OpenAI: The Contribution of 10,000 AI Agents to Navier–Stokes Remains Uncertain

Key Takeaways

1OpenAI mobilized around 10,000 model instances over 88 hours and 130 billion tokens

2Experts dispute the affiliation of the proof with the approach of Córdoba and Martínez-Zoroa

3Approximately 400 results were bulk-published on GitHub, irritating mathematicians

Why it matters : The actual contribution of AI agents and the claimed method remain uncertain, complicating evaluation by the community despite ongoing review by the Clay Mathematics Institute.

OpenAI has mobilized approximately 10,000 model instances, 88 hours of computation, and 130 billion tokens for a potential solution to the Navier–Stokes problem, but the contribution of this army of agents remains uncertain according to Noam Brown. Meanwhile, mathematicians are questioning the methodological affiliation claimed by the company and are frustrated by the scattered dissemination on GitHub, while the Clay Mathematics Institute continues its review.

The Contribution of the 10,000 Agents Remains Unmeasured, According to Noam Brown

OpenAI has mobilized around 10,000 model instances for nearly 88 hours and 130 billion tokens to develop its potential solution to the Navier–Stokes problem. According to Noam Brown, it is not actually established what this multitude of agents has concretely contributed to the outcome. Behavior at this scale has not been studied, as systematic comparative tests with thousands of agents would be too costly. No trial has measured the time it would take for a single agent to solve the same problem, and it remains nearly impossible to determine the precise advantage of 10,000 agents over 1,000. Noam Brown estimates that less than 10% of the success is attributable to the multi-agent configuration. He believes that these systems, still new and flashy, receive a disproportionate amount of credit, while the determining factor would rather be the underlying model, capable of handling long tasks and functioning in parallel.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Originality of the Contested Approach and Denial of Borrowing Accusations

OpenAI claims that its proof follows the approach of Diego Córdoba and Luis Martínez-Zoroa. Several groups of experts contest this lineage. For Gonzalo Cao-Labora (EPFL), OpenAI's argument differs in its details, and the similarities would only be superficial. Before the publication, Tristan Buckmaster (NYU) accused OpenAI of having used work conducted with Levent Alpöge (Anthropic) based on public models. OpenAI rejected this accusation, asserting that it never used Buckmaster's data. However, Camillo De Lellis (Institute for Advanced Study) believes that a definitive answer to this question is likely impossible, as language models remain black boxes.

Scattered Dissemination on GitHub and Ongoing Formal Review

At the beginning of October, OpenAI added nearly 400 new results regarding its research, complicating tracking for the mathematical community. The posting of a scattered set on GitHub has sparked discontent among many mathematicians, who feel that this way of sharing work undermines the value of their field. For its part, the Clay Mathematics Institute continues to analyze the proof since it was made public.

A Formalized Proof in Lean and a 166-Page Document

OpenAI's proof attempt is entirely written in Lean, a formal verification language that checks each logical step without producing a text directly readable by humans. The document spans 166 pages and was presented in early September.