Meta launched Muse, a personal AI agent that wants access to nearly everything on your phone. Muse is Meta's bid for the personal-agent category, designed to handle tasks like selling a car, booking flights, and managing calendars by connecting to email, payments, and health services. The Verge and Wired both covered the launch; Wired noted it was built with privacy controls, while TechCrunch framed the central question as whether users still trust Meta with that level of access. Meta positions Muse against other personal agent products already on the market. The trust question is not rhetorical - Meta's data practices have been litigated and regulated for years. (TechCrunch)
Brief
Wednesday, September 9, 2026 · 5 min read
OpenAI Claims Navier-Stokes Breakthrough, Academics Cry Foul
Good morning. OpenAI solved a 90-year-old math problem and somehow made enemies doing it. Meta launched a personal agent that needs your email, calendar, and payments, which is either very useful or a very familiar ask from a company with a complicated relationship to user data. And somewhere in the middle of all this, hackers figured out that stealing AI tokens is a real business now.
Today's reading time is 5 minutes.
INDUSTRY
The Navier-Stokes existence and smoothness problem has sat unsolved since the 1930s, carries a $1 million Clay Mathematics Institute prize, and is the kind of result that makes careers - which is exactly why the math community is not taking OpenAI's word for it.
Driving the news: OpenAI announced Tuesday that its agents had solved the Navier-Stokes problem, a set of equations describing how fluids flow that mathematicians have failed to crack for roughly 90 years. The Clay Mathematics Institute offers $1 million for a verified solution. Within hours, academics - including at least one NYU mathematician quoted by TechCrunch - accused OpenAI of procedural misconduct in how it handled and disclosed the result. MIT Technology Review noted the announcement has become mired in controversy before any independent verification has been completed.
- The Navier-Stokes problem is one of seven Millennium Prize Problems; only one other has ever been solved.
- OpenAI published a blog post claiming the result but has not yet submitted it through the standard peer-review process that would trigger the prize.
Zoom in: AI labs have been pushing into formal mathematics for a couple of years, mostly through proof assistants and reinforcement learning on competition problems. Solving an olympiad problem is one thing; a Millennium Prize Problem is a different category entirely, the kind of result that would normally go through months of peer scrutiny before any announcement. OpenAI skipped that queue and announced first, which is the core of the complaint.
- The Clay Mathematics Institute requires independent expert verification before awarding the prize - that process has not started.
- Wired reported the announcement has drawn accusations of impropriety from multiple academics, not just one.
Why it matters: For people building with AI, the immediate practical effect is zero - this does not ship as a feature. What it does affect is how much credibility AI labs get when they make large scientific claims going forward. If the result holds up under review, it changes what researchers trust AI agents to do in technical workflows. If it does not, every future 'AI solves X' headline gets a longer discount applied.
Bottom line: OpenAI may have solved one of the hardest problems in mathematics, or it may have just made the hardest problem in mathematics into a PR problem - the math community is not ready to say which.
TechCrunch ↗Get this every weekday
Two minutes, 7am ET. No fluff.
Also happening
Hackers are draining Claude subscribers' token balances without their knowledge. Anthropic has issued a warning to users after a Claude subscriber noticed last month that his account was burning tokens while he was not working. TechCrunch reports that hackers have found a way to steal API tokens from paying subscribers, effectively using their compute budget for free. Anthropic has not yet detailed the attack vector publicly. This lands on the same week Anthropic is also facing a class-action lawsuit from power users who say the company misled them about what top-tier subscriptions actually deliver. (TechCrunch)
ChatGPT Images 2.5 shipped with a Sketch feature that turns doodles into finished images. OpenAI released ChatGPT Images 2.5 on Tuesday, adding a drawing input called Sketch that lets users scribble directly inside the interface and have the model interpret the shape as a generation prompt. The Verge tested it with a rough cat drawing. The update also improves how the model handles reference photos and produces more personalized output. No separate subscription tier was announced; it appears to be rolling out to existing ChatGPT users. (The Verge)
A senior Anthropic safety researcher said AI has more than a 10 percent chance of killing all humans by 2030. The statement came hours after a colleague resigned, citing concerns that Anthropic and rival labs are racing to build superhuman systems they cannot control. The Verge reported both events together. A 10 percent estimate from inside one of the labs most focused on safety is a different kind of signal than the same number from an outside critic. Anthropic has not issued a formal response to either the resignation or the probability estimate. (The Verge)
The thread
Trust is the bill coming due
Three of today's stories are really the same story told in different rooms. Anthropic has subscribers suing over subscription promises and hackers draining their token balances. Meta is asking users to hand over email, payments, and health data to an agent from a company that has spent years in regulatory hearings about exactly that kind of access. OpenAI announced a historic math result and immediately faced accusations of cutting corners to get there first. None of these are capability failures - they are credibility failures, and they arrive at the moment when AI tools are asking for more access and more trust than ever before.
Try it yourself
Image Generator
ChatGPT Images 2.5 now accepts hand-drawn sketches as prompts inside its own interface; Snippt's image-generator lets you do the same thing from a text prompt across several models without a ChatGPT subscription.
On our radar
- Google DeepMind released AlphaGenome Atlas, a tool for analyzing the human genome that the company says could accelerate research into disease treatments.
- OpenAI opened a $5 million grant program funding independent research on how generative AI affects teen development and well-being.
- Anthropic is facing a class-action lawsuit from power users who say the company misled them about what top-tier subscription plans actually provide.
- DeepSeek Flash 4.1 is in internal beta via API, featuring a new architecture with native multimodal support and lower costs than the current version.
- A community build of Qwen3.8-Flash-Next running on Apple Silicon via MLX-serve now supports 1 million token context using 8-bit KV cache on an M5 Max with 128GB RAM.
- An MIT researcher published a case study showing GPT-5.6 Sol with Codex autonomously running quantum computing experiments, analyzing results, and calibrating qubits.
Get the brief in your inbox
Every weekday morning. Two minutes, no fluff.