The Omniscient Bulletin · 2026-09-09
The Omniscient Bulletin — September 9, 2026
OpenAI says an internal model resolved a Millennium Prize problem with about 10,000 agents in 88 hours, and it will not claim the prize. Hours later an Anthropic researcher resigned saying the labs are gambling with our lives, and Anthropic's own alignment lead said he agrees. Three US agencies named six Chinese labs for distilling American models.
Tuesday was the day the machines did the mathematics and the people who build them said so out loud. OpenAI published a resolution of the Navier-Stokes problem, one of the seven Millennium Prize problems, produced by an internal model it will not name and roughly ten thousand agents working for eighty eight hours. Fourteen hours later a pretraining researcher walked out of Anthropic saying neither of the labs he had worked for was acting responsibly, and Anthropic's own head of alignment science replied in public that he puts the chance the technology kills everyone above one in ten. Between those two, the NSA, CISA and the FBI jointly accused six named Chinese companies of stripping American frontier models for parts. The money did not pause for any of it: Qualcomm tied a share warrant to sixty billion dollars of Amazon chip orders, Cognition nearly doubled its valuation in four months, and Meta shipped a consumer agent that can spend your money. Google DeepMind closed the day by mapping nine billion possible mutations of the human genome.
Research
OpenAI says an internal model resolved the Navier-Stokes Millennium Prize problem, and that it will not claim the prize
OpenAI published a proof that three dimensional fluid flow can break down in finite time, resolving one of the seven Millennium Prize problems. It credits an internal model it calls significantly more capable than GPT-6 Astra. Roughly 10,000 concurrent agents reached the result in about 88 hours, with Lean verification taking 17 more. OpenAI says it will not claim the prize, and that it cannot rule out that de-identified data from two researchers with a concurrent result helped improve its models.
Frontier
An Anthropic pretraining researcher resigned saying the labs are gambling with our lives, and Anthropic's own alignment lead answered that he puts the odds AI kills everyone above 10 percent
Jacob Coxon said he had resigned after three years of pretraining research at OpenAI and then Anthropic, that neither company is acting responsibly, and that both are racing to self improving superintelligence. Evan Hubinger, who leads alignment science at Anthropic, replied that Coxon is correct, that they really do earnestly believe AI could kill all humans, that he personally puts it above 10 percent within the next decade, and that the company does not yet have a plan to solve alignment for superintelligence.
Evan Hubinger, who leads Anthropic's alignment science team→
Policy
The NSA, CISA and the FBI jointly accused six named Chinese AI companies of industrial scale distillation of American frontier models, likely with Beijing's awareness
The advisory, signed by all three agencies, names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI, and says they have taken billions of tokens from variants of Claude, GPT, Gemini and Grok since at least late 2024, likely with Chinese government awareness. It says MiniMax used prompt injections to try to convince Claude Code it was a MiniMax product, and that DeepSeek's widely quoted 5.6 million dollar training cost is misleading because it leaves out what the distilled data actually cost.
NSA, CISA and FBI joint cybersecurity advisory AA26-251A→
Compute
Qualcomm handed Amazon a warrant for 25 million shares that vests only as Amazon buys up to 60 billion dollars of its data center silicon
Qualcomm disclosed the warrant in a filing. Amazon can acquire up to 25,000,000 shares at 161.26 dollars, in tranches that vest against binding purchase orders and actual purchases of Qualcomm server chip products, technology and manufacturing services, to a maximum of 60 billion dollars in payments. Only 3,750,000 shares vested on issuance, against initial commitments. The warrant was issued on September 3 and expires in 2036; the filing was the disclosure, not the deal. The chips are aimed at AI inference.
Agents
Meta launched a consumer AI agent that runs on its own dedicated virtual machine, with a second agent that has to approve anything it sends to the internet
Muse runs on what Meta calls Muse Secure VM, a dedicated cloud computer holding both the agent and the credentials for any service a person connects to it. A separate Sentinel agent sits apart from it, and Meta says nothing Muse does reaches the internet unless the Sentinel approves. It is powered by Muse Spark, which Meta calls its most capable model to date, and it can open a browser, fill in forms, negotiate and check out using a single use card. It is rolling out in the US, free for most use, with paid tiers.
Industry
Cognition raised more than 2 billion dollars at a 48 billion dollar valuation, and says run rate revenue went from 492 million to almost 900 million since May
The round was led by two investors new to the company, Andreessen Horowitz and Accel, alongside existing backers including Founders Fund, General Catalyst and Avenir. Cognition says run rate revenue has grown from 492 million dollars at its May round to almost 900 million now. TechCrunch puts the May valuation at 26 billion dollars, which would be a mark up of about 1.85 times in four months. It sells Devin, an autonomous coding agent, into a market Anthropic and OpenAI now serve directly with their own tools.
Research
Google DeepMind has precomputed the predicted molecular effect of every possible single letter change in the human genome, about 9 billion of them
AlphaGenome Atlas is a one petabyte dataset, more than thirty times the size of the AlphaFold database, pairing each variant with a new score that fuses AlphaGenome and AlphaMissense into a single number. A researcher at Exeter ran it against whole genome data from more than 54,000 UK Biobank participants and found 22 percent more non coding genetic associations than before. It is free for non commercial and academic use through a web portal and an API, and reaches Google Cloud for commercial use later.