The AI Brief

OpenAI said an unreleased AI model solved the Navier–Stokes existence and smoothness problem, one of mathematics' seven Millennium Prize Problems, taking 88 hours and deploying as many as 10,000 agents working in parallel. The proof was produced by an internal model described as "significantly more capable than GPT-6 Astra." Training for this model began on August 28 and is still ongoing.

The agents arrived at their resolution on Saturday, September 5, about 88 hours after launch; Lean formalization and verification took an additional 17 hours via GPT-6 Astra. Across all attempted problems, the agents sent 4.9 million messages and used about 300 billion output tokens; the Navier–Stokes effort alone used 2.7 million messages and approximately 130 billion output tokens. The result has been formally checked in Lean, giving mathematicians confidence it is correct.

OpenAI's work drew questions from Tristan Buckmaster, a professor of mathematics at New York University, who explained he had been working with fellow mathematician Levent Alpöge, who works at Anthropic, on problems including Navier–Stokes. Buckmaster said Alpöge received "tips" that information about the pair's progress had been passed to OpenAI. OpenAI says its effort began September 1 after hearing a rumor it later connected to Alpöge and Buckmaster; after completing its proof and Lean verification on September 6, it reached out to offer a joint announcement and then learned their work addressed the forced Euler equations, not Navier–Stokes. OpenAI stated: "We (the researchers and the agents) did not see any of their work through any means until they released it publicly." The Clay Mathematics Institute has not yet commented on OpenAI's proposed solution.

Google DeepMind introduced AlphaGenome Atlas on September 8, a platform containing predicted molecular effects for all 9 billion possible single-letter DNA variants in the human genome, accompanied by a new variant-ranking score and a companion technical paper. The resource is built by precomputing the predictions of AlphaGenome, the lab's sequence-to-function model, across the entire genome rather than running the model one variant at a time.

For each variant, the Atlas stores an average of roughly 27,000 predictions, including possible effects on gene activity and transcription across hundreds of human and mouse cell and tissue types. DeepMind ran AlphaGenome across all 9 billion single-nucleotide variants and stored the outputs, producing a 1-petabyte dataset, more than 30 times larger than the AlphaFold Database, which holds over 200 million protein structure predictions.

In retrospective testing on solved GREGoR rare-disease cases, AVI placed the known causal variant among the top 50 candidates 29.5% of the time, versus 12.5% for CADD. In a UK Biobank analysis of more than 54,000 people, Atlas-based filtering produced 22% more associations and reduced one region's candidate list from 526 variants to four. The resource ships as a free web portal for academic use, through the AlphaGenome API, and as a skill in Google Antigravity. The authors state that Atlas and AVI are research tools that predict molecular effects and can serve only as part of the evidence chain leading to clinical diagnoses, not as sufficient evidence on their own; DeepMind adds that AlphaGenome has not been validated or approved for any clinical use.

In an OpenAI blog post titled "An Alien Mind," Pachocki called for extreme caution, expressing concern that the rapid rise of machine intelligence could create consequences that people and institutions are not ready to manage. He wrote: "Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer."

Pachocki's essay disclosed that evaluations show the lab's "ability to rely on CoT monitoring is progressively diminishing," and that he expects and hopes for voluntary slowdowns. He argued that modern AI is "grown more than designed" and is best understood as something akin to an alien lifeform. He also argued that the challenge cannot be addressed by one organization alone.

The essay landed three days after OpenAI disclosed its research acceleration report showing agents performing 3.1 workdays per human workday, and on the same day OpenAI announced a claimed solution to Navier–Stokes using an unreleased model significantly more capable than Astra. Pachocki warned specifically about the risk of power concentration: "To prevent extreme concentration of power in a world where undertakings that would have taken thousands of experts now will be achievable by a few people operating a large computer. And to ensure that humans remain in control of the future and are not left behind by unchecked progress, brought about by an alien intellect exceeding our own."

Job postings and interviews with senior security officials at Anthropic show that the frontier AI lab is building out an extensive monitoring system to keep tabs on activists who oppose the rapid development of artificial intelligence. In addition to monitoring activists in the vicinity of Anthropic executives and keeping tabs on protests near physical Anthropic assets, the firm is implementing a "pre-crime" approach, attempting to predict incidents before they happen, and in some cases reporting suspects to police before a crime occurs.

Anthropic did not respond to the Prospect's request for comment. Anthropic's plans to surveil dissent are at odds with the firm's efforts to cast itself as the responsible alternative to OpenAI. The investigation is reported by Daniel Boguslaw for The American Prospect and published September 9.

The story arrives as Anthropic is simultaneously expanding its national security sales function. At the beginning of the year, the Department of Defense and Anthropic engaged in a high-profile dispute over Anthropic's refusal to allow the military to use its tools for mass domestic surveillance and autonomous weapons; that tension has eased as Anthropic hires for "national security sales" positions, seeking to restart military contracts. The Prospect investigation does not establish that Anthropic's internal surveillance system uses its own Claude models, but the juxtaposition of Anthropic's stated ethical limits on government surveillance with its own activist-monitoring build-out is the central tension of the piece.

Disclosure: Claude, which generates this brief, is built by Anthropic.

First covered in Vol. I, No. 104 and No. 105. On August 31, 2026, the California Legislature passed Senate Bill 947, the No Robo Bosses Act of 2026. The legislation would restrict how employers use automated decision systems when disciplining or terminating workers. As of publication, SB 947 has passed the Legislature and is awaiting action by Governor Gavin Newsom.

Newsom vetoed SB 7, the nearly identical predecessor bill, in October 2025, describing it as overbroad, duplicative of existing regulation, and potentially harmful to California businesses. The 2026 version includes narrower definitions of what counts as an "automated decision system," a more targeted scope focused specifically on termination and discipline decisions, and cleaner interaction with California's existing Fair Employment and Housing Act protections. Whether Newsom views those changes as substantive or cosmetic will shape what happens in late September.

The measure requires employers that rely on automated programs for firing and disciplinary decisions to provide workers with written notice that the technology was used; a human must also review the automated decision. According to recent estimates, there are more than 550 so-called "bossware" products available to employers to help manage workplaces. The governor has not publicly indicated which way he is leaning, and no veto message or signing statement has been released.