OpenAI’s Navier-Stokes claim, Meta Muse, and Anthropic’s ITI exit

In this Brief

• OpenAI Navier-Stokes claim + credit fight; proof unpublished
• Meta Muse: Secure VM, Sentinel, $20 / $100 tiers
• Anthropic severs ITI over OVERWATCH / Chip Security / MATCH
• DeepMind AlphaGenome Atlas ~9B single-letter DNA predictions
• Cognition Series E >$2B at $48B; ~53× run-rate holds

OpenAI's Navier-Stokes claim

Setup: Emma Roth at The Verge writes that OpenAI's Tuesday blog claimed a solution to the Navier-Stokes problem, the equations that describe how liquids and gases flow, using an internal model more powerful than newly released GPT-6 Astra, alongside about 10,000 concurrent agents. Navier-Stokes is one of seven Millennium Prize Problems, each with a $1 million reward. Training of that internal model started August 28, per the company.

Madison Mills at Axios adds that the effort began Sep 1 after researchers heard rumors two Millennium problems had been solved, that the cost sat in the "millions of dollars," and that the claimed result is a singularity in finite time, math saying a smooth 3D flow can become infinite in finite time. The Next Web notes OpenAI described the result on a press call and had not published the proof as of Tuesday.

One day before OpenAI's announcement, NYU mathematics professor Tristan Buckmaster published related findings with Levent Alpöge, an Anthropic researcher, including Lean formalisations, machine-checkable proofs, on related problems. Buckmaster says they had been working with OpenAI's Codex and Anthropic's Claude as research notebooks, contacted OpenAI after learning the company had heard about their progress, and asked whether the model had trained on or accessed those sessions.

OpenAI said no specific user data was accessed, and that while unlikely it cannot rule out de-identified data from product use. Contested authorship and career-pressure allegations stay attributed and unsettled. OpenAI says it does not plan to take the $1m prize.

Take: A mathematical claim is not a result until someone can check it. Related Lean proofs are public. OpenAI's file is not. Who gets to use the lab's tools without competing against the lab is the fight underneath.

Longer cut: OpenAI claimed a Navier-Stokes Millennium breakthrough before publishing a checkable proof.

Live X: live Aaron_Harme X (Tue evening; status URL not on disk).

Meta Muse

Setup: Sarah Perez at TechCrunch writes that Meta launched Muse in the US, an always-on assistant that can send email, book travel, fill forms, and check out with payment rails. It connects to email, calendars, payments, health and fitness, smart home, dining, shopping, music, and events. Available at muse.ai, on iOS and Android, and in WhatsApp; AI glasses "soon." Free tier plus paid Power $20/month and Maximum $100/month; a payment card is required to start.

It runs in a dedicated Muse Secure VM, a private cloud computer spun up for one user, with a separate Sentinel agent on the same VM that can block or require approval before Muse reaches the internet or connected apps.

Purchases go through Link by Stripe with purchase protections; Shopify Shop Pay and 1Password are "coming soon."

Meta says Muse has no visibility into passwords or payment methods and that conversations are not shared with Meta ads systems. Those are company claims until independent review. Launch sits less than two weeks after Meta agreed to a massive $18 billion multistate settlement over social media consumer harms.

Ina Fried at Axios and Robert Hart at The Verge put the product under chief AI officer Alexandr Wang, note proactive and long-running tasks, and place Muse in a crowded field next to ChatGPT Work, Claude Cowork, Copilot Tasks, and others. "World's first personal AI agent built for everyone" is company language.

Take: Agent access is the product. Trust is the ask. A social feed never needed your card and your calendars in one box.

Live X: live Aaron_Harme X (Tue evening; status URL not on disk).

Anthropic severs ITI

Setup: Maria Curi at Axios reports that Anthropic is severing ties with the Information Technology Industry Council (ITI), the industry's main chip-and-tech lobby shop, over legislation that would curb foreign access to U.S. chips.

Last week ITI sent a letter to Senate and House Armed Services Committees urging lawmakers to strip three export-control bills from the annual defense policy bill (NDAA): the AI OVERWATCH Act, the Chip Security Act, and the MATCH Act. All three advanced out of House Foreign Affairs with substantial bipartisan support and are in the Senate NDAA managers' package after White House technical assistance. Anthropic supports all three and is ending ITI membership over the group's efforts to block them.

Per Axios: AI OVERWATCH would make Commerce notify Congress and allow a review period before export licenses for sensitive dual-use chips, tech with civilian and military uses, to adversarial nations (review period shortened from two years to 18 months). MATCH would push U.S. allies to restrict chip-manufacturing equipment sales to China. The Chip Security Act would require companies to track AI chips to counter smuggling.

ITI argues the bills would hurt U.S. dominance in global markets.

ITI members include Google, OpenAI, and Nvidia. Senate returns from recess Sep 14.

