AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Mathematicians say OpenAI doesn't understand its own Navier-Stokes proof

Collected Oct 17, 2026

In early September, OpenAI presented a 166-page potential solution to the Navier-Stokes problem, one of the Clay Mathematics Institute's Millennium Prize problems. The company said the work followed the approach of Spanish mathematicians Diego Córdoba and Luis Martínez-Zoroa. Several groups of experts now dispute that characterization, according to reporting by Scientific American. Gonzalo Cao-Labora, a mathematician at EPFL, says that when you get into the details it is a different approach, and that the resemblance is only surface-level.

The proof is written in Lean, a formal verification language. Lean checks each logical step mechanically, which makes a result unusually auditable by machine, but the file itself does not read like a human argument. A formalization establishes that a proof term typechecks against the stated theorem; it does not by itself explain the idea behind the argument or make the strategy legible to a reader. That gap is central here: reviewers can confirm the steps compile while still disagreeing about what method the proof actually uses. The Clay Mathematics Institute has been reviewing the submission since it was released.

There is also a dispute over provenance. Before publication, NYU mathematician Tristan Buckmaster accused OpenAI of building on work that he and Anthropic's Levent Alpöge had done with publicly available models. OpenAI denied the claim, saying it never used Buckmaster's data. According to Camillo De Lellis of the Institute for Advanced Study, a definitive answer is probably impossible, because language models are black boxes. In practice that means the question of where the underlying ideas came from may never be settled to anyone's satisfaction, regardless of how the mathematics itself is judged.

In early October, OpenAI released roughly 400 additional results, published as a loose collection on GitHub rather than as a curated paper. The pace and format angered many mathematicians, who felt it cheapened their field. This is a familiar tension: formal verification and automated search can generate certified statements far faster than a research community can read, contextualize and attribute them. A Lean-checked theorem is verifiable, but verification is not the same as understanding or peer review, and volume can make the social side of mathematics harder rather than easier.

On the compute side, OpenAI deployed about 10,000 model instances for roughly 88 hours, consuming around 130 billion tokens, to develop the potential solution. According to OpenAI researcher Noam Brown, nobody actually knows what that large agent swarm contributed to the result. Behavior at that scale has not been studied, because running systematic comparison tests with thousands of agents is too expensive. No one has measured how long a single agent would take on the same problem, and it is nearly impossible to say what concrete advantage 10,000 agents offer over 1,000. Brown attributes less than 10 percent of the success to the multi-agent setup, arguing that such systems are new and flashy and therefore receive disproportionate credit. What matters most, in his account, is the base model, which can work on tasks over long periods and run in parallel.

That admission is the most useful signal for teams building agent systems. Multi-agent orchestration is often sold as the source of capability, yet here the person closest to the run estimates the swarm contributed a small minority of the outcome and says the contribution cannot be isolated with current methods. A likely trade-off is that scaling agent count multiplies cost linearly while its benefit stays unmeasured, so the burden of proof sits with whoever claims the orchestration is doing the work. The practical test is an ablation: run the same task with one agent and with the full swarm, hold the base model and token budget comparable, and report both cost and quality.

Why it matters: Research groups and product teams evaluating multi-agent architectures now have a prominent example where the people running the system cannot quantify what the extra agents bought. For developers, that argues for treating agent count as a cost knob to be justified by controlled comparison rather than a default. For mathematicians, the episode raises questions about how automated, machine-verified output should be reviewed and attributed, though how the field resolves that remains an open inference rather than a settled outcome.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

OpenAI submitted a 166-page solution to the Navier-Stokes problem but apparently doesn't fully understand its own proof. Several groups of experts dispute the claimed methodology, saying the actual approach differs from what OpenAI described. And because language models are black boxes, where the underlying ideas came from will likely never be clear. The article Mathematicians say OpenAI doesn't understand its own Navier-Stokes proof appeared first on The Decoder .