of Sync features a conversation with Dario Amodei, Colossal Biosciences announcing a new deal with the United Arab Emirates to build “Biovaults”, Helion taking a step closer towards commercial fusion reactors, the story of digitising smell, and more!

Enjoy!

Anthropic vs Pentagon

Isaac Asimov’s First Law of Robotics says that a robot may not injure a human being. The principle has echoed through decades of AI development, shaping the safety frameworks and usage policies that companies like Anthropic have built around their models. Now, that principle is being tested.

The Wall Street Journal reported that the Pentagon used Anthropic’s Claude to help plan the military operation that captured former Venezuelan President Nicolás Maduro. The raid included bombing several sites in Caracas. Anthropic’s usage guidelines prohibit Claude from being used to facilitate violence.

That contradiction is now the centre of a dispute that could reshape the relationship between the AI industry and the US military.

Claude was deployed through Anthropic’s partnership with Palantir, the defence contractor whose tools are widely used across the military. When an Anthropic employee contacted a Palantir counterpart to ask how the model had been used in the operation, according to the Journal, the reaction was swift. A senior administration official told Axios the inquiry implied that Anthropic “might disapprove of their software being used, because obviously there was kinetic fire during that raid, people were shot.” Anthropic denies this, saying it has not discussed the use of Claude for specific operations with the Pentagon or Palantir outside of routine technical matters.

Either way, the damage was done, and the Pentagon announced it is reviewing its relationship with Anthropic.

The contract

Claude is currently the only large language model cleared for classified operations—a significant competitive advantage that made Anthropic an early favourite in the Pentagon’s AI push. The company signed a contract worth up to $200 million last summer alongside OpenAI, Google, and xAI, and built custom classified models.

But the contract has been contentious from the start. The Pentagon wants all four AI labs to allow their tools to be used for “all lawful purposes”—including weapons development, intelligence collection, and battlefield operations. Anthropic insists that two areas remain off limits: the mass surveillance of Americans and fully autonomous weaponry.

The Pentagon argues this is unworkable. There is “considerable gray area” around those categories, the senior official told Axios, and the military cannot negotiate individual use cases or risk Claude blocking applications mid-operation. OpenAI, Google, and xAI have all shown more flexibility. Pentagon CTO Emil Michael was blunt: “If any one company doesn’t want to accommodate that, that’s a problem for us.”

The threat is no longer theoretical. The Journal reported that Defence Secretary Pete Hegseth is close to designating Anthropic a supply chain risk—a classification typically reserved for foreign adversaries like Huawei, which would require any Pentagon contractor to certify it does not use Claude. Applying such a designation to a US company would be unprecedented, and Dean Ball, a former Trump AI policy adviser, told the Journal it would be "the most strategically unwise move" the military could make in the AI competition.

The politics

But this is more than a contractual dispute—the two sides simply do not like each other. Dario Amodei, CEO and co-founder of Anthropic, compared Trump to a “feudal warlord” in a now-deleted pre-election Facebook post and has criticised the administration’s chip export policies. Anthropic employs several former Biden officials, and AI czar David Sacks has accused the company of being “AI doomers.” Hegseth’s Pentagon has framed the standoff in culture-war terms, casting “woke” tech companies as a liability.

The broader context makes Anthropic’s position all the more unusual. In 2018, thousands of Google employees signed a petition protesting the company’s military AI work under Project Maven. In 2026, every major AI lab is competing for Pentagon contracts. Anthropic is the outlier not for seeking military work, but for placing any limits on it at all.

The company has tried to build bridges—most recently adding Chris Liddell, a deputy chief of staff for policy coordination during Trump's first term, to its board. It has not changed the dynamic. Hegseth is using the dispute to send a message, a defence official told the Journal—not just to Anthropic, but to every AI company that might consider placing limits on military use.

The choice

Beneath the politics lies a genuinely difficult problem. If artificial intelligence is the most powerful technology ever invented—as every major lab claims—then integrating it into military operations requires extraordinary care. Anthropic made its name championing the safe and responsible use of AI.

Yet the company also chose to pursue military contracts, becoming the first AI lab on classified networks. That created an inherent tension: you cannot be both the industry’s most safety-conscious company and a defence contractor without, eventually, being forced to choose.

