OpenAI's biggest math breakthrough is getting ugly...

Fireship recounts a disputed account of AI-assisted fluid-dynamics research and credit between mathematicians and OpenAI. This summary uses English auto-generated captions, not the papers, statements, or private call; names may be mistranscribed. Neither the claimed proofs nor allegations about training data and threats are independently verified here.

Key ideas

• Fireship introduces Navier–Stokes as equations describing fluid motion and frames the mathematical question as whether solutions can break down. His explanation is deliberately simplified and should not be treated as the precise statement or verification requirements of the Millennium Prize Problem. 00:00–03:01(opens in a new tab)
• He reports that Tristan Buckmaster and a collaborator working at Anthropic used coding agents while pursuing a related Euler-equation result. The video distinguishes their reported result from the full Navier–Stokes prize problem rather than treating the two as interchangeable. 01:59–03:01(opens in a new tab)
• Fireship then describes OpenAI's claimed solution through a large agent search and recounts Buckmaster's concern that the method overlapped with work submitted through Codex. A question about possible training on user sessions is not evidence that such training occurred; the report says OpenAI denied accessing user data and described the proofs as different. 01:59–04:53(opens in a new tab)
• The reported phone-call narratives conflict over proposed attribution arrangements and threats. Fireship presents Buckmaster's accusations alongside OpenAI's denial and Sam Altman's defense of its team. The transcript does not resolve the dispute or establish wrongdoing by either side. 03:01–04:53(opens in a new tab)
• Fireship attributes to Terence Tao a warning that agent swarms racing to act on research rumors could discourage mathematicians from sharing preliminary ideas. This is an argument about research incentives, distinct from whether any claimed proof is valid. 04:53–05:44(opens in a new tab)

Practical implications

Editorial: Keep proof verification, research credit, confidentiality, and training-data policy as separate questions. An agent-assisted result does not settle how its inputs were obtained or whether its mathematical claims withstand expert scrutiny.

Questions and limits

Editorial: The report supplies no independent proof review and leaves the data-use allegation unresolved. The fictional curse passage is satire. Railway sponsors the closing hosting recommendation; it is not evidence that Railway resolves the confidentiality issues discussed. 04:03–05:44(opens in a new tab)

Full video: https://www.youtube.com/watch?v=aspmNhKAFMc(opens in a new tab)