OpenAI's 10,000 Agents Crack Navier-Stokes in 88 Hours, Then a Credit Storm Erupts
新智元 · wechat · 2026-09-09
OpenAI Claims a Millennium Problem
OpenAI announced that an internal model, far beyond GPT-6 Astra, produced a complete proof for the Navier-Stokes existence and smoothness problem, formally verified in Lean — the second Millennium Problem ever solved.
- The proof: A constructed vortex flow blows up in finite time (velocity → ∞) with bounded kinetic energy, matching Clay's options C/D. The 165-page paper is signed only "OpenAI".
- The process: 10,000 concurrent agents ran from Sept 1; 100 agents first solved the forced-free Euler blowup in 50 hours as a stepping stone, with the final breakthrough 88 hours in. Total: 2.7M messages, 130B output tokens, millions of dollars in compute; Lean verification took 17 hours.
- Credit storm: NYU's Tristan Buckmaster and Anthropic's Levent Alpöge had already solved related blowup problems (porous medium, Boussinesq, forced Euler) with LLM help, endorsed by Terence Tao. Buckmaster alleges OpenAI started its project only after hearing rumors of their work, and that Sebastien Bubeck pressured him to drop Levent (an Anthropic employee) from authorship, including a "why ruin your career" remark.
- Two sides: Bubeck admits the poor wording but calls it coordination, not poaching; Sam Altman backed the team; Mark Chen denied seeing their work, while conceding anonymized data leakage can't be fully ruled out.
- What's at stake: Clay's rules require journal publication plus two years of acceptance — a company-signed AI proof directly challenges academic norms and raises the question of whether a machine's correct proof equals a mathematician's.
Related event: OpenAI claims Navier-Stokes breakthrough amid authorship controversy(243 posts)→
More from Companies & People
- Lenny shares: the scarce PM skill today is elevating other people's ambition — lennysan · 2026-09-09
- a16z GP: agents can't 'make me a million dollars' — human direction still unlocks the next hill — lennysan · 2026-09-09
- How Cursor's Shipping Velocity Broke GitHub's Moat — and No One Owns SOTA Now — pzakin · 2026-09-09
- Alexandr Wang hypes Meta's Muse agent: hotter and faster than rivals — mon__lim · 2026-09-09
- Turn your team's project into an organization: four reasons formal structure wins — AaronBergman18 · 2026-09-09
- Researchers warn: don't feed commercially sensitive data to cloud LLMs as labs eye drug discovery — SumitGup · 2026-09-09