The Pentagon’s position is straightforward. As Michael argued, if a drone swarm is incoming and human reaction time is insufficient, restrictions on AI become a liability. Anthropic’s position is that some lines should not be crossed—that mass surveillance and autonomous killing are categorically different from document summarisation and intelligence analysis.

Anthropic is standing by its principles, and that is commendable. But principles come at a price. The company risks losing not just a lucrative $200 million contract, but its privileged position as the only AI lab on classified networks. The repercussions are already visible beyond the Pentagon. When Anthropic recently approached 1789 Capital, a pro-Trump venture firm where one of the president's sons is a partner, the firm declined on ideological grounds, the Wall Street Journal reported. Given how vengeful and transactional the current US administration has shown itself to be, the consequences may not stop there.

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

---

🧠 Artificial Intelligence

Introducing Claude Sonnet 4.6
Anthropic released Claude Sonnet 4.6, its most advanced Sonnet model yet, with major improvements in coding, long-context reasoning, computer use, planning, and design, plus a new 1 million token context window in beta. According to Anthropic, it shows significant gains in computer-use benchmarks, has strong safety results and offers performance close to Opus 4.5 at a lower cost. It is now the default model for Free and Pro users at the same price as Sonnet 4.5.

Nvidia and OpenAI abandon unfinished $100bn deal in favour of $30bn investment
Nvidia is close to making a $30 billion equity investment into OpenAI as part of a broader funding round, according to the Financial Times, with the deal potentially being finalised soon. This comes after a previously announced $100 billion deal was reportedly put on hold.

India hungry to harness US tech giants’ technology at Delhi summit
At the AI Impact Summit in Delhi, Narendra Modi described AI as a major turning point for civilisation and key to boosting India’s economic growth. US companies such as OpenAI, Google and Anthropic are expanding in India under a new US–India technology agreement, as Washington competes with China for AI influence. In total, India bids to attract over $200 billion in AI infrastructure investment by 2028. While AI could drive rapid growth and transform services and industry, there are concerns that relying too heavily on foreign technology could weaken India’s independence. TechCrunch has a good list of all announcements made at the summit. Additionally, the event produced a wonderfully awkward photo of Sam Altman and Dario Amodei apparently refusing to hold hands for a photo opportunity.

Gemini 3.1 Pro: A smarter model for your most complex tasks
Google has announced Gemini 3.1 Pro, an upgraded core model behind its latest Gemini 3 Deep Think update, aimed at solving complex problems in science, research and engineering. The company says 3.1 Pro offers much stronger reasoning, scoring a verified 77.1% on the ARC-AGI-2 benchmark—more than double the reasoning performance of 3 Pro—and can better handle complex tasks, combine data, and explain difficult topics clearly. It is rolling out in preview for developers through the Gemini API (via Google AI Studio, Gemini CLI, Antigravity and Android Studio), for enterprises in Vertex AI and Gemini Enterprise, and for consumers in the Gemini app and NotebookLM, with higher limits and exclusive access for Google AI Pro and Ultra users ahead of general availability.

Lyria 3

Lyria 3 is DeepMind’s most advanced AI music generation model yet, designed to create high-fidelity music and audio. It can turn a simple text prompt (or even an uploaded image) into a short, polished music track, including instruments, vocals and lyrics in a wide range of styles. It also gives users more control over things like tempo, mood and genre, and DeepMind says outputs are watermarked (using SynthID) so AI-made music can be identified. It’s currently available in beta, rolling out through the Gemini app.

▶️ Dario Amodei — “We are near the end of the exponential” (2:22:19)

Dwarkesh Patel sits down with Dario Amodei, founder and CEO of Anthropic, to discuss AI scaling laws, reinforcement learning, continual learning, economic diffusion, and the capital dynamics of frontier AI labs. Amodei argues that the “Big Blob of Compute” hypothesis continues to hold, that we are approaching the “end of the exponential”, and that the public underestimates how close we may be to a “country of geniuses in a data centre”. He defends a soft take-off model of rapid but smooth capability growth, explains why trillions in revenue before 2030 are highly likely, outlines why the industry may stabilise into a high-margin, few-player equilibrium, and discusses why even with short AGI timelines, compute investment must be balanced against bankruptcy risk and diminishing returns. Kudos to Dwarkesh for challenging Dario on some points.

