Experts say OpenAI’s Navier–Stokes proof took a different route

Experts examining OpenAI’s AI-generated Navier–Stokes proof say its approach differs from the method developed by two Spanish mathematicians. That challenges the early impression that OpenAI’s result followed their work, though it does not settle every question about how the AI reached its answer.

OpenAI Watch analysis

What happened

OpenAI published a 166-page solution to the Navier–Stokes problem in September. Early accounts, including statements from the company, suggested it used an approach associated with Spanish mathematicians Diego Córdoba and Luis Martínez-Zoroa. That perceived similarity helped fuel a dispute with mathematicians Tristan Buckmaster and Levent Alpöge, who had been pursuing related work with the help of language models.

Scientific American reports that several groups of experts working through the proof now consider the resemblance between the two approaches superficial. The article quotes mathematician Gonzalo Cao-Labora saying that, in the details, the approach is different. Both methods exploit a loophole in the formal problem by constructing an unusual force that makes the equations produce a blowup, but OpenAI’s proof reportedly constructs that force differently.

Why it matters

This is a meaningful correction to the story around the proof, not a verdict on its ultimate significance. The report says OpenAI’s initial, smaller group of AI agents may have followed the Spanish approach while working on an intermediate result; a later swarm of about 10,000 agents apparently took another route to finish the proof. The distinction matters when assessing both the work’s novelty and the earlier claims of resemblance.

The proof was written in Lean, which can check that formalised logical steps follow, but that does not make the result easy for people to understand. Scientific American says researchers are still using the Lean code and other language models to build a clearer account of the mathematics. More than a month after the announcement, experts still do not fully understand the proof or the model’s path to it.

Our read

The strongest new point is that experts now see a substantive difference between the completed proof and the Spanish program, even if an earlier intermediate result may have drawn on that approach. That narrows one part of the controversy; it does not reveal exactly how the AI system reasoned, or resolve every question about what influenced it. Scientific American reports that experts do not expect the model’s precise chain of reasoning to be recoverable.

For readers following AI-assisted mathematics, this is a useful reminder that checking a proof and explaining how it was found are separate jobs. Lean can help with the first. The second still requires mathematicians, time and, apparently, a healthy tolerance for 166 pages.

What to watch

• Whether researchers publish a clearer account of the proof’s mathematical route.
• Whether further analysis distinguishes the intermediate result from the final proof.
• How OpenAI describes the work’s origins as expert scrutiny continues.

Discussion spark: If experts can verify an AI-assisted proof but cannot reconstruct how the system found it, is that enough to establish the result’s scientific value?

OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.