AI Browser Wars Have Begun - Sync #542
Plus: Nike unveils an exoskeleton; Statement on Superintelligence; Amazon plans to replace 600k US workers with robots; Karpathy on AGI; DeepSeek-OCR; Unitree H2; and more!
Hello and welcome to Sync #542!
This week, OpenAI launched ChatGPT Atlas, the latest addition to the growing cohort of AI browsers promising to rethink how we search, read, and get work done online.
Elsewhere in AI, Google and Anthropic have signed a cloud deal worth tens of billions of dollars, Apple is reportedly struggling with its AI strategy, and Meta has announced job cuts across its AI operations. Meanwhile, a group of experts, scientists, and public figures is calling for a temporary halt to the development of artificial superintelligence.
Over in robotics, Amazon plans to replace 600k US workers with robots, Unitree shows a new H2 humanoid robot, and a new racing drone has broken the world speed record.
This week’s issue of Sync also features Project Amplify, Nike’s futuristic powered exoskeleton designed to help everyday runners and walkers move faster and farther with less effort, how ageing clocks can help us understand why we age, why AI companies are racing to build a virtual human cell, Andrej Karpathy sharing wisdom for 2 hours and 26 minutes, and more!
Enjoy!
AI Browser Wars Have Begun
With ChatGPT Atlas, OpenAI joins the growing cohort of AI browsers promising to rethink how we search, read, and get work done online
For years, Chrome, Safari, and other browsers have been the gateways to the internet. Now, a new wave of AI browsers is starting to challenge that status quo. The latest of these is ChatGPT Atlas, OpenAI’s new AI-powered browser.
Atlas launches first on macOS for Free, Plus, Pro, and Go users, with Windows and mobile versions coming later. At first glance, Atlas looks like a standard browser. But a built-in ChatGPT side panel is always available to explain, summarise, or dig deeper into what you’re viewing. There’s no need to copy and paste text into a chatbot—Atlas already knows the page you’re on. It also remembers browsing history, which allows for more personalised responses over time. There’s also agent mode, which can carry out small tasks on your behalf.
Atlas joins a new cohort of AI browsers, such as Perplexity’s Comet. Then there’s Microsoft’s Edge, now positioning its Copilot Mode as “an AI browser” in its own right. Google has done a similar thing with Chrome and Gemini earlier this year. There is also Dia from The Browser Company (now owned by Atlassian), Opera Neon, and more.
All of them promise to rethink how we search, read, and get work done online. The idea of an “agentic web”—where AI handles the tedious bits of online life—is compelling. The AI agent built into the browser will have access to pages and accounts that you have access to, opening interesting possibilities for delegating or automating some tasks.
However, we are still in the early stages of AI browsers, and the question now is whether any of this is genuinely useful and actually changes how most people browse the internet.
There’s also the matter of trust. AI browsers rely on seeing what you see, which means allowing them to read pages, tabs, and sometimes even your full browsing history. That level of visibility goes far beyond what most people expect from a traditional browser. OpenAI, Microsoft, and Perplexity all say that data is handled responsibly, but the details can be opaque, and policies can change. For many users, the question isn’t just “Can this browser help me?” It’s also “Do I want a company—or its AI model—looking over my shoulder as I browse?”
Security is a concern, too. Researchers have shown how easy it is to trick models like ChatGPT and Gemini with malicious prompts. AI browsers can expose their users to new attacks and vulnerabilities. This isn’t hypothetical. Security researchers at Brave recently demonstrated that AI browsers are vulnerable to prompt injection attacks via text and images.
For now, AI browsers feel like early versions of a future that hasn’t fully arrived. But the direction is clear. The browser is shifting from a window onto the web to a conversational layer that sits on top of it. Whether that becomes a genuine upgrade or just another layer of complexity depends on whether these tools can prove useful beyond the demo stage.
If you enjoy this post, please click the ❤️ button or share it.
🦾 More than a human
Nike Unveils Project Amplify, the World’s First Powered Footwear System for Running and Walking
Nike has unveiled Project Amplify, a powered exoskeleton designed to help everyday runners and walkers move faster and farther with less effort. The system uses a small motor and battery to support the lower leg, making movement feel easier, particularly at slower running or walking speeds. Nike says over 400 athletes have tested multiple versions, covering more than 2.4 million steps, with some reporting that hills feel like flat ground and that the system can help improve from a 12-minute mile to a 10-minute mile. Project Amplify is still in testing, with a wider release planned in the coming years.
Vision Restored Using Prosthetic Retinal Implant
Researchers have successfully restored limited vision in older patients with advanced age-related macular degeneration using a tiny wireless retinal implant paired with camera-equipped glasses. In a study, most participants regained enough sight to read, although the vision remained blurry, black-and-white and narrow in field. While not a cure, experts say the technology offers new hope where previous treatments could only slow deterioration.
