Ten thousand agents claim a million-dollar proof; Meta ships Muse
Edition of 2026-09-09
# AI For All ### Wednesday, September 9, 2026
As of today, a professional who does confidential work inside an AI coding tool must deal with a documented new risk: OpenAI admits it launched a ten-thousand-agent race at a famous open problem after hearing rumors of a rival team's progress, and that rival had been drafting inside OpenAI's own products for almost a year.
The rumor that launched ten thousand agents
For more than a century, mathematicians have asked whether the equations that describe fluids can break down: whether a solution, left to itself, can explode in a finite time. The Clay Mathematics Institute attached a prize of one million dollars to the question in 2000. On Tuesday morning, OpenAI said the answer is yes, and that ten thousand autonomous agents, running on an internal model, found the explosion. The proof was checked in Lean, a language machines use to verify mathematics, and OpenAI says the agents did the work in eighty-eight hours.
The announcement came within hours of two mathematicians publishing their own results on closely related problems. Tristan Buckmaster of NYU and Levent Alpoge of Anthropic had worked for almost a year, drafting inside OpenAI's own coding tools, and they say OpenAI's first prompt came only after word of their progress leaked. OpenAI admits the rumors inspired its effort, concedes the Euler result to the pair, and claims the Navier-Stokes result for itself. Buckmaster's public statement suggests OpenAI may have benefited from work done inside its own products.
The math may well be right. That is not the part that should worry you. The part that should worry you is the sequence: a rumor, a swarm, a proof, a dispute, all inside a week, all built on the private drafts of a rival team that paid OpenAI for its own tools. Anyone feeding confidential work into an AI assistant now has a documented reason to ask where it ends up. Quanta reports the two sides' accounts conflict, and sorting out the timeline will take time.
Quanta Magazine · Tristan Buckmaster's statement
Meta's new agent asks for your chores, not your passwords
Meta launched Muse on Tuesday: a personal agent that books travel, fills forms, negotiates, and keeps working after you close the app. It asks before sensitive actions, like sending an email or buying something. Meta's pitch is that Muse lives in its own secure machine, that a separate agent called Sentinel controls what reaches the internet, and that Muse never sees the passwords it uses.
The offer is free for most needs, with subscriptions at twenty or one hundred dollars a month depending on usage. For now it is available to American consumers only, on iOS, Android and the web, and inside WhatsApp. Meta is betting that people will pay an agent to run errands, at a moment when trust in agents is low and Meta itself just settled with a coalition of state attorneys general.
The architecture is genuinely more careful than most: a dedicated machine, an audit trail, an opt-out from training. Hold the applause. The version where even Meta cannot read your conversations arrives later this year, and the shopping habits Muse learns are a business model waiting to happen. If you are American and time-poor, this is available today. If you are cautious, the free tier costs nothing but your data habits.
The silent ninety-eight percent of your DNA now has a map
Your genome is a long text, and for most of it, scientists could not tell you what a single-letter change would do. That is not laziness. Only about two percent of the genome codes for proteins, the part biology understood best. The rest, the part that orchestrates when and where genes switch on, was mostly guesswork.
Google DeepMind published AlphaGenome Atlas on Tuesday: predicted effects for nine billion single-letter changes, every one physically possible, including variants no patient has ever carried. The catalogue is thirty times larger than the AlphaFold database, and it is free today through a web portal. At the Broad Institute, a team already used it to find the variant behind an unsolved epilepsy-linked condition, and lab screens confirmed the prediction.
For a clinical geneticist chasing a rare disease, this turns a search measured in years into a lookup. These are model predictions, not lab results, and the confirmation step still belongs to humans. But the direction is unmistakable: the genome's dark matter just got a map, and the map is public.
From napkin sketch to finished image
OpenAI upgraded the image generator inside ChatGPT, and the most useful change is the least flashy. You can now draw a rough sketch directly in the chat, and the model turns it into a finished image, keeping the shape and composition you drew. You can also comment on a generated image to change one element, a chair for a bench, without rebuilding the whole scene.
OpenAI says generation is up to fifty percent faster than the previous version, and the rollout started today across every tier of ChatGPT, free included, on desktop, mobile and web. Two companion models arrived in the API for developers.
Someone paying for ChatGPT can now go from a napkin idea to a usable visual in one conversation. The caveat is the same one OpenAI prints in its own safety card: sharper realism means more convincing fakes, which is why these images now carry Google DeepMind's invisible SynthID watermark in addition to standard metadata. Faster tools, visible receipts.
Weak signal
Terence Tao, the mathematician, watched the Navier-Stokes race and wrote that even the rumor of someone working on a problem can now trigger a massive amount of AI-powered effort to solve it first. The incentive, he warns, may soon point to not sharing promising research at all, which would reverse centuries of open science. Watch whether mathematicians stop talking about work in progress before it is done. The seminar room may be the first casualty.
Terence Tao, writing on September ninth
The next time a rumor of a proof spreads, the question will not be whether the AI can solve it, but whether anyone still dares to share the rumor in the first place.