OpenAI resets spending expectations, tells investors compute target is around $600 billion by 2030
According to CNBC, OpenAI has told investors it now plans to spend about $600 billion on computing infrastructure by 2030, much less than the $1.4 trillion previously mentioned, as it faces questions about whether it can earn enough to cover its costs. The company expects revenue to reach more than $280 billion by 2030, up from $13.1 billion in 2025, with growth coming from both consumer and business products such as ChatGPT and its coding tool, Codex.

What’s behind the mass exodus at xAI?
The Verge reports on a wave of departures and internal turmoil at xAI following its recent merger with SpaceX. Several cofounders and employees have left, with former staff claiming the company is focused on NSFW Grok content, lacks safety oversight, and is largely playing catch-up to rivals like OpenAI. The piece also highlights Musk’s ambitious “space-based AI” vision and how some departing employees are using their equity and experience to start new AI companies.

OpenClaw, OpenAI and the future
Peter Steinberger, the creator of OpenClaw, is joining OpenAI. While he believes OpenClaw could become a big company, that isn’t his goal, so he’s ensuring it stays open source by moving it into an independent foundation supported by OpenAI.

OpenClaw shows us the future of AI—and why it is not ready yet

Anthropic-funded group backs candidate attacked by rival AI super PAC
A political fight over AI funding has broken out in New York’s 12th district, where Assembly member Alex Bores is being attacked by the pro-AI super PAC Leading the Future, which has raised over $100 million from major tech backers and spent $1.1 million criticising him for supporting a bill that would require AI companies to share safety information. In response, another group, Public First Action—backed by a $20 million donation from Anthropic—is spending $450,000 to support Bores, promoting a pro-AI approach focused on transparency, safety and public oversight.

Introducing Manus in Your Chat
Manus has launched Manus Agents, which lets users access Manus directly inside messaging apps, starting with Telegram, and run full multi-step tasks through chat, including research, writing reports, and creating documents. Users can also send voice notes, images, and files, choose how the agent responds, and switch between faster or more powerful models. Manus Agents are available for all users across all subscription tiers.

Meta buys Manus for over $2 billion - Sync #552

AI Pioneer Fei-Fei Li’s Startup World Labs Raises $1 Billion
World Labs, an AI startup founded by an AI pioneer Fei-Fei Li, has raised $1 billion in new funding from investors including Autodesk, Andreessen Horowitz, Nvidia and AMD. The company is working on “world models”, a type of AI designed to understand and make decisions in the 3D real world. It recently launched a product called Marble that can create 3D worlds from text or image prompts. World Labs plans to use the money to improve tools for robotics and scientific research, but it did not share its valuation.

David Silver is chasing superhuman intelligence with a $1bn seed
David Silver, an AI researcher who helped build DeepMind’s AlphaGo and AlphaStar, is raising a huge $1bn seed round for his new startup, Ineffable Intelligence. The round is reportedly being led by Sequoia, with Nvidia, Google and Microsoft also in talks, and could value the company at about $4bn despite it not having launched a product yet. If it goes through, it would be the biggest seed round ever in Europe and a major sign that investors are backing bold AI research in the region.

Saudi Arabia’s AI venture Humain invests $3bn in Elon Musk’s xAI
Saudi Arabia’s state-backed AI company Humain has invested $3bn in Elon Musk’s xAI, becoming a significant minority shareholder before its stake was converted into shares in SpaceX following xAI’s merger with the rocketmaker. The move forms part of the kingdom’s broader strategy to diversify its economy and establish itself as a global AI hub, backed by heavy investment in data centres and AI models.

Micron Is Spending $200 Billion to Break the AI Memory Bottleneck
Micron, America’s biggest maker of memory chips, is racing to build more factories in response to the major memory shortage caused by the AI boom. The company is spending $50 billion to expand in Boise, Idaho, and is also starting a $100 billion project in New York, with new production expected from 2027 onwards.

Meta expands Nvidia deal to use millions of AI chips in data center build-out, including standalone CPUs
Meta has made a big new multi-year deal with Nvidia to use millions of its chips in Meta’s AI data centres, including Nvidia’s Grace standalone CPUs, next-generation Vera Rubin systems, and networking technology. Although no price was shared, analysts say it is likely worth tens of billions and fits into Meta’s wider plan to spend up to $600 billion in the US by 2028.

U.S. court bars OpenAI from using ‘Cameo’
A court in Northern California ruled that OpenAI must stop using the name “Cameo” for a feature in its Sora 2 video app because it could confuse people with the real Cameo platform. OpenAI had already renamed the feature to “Characters”, but said it disagrees with the ruling and plans to continue fighting the case.