How aging clocks can help us understand why we age—and if we can reverse it
The article explores “ageing clocks”, which estimate a person’s biological age by measuring internal markers such as DNA methylation rather than simply counting years. Scientists hope these clocks can help them understand how ageing works and whether it can be slowed or even reversed. However, the clocks are still not accurate enough for personal use, even though influencers and wellness companies are already using them to make bold anti-ageing claims.
🧠 Artificial Intelligence
▶️ Andrej Karpathy — AGI is still a decade away (2:26:07)
It’s always worth making time to listen to what Andrej Karpathy has to say, and this conversation with Dwarkesh Patel is no exception. It’s dense with insights, and each topic discussed could easily be its own podcast episode. Karpathy explores the current state and future trajectory of AI, arguing that we are entering a “decade of agents,” not a “year of agents.” He outlines the limitations of today’s AI systems—especially in memory, reasoning, reliability and tool use—and critiques current LLMs for memorising too much and reasoning too little, along with a pointed criticism of reinforcement learning as a training method, while also proposing ways to address these issues. Finally, he reflects on the future of education and why we need to build new learning institutions for the AI era.
Statement on Superintelligence
A group of experts, scientists, public figures, and over 46,000 other people have signed a statement organised by the Future of Life Institute calling for a temporary halt to the development of artificial superintelligence (ASI). The statement warns that creating systems more intelligent than humans could pose serious risks, including job losses, threats to civil liberties, national security concerns, and even the potential for human extinction. It urges that ASI development be paused until there is strong scientific confidence and public agreement that it can be done safely.
Google and Anthropic announce cloud deal worth tens of billions of dollars
Anthropic and Google have announced a major cloud partnership that will give Anthropic access to up to one million of Google’s TPU chips, adding more than a gigawatt of AI computing capacity by 2026. The deal, worth tens of billions of dollars, supports Anthropic’s multi-cloud strategy, which balances Google’s TPUs, Amazon’s Trainium chips, and Nvidia GPUs to optimise performance and cost while maintaining independence from any single provider.
Alibaba Cloud says it cut Nvidia AI GPU use by 82% with new pooling system
Alibaba Cloud says its new Aegaeon system can cut the number of Nvidia GPUs needed to run large language models by up to 82%. Instead of dedicating one GPU to one model, Aegaeon shares GPU time at a very fine level, letting several models run on the same hardware at once. In tests, this reduced the number of GPUs needed from 1,192 to 213, while also increasing overall performance. These results, however, may depend on Alibaba’s tightly integrated infrastructure and network stack.
Meta slashes jobs in its AI operations
Axios reports that Meta is cutting about 600 jobs in its older AI teams because the company felt the groups had become too slow and complex. The move is meant to make decision-making faster and give employees more responsibility. However, Meta is still growing its new TBD Lab, which focuses on more ambitious AI work, and is continuing to hire new talent for it.
Apple loses another AI exec to Meta
Ke Yang, a key executive leading Apple’s efforts to build an AI-driven web search, is moving to Meta. His exit, along with those of former AI models head Ruoming Pang and around a dozen others—some of whom have joined Meta’s new Superintelligence Labs—raises concerns about Apple’s ability to deliver its anticipated Siri upgrade, with some staff expecting more to leave soon.
Apple employees have ‘concerns’ over Siri performance in early iOS 26.4 builds
Apple’s AI division is experiencing significant internal problems, Mark Gurman reports. Delays and setbacks in improving Siri have led to key staff and leaders leaving the company. Additionally, there is uncertainty over who will guide Apple’s AI strategy next. While iPhone sales remain strong for now, Apple may struggle as AI becomes more important to consumers.
DeepSeek-OCR
DeepSeek has released an open-source model, DeepSeek-OCR, that processes text by converting it into images instead of using traditional text tokens. This visual approach achieves around 10 times greater efficiency and preserves formatting and layout, while avoiding many issues associated with tokenisers. The research suggests it could eventually enable language models with much larger context windows, potentially reaching tens of millions of tokens. However, it is still unclear how well models will be able to reason over this compressed, image-based representation of text. DeepSeek-OCR is available on HuggingFace.
Is China Serious About AI Safety?
The article discusses the launch of the China AI Safety and Development Association (CnAISDA) and questions whether it marks a genuine effort by China to address risks from advanced AI or is largely symbolic. It explains that while China has begun to adopt some global AI safety ideas, these efforts remain early, limited, and more focused on international image and cooperation than on practical domestic safeguards. The authors conclude that the real importance of CnAISDA will depend on whether it ultimately leads to concrete action to manage high-risk AI.
Open source agentic startup LangChain hits $1.25B valuation
LangChain, an open source framework for building AI agents, has raised $125 million at a $1.25 billion valuation. Started in 2022, the project became popular for helping developers work more easily with early large language models. As the company has grown, it has expanded into a full platform for building and managing AI agents.