Take: Same industry. Different vote on the export wall. Anthropic is choosing the control side while peers stay in the club.

Live X: live Aaron_Harme X (Tue evening; status URL not on disk).

DeepMind AlphaGenome Atlas

Setup: Robert Hart at The Verge writes that Google DeepMind unveiled AlphaGenome Atlas, a predictive map of every possible DNA letter change in the human genome. DNA's alphabet is A/C/G/T; the human genome has about 3 billion letter pairs; that yields roughly 9 billion potential single-letter substitutions, one-base flips. Atlas predicts molecular effects, such as how much of a protein is produced, including in noncoding stretches that can turn genes up or down.

Access: web portal; a skill in Antigravity, Google's product for AI agents; the AlphaGenome interface. A Variant Impact Score (AVI) ranks variants using other Google models.

The dataset is about 1 petabyte, roughly a million gigabytes. Available to researchers for noncommercial use via the website; commercial use on Google Cloud "soon." Builds on AlphaGenome (model released last year) and AlphaMissense.

Ziga Avsec, DeepMind genomics lead, said the underlying model was already out and that precomputing and analyzing this many variants took time because the space is so big.

Training used public human and mouse genome databases.

Disease-treatment language in company framing is forward-looking; do not claim Atlas cures anything.

Take: Nine billion flips, scored in advance. A research catalog, not a clinic. The ship is the map.

Longer cut: DeepMind published predicted effects for about nine billion single-letter DNA changes.

Live X: live Aaron_Harme X (Tue noon; status URL not on disk).

Cognition's flat multiple

Setup: Marina Temkin at TechCrunch writes that Cognition, the company behind the coding agent Devin, raised $2 billion at a $48 billion valuation, about four months after a prior fundraise at $26 billion. Round led by Andreessen Horowitz, Accel, Founders Fund, General Catalyst, and Avenir.

Since the May announcement, annualized run-rate revenue, recent sales stretched into a yearly number, moved from $492 million to $900 million on company figures. The Next Web puts both rounds near 53× run-rate revenue, a flat multiple: investors paid about the same sky-high price per dollar of revenue for a bigger book.

NVIDIA joined as an investor and is named as a Devin customer in chip design. Named customers elsewhere include Mercedes-Benz, NASA, Goldman Sachs, and Citi. Burn and year-end ARR, annual recurring revenue, forecasts in coverage stay attributed to secondary reporting (The Information via TechCrunch).

Take: Bigger book. Flat multiple. Coding agents still priced as a category race, not a settled winner-take-all.

Live X: live Aaron_Harme X (Tue evening; status URL not on disk).

Also noted

The Next Web writes that Mistral raised €3bn Series D at more than €21bn, led by Samsung with the Commission-backed Scaleup Europe Fund at the front of the check. Longer cut: Mistral raised €3 billion with Samsung and a Commission-backed fund at the front of the check. Live X: live Aaron_Harme X (Tue morning; status URL not on disk).

The Next Web summarizing Reuters writes that Google rolled out EU Search changes under the Digital Markets Act after a €460m self-preferencing fine, and that an unnamed Google official called it the largest quality-of-service drop in Search's 29-year history. Live X: live Aaron_Harme X (Tue noon; status URL not on disk).

ASML and TSMC announced an industry push to 12-inch EUV photomasks for High NA, with a pilot line target by 2031 and system readiness by 2033; Reuters notes Samsung aiming High NA DRAM volume by 2028. Live X: live Aaron_Harme X (Tue noon; status URL not on disk).

HotHardware writes that Qualcomm and Amazon will co-design custom AI inference chips, silicon that runs finished models, with a warrant for up to 25 million Qualcomm shares that fully unlocks only if Amazon spends up to $60 billion through 2036. That $60B is a ceiling, not a firm order. Live X: live Aaron_Harme X (Tue noon; status URL not on disk).

The Next Web summarizing Bloomberg writes that Anthropic walked from ~$6bn Decart talks after diligence; the interest was reportedly Decart's chip-optimisation stack, not world-model demos. Live X: live Aaron_Harme X (Tue morning; status URL not on disk).

Sam Biddle at The Intercept published FOIA paperwork on DoD contracts with OpenAI, Anthropic, Google, and xAI, each worth up to $200 million, to prototype tools for military advantage and decision-making. Live X: live Aaron_Harme X (Tue morning; status URL not on disk).

Josh Taylor at The Guardian writes that Australia's exposure draft "my feed, my way" would force platforms to keep offering algorithm opt-out, with fines of more than A$100m. Live X: live Aaron_Harme X (Tue morning; status URL not on disk).

The Tuesday Brief covered Cognizant's $4.5tn wage bill next to 1,500 US grads, Liquid's ~$320M federation pause, Lake Mariner's four logos, DeepMind's math swarm, and Clifford quitting Aria.