Mistral AI buys cloud startup Koyeb
Mistral AI has bought French startup Koyeb to strengthen the infrastructure needed to run and scale its AI models in real-world use. Koyeb builds a serverless cloud platform that lets developers deploy AI apps without worrying about managing servers. The move shows Mistral wants to control more of the full AI stack, while offering a more European alternative to US cloud giants like AWS, Microsoft, and Google.

AI Is Getting Scary Good at Making Predictions
The article suggests a possible new era where humans may no longer be the best predictors of the future—and we may rely on AIs even when we don’t understand how they reached their conclusions. It explains how AI systems are rapidly improving in forecasting tournaments and prediction markets, now competing with—and sometimes beating—elite human forecasters. By combining multiple language models and processing huge amounts of information quickly, these tools may soon become the world’s most trusted guides to what happens next.

Sony patents AI-generated podcasts voiced by characters from its games
Sony has patented an idea for AI-made podcasts that are personalised for each gamer and voiced by characters from video games. The podcast would use things like what the player has been playing, what their friends have done, and console updates to give daily news and tips. Players could choose what they want included, and it might even add jokes.

AI Is Killing the Need for Lawyers, But Increasing the Demand for Legal Thinkers
Tobias Mark Jensen of Futuristic Lawyer (good Substack, btw) argues in this post that AI will rapidly automate most legal work, making the memorisation-and-regurgitation skills law school rewards today largely obsolete. He believes the future belongs to “legal thinkers” who work at a higher level of abstraction—grappling with meaning, ethics, and real-world impact.

How Nvidia became the first $5 trillion company, in 4 charts
This article presents four charts showing how Nvidia has grown in the last few years, all thanks to the AI boom and surging demand for its GPUs.

Qwen3.5: Towards Native Multimodal Agents
The Qwen team has launched Qwen3.5, a native multimodal model that performs strongly in reasoning, coding, agent tool use and multimodal tasks, with the aim of improving productivity for developers and enterprises. The company says the upgrade is driven by large-scale reinforcement learning across a wide range of agent environments, broader multilingual support, early text–vision integration, and a 1M-token context window in the hosted Qwen3.5-Plus version. According to benchmark results shared by the team, Qwen3.5 is competitive with leading frontier models across reasoning, coding, agent tasks and long-context evaluation.

🤖 Robotics

▶️ Martial arts robots dazzle at 2026 Spring Festival Gala (4:50)

The Chinese Spring Festival Gala, a hugely popular televised variety show broadcast on Chinese New Year’s Eve, continues featuring performances by humans and humanoid robots. In this year’s show, a group of Unitree robots performed synchronised martial arts routines alongside children. The results are impressive and show how much progress Chinese robotics companies have made in just one year. Unitree also released a behind-the-scenes video showing the robots preparing for the show.

Agility Robotics Announces Commercial Agreement with Toyota Motor Manufacturing Canada
Agility Robotics has signed a commercial agreement with Toyota Motor Manufacturing Canada (TMMC) to deploy its humanoid robot, Digit, across Toyota’s Canadian facilities following a successful pilot. Digit will support manufacturing, supply chain and logistics operations by taking on repetitive and physically demanding tasks, improving safety and operational efficiency while enabling employees to focus on higher-value work. TMMC joins a growing list of major partners using Digit, including GXO Logistics, Schaeffler and Amazon.

Allonic’s 3D Tissue Braiding—a new, simpler way to build robotics

Allonic, a Hungarian robotics start-up, is bringing a new method for building robots to the table—one that involves weaving. The company showcased its 3D Tissue Braiding by building a fully biomimetic robotic hand with braided tendons and pulleys. According to Allonic, this method is cheaper and quicker than traditional methods, while producing stronger, monolithic robotic bodies in a single automated process, at any scale and level of complexity.

Waymo defends use of remote assistance workers in US robotaxi operations
Waymo told US lawmakers that its remote staff do not drive or control its robotaxis on public roads. Waymo said remote workers only provide advice when the vehicle’s automated system requests help in ambiguous situations, and that only a US-based event response team could potentially move a stopped vehicle at very low speed—something it says has not happened outside of training. Lawmakers have raised safety and national security concerns, and the US Transportation Department has been asked to look into the issue.

