OpenAI’s Navier–Stokes result solves a narrower problem

What OpenAI actually established

The Navier–Stokes equations describe how fluids move, from air currents to ocean flows. Scientists have used them for almost two centuries and can obtain reliable results with them, but they still do not fully understand why the equations work in every case.

The problem is to determine whether the equations can produce a physically nonsensical outcome—for example, a mathematical description in which a fluid accelerates without limit. In 2000, the Clay Mathematics Institute placed the problem on its list of six Millennium Prize Problems, offering $1 million for a solution to any one of them.

OpenAI said its agents found a scenario leading to an infinite fluid velocity, creating an “explosion”: a singularity that cannot exist in the physical universe.

The result has already triggered two separate disputes:

The question is not only whether the proof is technically correct. It is whether a proof that specialists cannot readily interpret can serve the role mathematics normally demands of it.

The version of the problem matters

The Navier–Stokes problem has two versions that qualify for the Clay Institute’s prize:

Specialists are primarily interested in the cleaner scenario: whether an explosion can occur without external intervention and under the action of the fluid’s internal forces alone.

OpenAI’s result relies on an external force. A recent proof published by mathematicians argues that this approach cannot, in principle, solve the full problem.

Luis Silvestre, a University of Chicago mathematician, told Scientific American that the central problem remains unsolved. The Clay Institute’s version may be closed, he said, but the main Navier–Stokes problem is not.

That distinction is easy to lose in the phrase “solved the Navier–Stokes problem.” OpenAI appears to have produced a valid result within an eligible formulation, not settled the version that many mathematicians consider the deeper test.

A proof is also a research artifact

The technical correctness of OpenAI’s work is not the only controversy. Hours before the announcement, New York University mathematician Tristan Buckmaster accused the company of using his research without warning or permission and then extending his ideas into the final result.

Buckmaster said OpenAI’s work followed a path similar to his own and that nobody else in the field had pursued the same approach. He had used OpenAI’s Codex in his research and suggested that the company might have secretly accessed his notes.

According to Buckmaster, when he contacted OpenAI, the company offered to list him as an author if he agreed not to collaborate with Levent Alpoege, who worked at Anthropic. OpenAI rejected the allegations but acknowledged that it could not rule out the possibility that anonymized data from its products had helped improve its models.

My view is that the attribution dispute and the opaque proof point to the same weakness: the result is being presented as an endpoint when it may be more useful as raw material. Mathematicians agree that the work is technically complex but correct, yet understanding the AI’s reasoning could take many more weeks. If a machine can produce a valid proof only after humans reconstruct its meaning, the bottleneck has moved from finding an answer to making the answer part of mathematics.