▶️ AI Is Changing the Game in Astronomy (17:17)
In this video, Jon McColgan from Astrum discusses how AI is changing science, with a focus on physics and astronomy. McColgan explains that AI and machine learning are already being used by scientists to help them process the huge amounts of data produced by modern experiments. He also highlights the potential for large language models to search vast collections of research papers for insights and ideas, to improve and translate scientific writing, and even to suggest new concepts or experiments. However, McColgan cautions that we must use AI tools wisely if we want them to help us advance our understanding of the Universe.
Anthropic Has a Plan to Keep Its AI From Building a Nuclear Weapon. Will It Work?
Anthropic worked with US nuclear agencies to create a filter that stops its AI chatbot Claude from giving help on building nuclear weapons. Supporters say it’s a sensible safety step as AI becomes more powerful, while critics argue the model likely didn’t have access to sensitive nuclear information anyway and that the measure may be more symbolic than necessary. Anthropic hopes other AI companies will adopt similar safeguards.
🤖 Robotics
Amazon hopes to replace 600,000 US workers with robots, according to leaked documents
The New York Times reports that Amazon is planning to significantly expand its automation efforts, with internal documents suggesting it aims to automate around 75% of its operations and avoid hiring more than 600,000 US workers by 2033, potentially saving the retail giant billions of dollars. While Amazon denies that these plans reflect its broader hiring strategy, the company has considered ways to soften public concern about job losses, including using friendlier language such as “advanced technology” and “cobots.” Critics warn that if Amazon succeeds, it could shift from being a major job creator to a net job eliminator in the US.
▶️ World’s FASTEST Drone V3 (16:21)
This video follows Luke Maximo Bell on his journey to reclaim the world record for the fastest drone. He shows how he built Peregreen 3, from a fully 3D-printed frame to a water-cooling system, sharing the fails and crashes along the way—and eventually hitting 585 km/h (363 mph) to beat the previous record.
▶️ Unitree H2 Destiny Awakening (0:49)
In this video, Unitree shows what H2, its newest humanoid robot, can do, from dancing and fighting to catwalking. The robot’s movements are very smooth, natural and impressive.
DoorDash and Waymo launch autonomous delivery service and introduce limited-time Waymo promotion for DashPass members
DoorDash is partnering with Waymo to test driverless delivery in Metro Phoenix. DashPass members in Los Angeles, San Francisco, and Phoenix can also get $10 off one Waymo ride per month until the end of 2025. The delivery service will start with DashMart orders and may expand to more stores later. The collaboration is positioned as a step toward a more automated and convenient local commerce experience.
AI drones are America’s newest cops
Police departments across the US are increasingly using AI-powered drones for surveillance, investigations and emergency response. Drones are cheaper than helicopters and provide extensive real-time data, but their growing use has raised major privacy and legal concerns, including questions around warrantless surveillance and how collected data is stored and used.
Researchers from China present a research project in which they trained a humanoid robot to act as a goalkeeper. Although the robot’s abilities are still far from those of a human goalkeeper, the project represents a step towards enabling humanoid robots to perform dynamic, whole-body, real-time interactions with fast-moving objects.
🧬 Biotechnology
Why AI Companies Are Racing to Build a Virtual Human Cell
Scientists have long dreamed of “virtual cells”—a detailed computer model of a cell that would allow them to observe and test biological processes digitally. Advances in AI and large datasets are now making this idea more realistic. These tools could speed up drug discovery and deepen understanding of how cells work. However, challenges remain. A fully functional virtual cell is still likely a decade away, but researchers are increasingly hopeful.
Why the next generation of mRNA vaccines is set to be even better
Scientists have developed a new type of mRNA vaccine that creates virus-like nanoparticles in the body, leading to a much stronger immune response. In mice, it generated far more antibodies than current COVID-19 mRNA vaccines, which could allow for smaller doses with fewer side effects. This approach combines the fast, low-cost production of mRNA vaccines with the high effectiveness of nanoparticle designs, and researchers are now exploring its use for other viruses.
Metagenomi taps Amazon’s custom chips to develop gene-editing tools
Metagenomi, a biotech company, has begun using Amazon Web Services’ Inferentia AI chips to help develop new gene-editing tools. The company says these chips are cheaper than those from Nvidia and allow them to test a large number of proteins more cost-effectively. This helps them search more widely for potential treatments for genetic diseases.
💡Tangents
Humans extend forgiveness to machines just as they do to people, study reveals
New research suggests that users also forgive technology, similar to how they forgive people. The study found that when devices or apps fail, people may feel annoyed or disappointed, but they often choose to keep using them. Some do this because the technology is still useful overall, others blame human error rather than the machine, and some simply accept that technology is not perfect. People who are more confident with technology tend to forgive more easily, while those less comfortable may stop using it.
Thanks for reading. If you enjoyed this post, please click the ❤️ button or 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!"