Waymo’s big ten days - Sync #558

Tesla Begins Cybercab Production. Now Comes The Hard Part
Tesla says its first purpose-built driverless vehicle, the Cybercab, has rolled off the production line at its Texas factory. The two-seat car has no steering wheel or pedals and is meant to run fully autonomously using Tesla’s Full Self-Driving software.

Robot Dogs Are on Going on Patrol at the 2026 World Cup
Authorities in Mexico’s Guadalupe, Nuevo León, have introduced four unarmed robot dogs to help with security at Monterrey’s BBVA Stadium ahead of the 2026 World Cup. The robots have cameras, night vision and speakers, and are controlled by operators to help spot suspicious activity, manage crowds, and warn people if needed. They have already been used at a recent football match, and are part of a wider plan that also includes surveillance drones and anti-drone technology.

🧬 Biotechnology

Colossal BioSciences, UAE To Launch Biovault And Lab At Museum Of The Future
Colossal Biosciences, the startup known for ambitious efforts to revive extinct animals and protect endangered species, has announced a nine-figure partnership with the United Arab Emirates to build a major genetics lab and “BioVault” in Dubai’s Museum of the Future. Backed by the UAE’s crown prince, the project will store millions of genetic samples from over 10,000 species and serve as the first in a planned global network designed to safeguard biodiversity against ecological disasters and species loss.

Loyal Raises $100M Series C, Led by age1, to Advance the First Canine Longevity Drug
Loyal, a company developing lifespan-extension drugs for dogs, has raised $100M in Series C funding, bringing its total funding to over $250M. The capital will help Loyal finish the remaining FDA steps and prepare to launch its lead drug, LOY-002. Loyal has already completed two of the three main requirements for FDA Expanded Conditional Approval, and its large STAY trial is now fully enrolled, tracking 1,300 dogs across 70 veterinary clinics. If approved, LOY-002 could become the first FDA-approved drug designed to extend lifespan in any species.

Souped-Up CRISPR Gene Editor Replicates and Spreads Like a Virus
Scientists have created a new CRISPR system called NANITE that can move from edited cells to nearby cells, helping solve a major problem in gene therapy: reaching enough target cells to make a difference. NANITE packages CRISPR into virus-like carriers so the tool can spread, making it about three times more effective in lab-grown cells and greatly reducing a harmful disease protein in mice. This could mean lower doses, fewer side effects, and gene-editing treatments that are safer and easier to use for more genetic diseases.

New gene-editing method could advance regenerative medicine
Scientists at the University of Illinois Chicago have developed a new CRISPR method that edits genes in a specific order over time, rather than all at once. Their system uses a chain of guide molecules that switch each other on one by one, helping cells follow the same step-by-step process used in natural development. The researchers showed it works in human stem cells and say it could make it easier to create large numbers of useful cells for treatments, such as insulin-producing cells for people with type 1 diabetes.

Studies test whether gene-editing can fix high cholesterol. For now, take your medicine
Scientists are testing a new gene-editing treatment that could lower high cholesterol with a single dose, instead of needing pills for life. Early small studies suggest it can cut “bad” LDL cholesterol by about half, but researchers still need much bigger and longer trials to make sure it’s safe and works long-term.

💡Tangents

Fusion startup Helion hits blistering temps as it races toward 2028 deadline
Helion, a fusion energy start-up backed by Sam Altman, says its Polaris prototype has reached plasma temperatures of around 150 million°C, which is about three-quarters of what it thinks it needs for a commercial reactor. It has also begun testing with deuterium–tritium fuel, something only a few private companies have done. The milestone supports Helion’s very ambitious plan to start selling fusion electricity by 2028, though it still has major technical hurdles to clear and is racing against other fusion firms aiming for the early 2030s.

Scent, In Silico
We have digitised vision and hearing, but smell has so far eluded digitalisation. This article tells the story of smell: its evolutionary history, our attempts to understand it, and how we are now trying to digitise it. If successful, computational smell could help detect threats and information invisible to cameras, such as gas leaks, food spoilage, disease markers in breath, counterfeit products, and more.

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

Humanity Redefined sheds light on the bleeding edge of technology and how advancements in AI, robotics, and biotech can usher in abundance, expand humanity's horizons, and redefine what it means to be human.

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!

My DMs are open to all subscribers. Feel free to drop me a message, share feedback, or just say "hi!"