Particle.news

OpenAI Says Internal AI Solved Navier–Stokes and Faces Credit Dispute

OpenAI published a Lean‑verified proof produced with heavy compute using thousands of agents, prompting questions about priority, data use, peer review.

Overview

  • OpenAI announced on Tuesday that an unreleased internal model produced a Lean‑formalized proof showing a finite‑time singularity for the three‑dimensional Navier–Stokes equations after roughly 88 hours and using up to 10,000 concurrent AI agents.
  • Two independent researchers, Tristan Buckmaster and Levent Alpöge, had posted related AI‑assisted results the night before and say they worked for months using Codex and other models, alleging OpenAI learned of their progress before public release.
  • OpenAI denies directly accessing the researchers’ private drafts and says its team and agents did not see those materials, but the company concedes it cannot completely rule out incidental improvements from de‑identified user data.
  • The Clay Mathematics Institute and the mathematics community will require peer‑review, two years of community validation, and formal scrutiny before any claim can qualify for the Millennium Prize or be accepted as settled.
  • Beyond the technical claim, the episode highlights new tensions over how large compute pools and model training practices reshape incentives for sharing unfinished work, assigning credit, and verifying machine‑produced mathematics.