AI-assisted Navier-Stokes research sparks an OpenAI priority dispute

Update

What changed

Sam Altman has added his reaction to the reported Navier–Stokes development, calling it one of the most amazing moments in OpenAI history. His post also attributes the claimed proof to a group of agents using a next-generation OpenAI model. The proof’s validity and standing as a solution to the Millennium Prize Problem remain unverified in the supplied evidence.

Sources and evidence

• Sam Altman, via X: Sam Altman described the reported Navier–Stokes development as one of the most amazing moments in OpenAI history.

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

OpenAI's Mark Chen says no human or agent looked at user data as part of the Navier-Stokes effort. He separately says the company uses user feedback and de-identified data to improve ChatGPT and Codex generally.

The clarification matters because it distinguishes the specific research project from OpenAI's broader model-improvement practices. WittyWires has not independently verified the underlying data-handling process.

What to watch

– Whether OpenAI provides more detail on the datasets and safeguards involved. – Whether the mathematicians involved accept the distinction between the research effort and general model improvement.

Sources and evidence

• Mark Chen on X: Mark Chen says no human or agent looked at user data as part of the Navier-Stokes effort.

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

Update: the timeline and data question come into focus

Simon Willison’s account adds a much more specific chronology to the priority dispute. He reports that NYU mathematician Tristan Buckmaster and Levent Alpöge, who works for Anthropic, spent almost a year on related problems using Claude and Codex and reached a breakthrough on 15 August.

The OpenAI account reproduced by Willison says the company began its own effort on 1 September after hearing rumours that two Millennium Prize problems had been resolved. OpenAI says its agents arrived at a Navier–Stokes resolution on 5 September, about 88 hours after launch, followed by another 17 hours of Lean formalisation and verification using GPT-6 Astra.

The reported scale is extraordinary: 4.9 million agent messages and about 300 billion output tokens across all attempted problems, including 2.7 million messages and roughly 130 billion tokens on Navier–Stokes alone. Those are OpenAI’s figures, not independently verified measurements.

The thorniest new detail concerns training data. OpenAI reportedly says its researchers and agents did not see Buckmaster and Alpöge’s work before publication and that no specific user data was accessed to solve the problem. It also says it cannot rule out de-identified data derived from their product usage having helped improve its models.

That does not establish that either team’s work trained the system or determined OpenAI’s result. It does explain why the dispute matters beyond academic credit. Researchers using frontier models need a far clearer account of whether private, de-identified or aggregated interactions can later shape systems used by competitors.

The mathematical claims still require expert scrutiny, and the competing priority accounts remain disputed. For now, the useful shift is from vague suspicion to three testable questions: when each approach began, what information moved between people and systems, and precisely how customer research data may influence later models.

Sources and evidence

• Simon Willison’s account of Buckmaster’s published complaint: Simon Willison reports that Tristan Buckmaster and Levent Alpöge worked on related Navier–Stokes problems for almost a year and reached a breakthrough on 15 August after using Claude and Codex extensively.

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

Update: OpenAI details the timeline and scale

The dispute now has a much clearer chronology and a rather enormous compute bill attached. According to OpenAI’s account quoted by Simon Willison, the lab launched its effort on 1 September after hearing rumours that two Millennium Prize problems had been resolved. It says its agents reached a Navier–Stokes result on 5 September, about 88 hours later, before another 17 hours of Lean formalisation and verification using GPT-6 Astra.

OpenAI reports that agents sent 4.9 million messages and produced roughly 300 billion output tokens across all attempted problems. The Navier–Stokes work alone reportedly involved 2.7 million messages and about 130 billion output tokens. Those are OpenAI’s figures, not independently audited measurements, but they reveal how aggressively a well-resourced lab can pursue a mathematical lead once the rumour mill starts humming.

The data question remains unresolved in an important way. OpenAI says its researchers and agents did not see Tristan Buckmaster and Levent Alpöge’s work before it became public, and that no specific user data was accessed to solve the problem. It also says it cannot rule out de-identified data from their use of its products having helped improve its models.

That is narrower than Buckmaster’s concern, not a complete answer to it. His account says the pair had worked on related problems for almost a year using Claude and Codex, reached a breakthrough on 15 August, and later pressed OpenAI about when its prompting began and whether their sessions could have influenced training.

WittyWires has not independently reviewed the underlying PDF or validated either proof. The useful new fact is the shape of the disagreement: OpenAI denies direct access while leaving open a more diffuse route through model improvement. In AI-assisted research, that distinction could become as consequential as the priority date itself.

Sources and evidence

• OpenAI, as quoted by Simon Willison: Simon Willison reports that OpenAI said it began the effort on 1 September after hearing rumours of breakthroughs on two Millennium Prize problems.

Independent WittyWires Watcher; not an official account or feed.

Update

What changed

Update: the row is becoming a trust problem

The priority dispute around OpenAI’s proposed Navier-Stokes breakthrough is widening beyond who reached the result first. Mathematicians interviewed by The Verge say the episode could make researchers more guarded about sharing incomplete ideas and using AI tools while work is still unpublished.

OpenAI says its researchers and agents did not see Tristan Buckmaster and Levent Alpöge’s work before publication and that no specific user data was accessed to solve the problem. It also says it cannot entirely rule out the possibility that de-identified data derived from their use of its products indirectly helped improve its models. That is not evidence that their work influenced the result, but it leaves the provenance question less neatly closed than a flat denial might suggest.

The Verge also adds detail to the chronology. OpenAI says it launched the roughly 10,000-agent effort after hearing social-media rumours that researchers were progressing on Millennium Prize problems, and only later realised the rumours concerned Buckmaster and Alpöge. Buckmaster’s account of hostile exchanges remains disputed, with OpenAI mathematics lead Sébastien Bubeck denying parts of it.

The practical lesson is already emerging. Researchers considering AI tools for confidential work need explicit, auditable answers about retention, training use and access to their sessions. Otherwise the cleverest assistant in the room may also become the reason nobody discusses half-formed ideas in the room at all.

Sources and evidence

• The Verge, citing OpenAI’s announcement: The Verge reports that OpenAI said an unreleased model, supported by roughly 10,000 AI agents, produced a proposed Navier-Stokes solution in 88 hours.

Independent WittyWires Watcher; not an official account or feed.