OpenAI’s Navier–Stokes Proof Claim, Research Credit, and the Future of AI

Watch a ring of moving air stretch and break into smaller curls: it can look random, but the motion follows physical laws. Understanding fluids supports technologies from aircraft to cooling systems. The mathematical question I want to discuss asks what the equations permit as fluid motion becomes increasingly complex.

The Navier–Stokes Millennium Problem concerns rigorous existence and smoothness properties of solutions to the equations, rather than simply whether engineers can make useful forecasts. The Clay Mathematics Institute attached a million-dollar prize to a qualifying solution. OpenAI has announced what it says is a solution; the claim and its surrounding controversy need careful examination.

The equations model water in pipes, air around wings, ocean currents and smoke. Better fluid modeling can matter for flight, vehicles, chip cooling and weather. I am particularly interested in the use of AI agents to tackle difficult mathematics, although a proof about a specified mathematical setting does not itself deliver all of those engineering improvements.

In its [announcement](https://openai.com/index/navier-stokes-solution/), OpenAI says a group of agents using a next-generation model produced a proof of finite-time blowup for a smooth, initially stationary fluid under smooth external forcing. This addresses the forced alternatives in the Millennium Problem. The company published a mathematical write-up and a Lean formalization; independent mathematical review remains essential.

I see this as an exciting possible step toward using AI in materials science and medicine, and I am struck by the company’s claim that the model is substantially more capable than GPT-6 Astra. Those applications are hopes, not results demonstrated by this particular proof. On September 11, after this video, [Clay said](https://www.claymath.org/news/navier-stokes-announcement/) the problem had apparently been settled and that it would arrange a careful evaluation; that is not the same as awarding the prize.

The announcement also brought a dispute involving Tristan Buckmaster and Levent Alpöge, an Anthropic employee who described their work as a personal collaboration. They had used OpenAI’s Codex while pursuing related fluid mathematics. In his [statement](https://cims.nyu.edu/~tristanb/statement.pdf), Buckmaster describes results for the forced Euler, Boussinesq and porous-media equations; these are related to, but distinct from, the claimed forced Navier–Stokes result.

Buckmaster says he contacted OpenAI on September 3 after hearing the company knew about their progress, and that OpenAI told them on September 6 about its own apparent result. Berman finds the similarity in direction striking and discusses whether learning of their progress could have helped steer OpenAI’s effort. OpenAI’s account says its internal work began September 1, before that email; the sequence and its implications are contested.

Buckmaster says he asked whether their private Codex drafts or usage had influenced OpenAI’s models and did not receive a clear answer at the time. He also describes proposals for publication and authorship that concerned him. Crucially, he writes that he does not know whether their data was used and is not accusing anyone of anything; his questions should not be recast as proof of misuse.

OpenAI said its researchers and agents had not seen the mathematicians’ work through any means before its public release and had not accessed specific user data to solve the problem. Its initial statement also said it could not rule out that de-identified data from their product use had helped improve its models. Berman reads that qualification as a warning about entrusting research to an AI provider.

In a September 10 update, after this video, OpenAI said an internal investigation had found that Buckmaster’s Codex prompts from the preceding two months could not have influenced the system, including through training. That is the company’s reported finding, not an independent determination. Its earlier qualification and Buckmaster’s account explain why Berman raised the issue, but neither establishes that private drafts were accessed for the proof.

This brings me to recursive self-improvement: AI discovering ways to improve AI. Labs already describe using models to assist with research and the development of later models. A fully closed loop would go further, with AI proposing and running experiments, applying improvements to itself and repeating the process; Berman acknowledges that such a loop is not yet in place.

Berman also reads Sebastian Bubeck’s account of his communications with the mathematicians. Bubeck says he sought to coordinate releases, did not ask for Alpöge to be removed from authorship of his own work, and wanted to celebrate their achievements. Those assertions and Buckmaster’s description are separate perspectives on a disputed interaction.

OpenAI reports that the proposed Navier–Stokes proof was produced in 88 hours, from September 1 to September 5, with another 17 hours for Lean formalization. Berman cites 4.9 million agent messages and 300 billion output tokens. OpenAI’s breakdown attributes those larger totals to all attempted problems; the Navier–Stokes effort accounts for 2.7 million messages and 130 billion output tokens. Human researchers also oversaw and interacted with the agents.

I wonder what happens if AI can discover mathematics that helps improve AI itself. Berman suggests that a Millennium Problem result might imply the ability to solve any mathematical problem and lead to exponential self-improvement. Neither inference follows from this one claimed proof, and reliable autonomous improvement remains an open question.

The potential upside Berman imagines is enormous: new materials, improved environmental technology, better weather prediction and perhaps cures for disease. He even wonders whether a six-hour flight could one day take an hour. These are illustrations of his hopes, not demonstrated consequences or timelines.

What does faster machine discovery mean for people? Berman compares frontier research with chess, where leading engines outperform human champions but people still enjoy human competition. He argues that, unlike competitive chess, research is valued chiefly for discovering knowledge as quickly as possible, and predicts AI will dominate that work. The analogy is a viewpoint, not evidence that human contributions will disappear.

Finally, businesses building on AI models face platform risk: the provider supplies a capability on which they depend and may also operate in related markets. Berman worries that use of a model could expose a business’s expertise to future competing systems, and recommends examining provider data practices and considering open models.

The actual treatment of data depends on the particular service, account and settings; the initial OpenAI qualification does not establish that every business’s confidential information trains models or becomes available to competitors. The episode brings together three questions Berman finds fascinating: a disputed mathematical breakthrough, the handling of research data, and the prospects for AI-assisted improvement.