OpenAI's Navier-Stokes Proof Sparks a Mathematician Plagiarism Dispute
OpenAI says its model proved a Navier-Stokes singularity result. Two mathematicians say it may have echoed their unpublished, private work.
What happened with OpenAI’s Navier-Stokes claim?
OpenAI announced that an internal AI system produced a formally verified proof related to Navier-Stokes blowup, the question of whether a smooth, well-behaved fluid can spontaneously develop infinite speeds in finite time. The company said the result was checked using Lean, a proof assistant that verifies mathematical arguments step by step, and that the work involved roughly 10,000 AI agents exchanging close to 5 million messages over about 88 hours. Days earlier, two mathematicians, Tristan Buckmaster of NYU and a researcher named Elgindi who works at Anthropic, had published their own year-long results on a closely related cluster of equations. The overlap in timing and approach triggered a dispute over whether OpenAI’s model had somehow drawn on their private, unpublished work.
TL;DR
- Navier-Stokes blowup is one of the seven Millennium Prize Problems, unsolved for roughly 90 years, asking whether fluid described by the equation can spiral into a singularity rather than stay smooth forever.
- OpenAI claims an internal model generated a Lean-verified proof of forced blowup using massive compute, described as around 10,000 agents and nearly 5 million messages over 88 hours.
- Mathematicians Tristan Buckmaster and Elgindi had spent about a year working a narrow, unusual angle on related equations (Euler, Boussinesq, porous media flow) before OpenAI’s announcement.
- Buckmaster says Sebastian Bubeck, who leads OpenAI’s math efforts, told him an internal model had already produced a roughly 100-page blowup proof using the same uncommon approach almost nobody else in the field was pursuing.
- The dispute centers on whether OpenAI’s model had access to a private Codex session where Buckmaster and Elgindi were drafting their unpublished work, a question Buckmaster says went unanswered when he pushed specifically on training data.
- Buckmaster’s account describes pressure over authorship, including suggestions to give the AI model credit and to leave his collaborator off a paper because of a competing employer, plus a warning about career risk if he went public.
- The episode raises a broader concern for anyone doing serious original work with AI coding and research tools: if private drafts can influence a model’s output with no way to prove it, attribution disputes like this one may become more common.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
Why does the Navier-Stokes blowup problem matter?
The Navier-Stokes equations describe how fluids move: water in a pipe, air over a wing, blood through a vessel, weather systems churning across a continent. Engineers and forecasters rely on them constantly. But mathematicians have never been able to prove, in three dimensions, that solutions to the equations always stay smooth. The open question is whether a fluid that starts out calm and well-behaved can, under the equation’s own dynamics, generate infinite velocity in a finite amount of time. That kind of blowup would mean the equations break down exactly when things get extreme, a serious theoretical gap even though the equations work fine in everyday practice.
This is one of the seven Millennium Prize Problems, the set of hardest unsolved questions in mathematics, each carrying a million-dollar prize. Before this announcement, nobody had resolved it. A verified proof, even for a specific case (forced blowup, meaning blowup that occurs under an external forcing term rather than the fully general unforced equation), would be a landmark result.
What did OpenAI actually claim?
OpenAI’s announcement described an internal system producing a proof of forced Navier-Stokes blowup, formally checked in Lean, a tool that verifies each logical step of a mathematical argument with no room for hand-waving. The company framed this as the product of enormous parallel compute: roughly 10,000 AI agents working together, generating close to 5 million messages, arriving at the result in about 88 hours. On its face, that is a striking claim, an AI system reportedly closing part of a 90-year-old open problem through brute-force collaborative search rather than a single elegant human insight.
How does the mathematicians’ dispute complicate the story?
A few days before OpenAI’s announcement, Tristan Buckmaster and Elgindi (who works at Anthropic) had published their own findings, the product of roughly a year of work using AI tools to assist with three related equations: Euler, Boussinesq, and porous media flow. According to Buckmaster, once rumors circulated that he and Elgindi had a major result in hand, he received a call from Sebastian Bubeck, who leads math efforts at OpenAI. Bubeck reportedly told him an internal model had already generated a roughly 100-page proof of forced Navier-Stokes blowup, using the same narrow and unusual approach that Buckmaster and Elgindi had been quietly developing, an angle Buckmaster says almost no one else in the field was pursuing at the time.
The exchange got more specific from there. Buckmaster says Elgindi asked OpenAI for the precise claim being made and received a description matching existence of forced blowup in the R3 and T3 settings (standard notations for three-dimensional space and a three-dimensional periodic domain in this area of math). Screenshots of these exchanges were later shared publicly by Bubeck on X.
One coffee. One working app.
You bring the idea. Remy manages the project.
Buckmaster’s account, laid out in a detailed blog post, says he was presented with two options: a joint release with OpenAI, or writing up the result himself while crediting OpenAI’s model as the source. He also says there was pressure to leave Elgindi off any resulting paper because Elgindi works at a rival lab, Anthropic. When Buckmaster indicated he might go public about the situation, he says he was warned about damaging his own career.
What is the unanswered question at the center of this?
The detail that matters most, according to Buckmaster, is whether OpenAI’s model had any access to a private Codex session where he and Elgindi had been drafting their unpublished work over the preceding year. He says he asked directly and was told the model does not look up user data. When he pressed further, asking specifically about whether such sessions could have been used in training, he says he got no clear answer.
That distinction matters. A coding assistant “looking up” a user’s live session is different from that session’s content later being incorporated into a training run. OpenAI’s account, as described in the dispute, addressed the former but reportedly left the latter unresolved. Without a clear answer, there is no way for outside observers, or for Buckmaster himself, to confirm whether the overlap in approach was coincidence, convergent mathematical reasoning, or something closer to unacknowledged use of private work.
Is this a bigger issue than one dispute?
The specifics of this case involve two individual mathematicians, a narrow technical approach to a hard equation, and one company’s internal tooling. But the underlying question extends well past this incident. Researchers, engineers, and writers increasingly draft private, unpublished work inside AI-assisted coding and writing tools. If those tools, or the models behind them, can absorb that material in ways that later surface in a company’s own output, with no reliable way to trace or disprove it, that is a structural problem for anyone doing original work with AI assistance. It also puts AI labs in an awkward position: even a completely independent internal result can look like it was lifted, simply because there’s no transparent audit trail connecting training data to output.
Frequently Asked Questions
What is the Navier-Stokes blowup problem?
It asks whether a fluid governed by the Navier-Stokes equations, starting from smooth, well-behaved initial conditions, can develop infinite speed in a finite amount of time. It’s one of the seven Millennium Prize Problems and had gone unsolved for about 90 years before this dispute arose.
What did OpenAI say its AI system did?
OpenAI said an internal model produced a proof of forced Navier-Stokes blowup, verified using the Lean proof assistant, generated through a large-scale process involving around 10,000 AI agents and nearly 5 million messages over roughly 88 hours.
Who are Tristan Buckmaster and Elgindi?
Tristan Buckmaster is a mathematician at NYU. Elgindi is a mathematician who works at Anthropic. The two spent about a year working on related fluid equations (Euler, Boussinesq, and porous media flow) using AI assistance, publishing their results just days before OpenAI’s announcement.
What is the core allegation against OpenAI?
Buckmaster says OpenAI’s model arrived at a proof using the same unusual, narrow approach he and Elgindi had been developing privately, and that OpenAI never clearly answered whether the model had been trained on their unpublished Codex drafts.
Has OpenAI confirmed or denied using private data?
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
According to Buckmaster’s account, OpenAI told him the model does not look up user session data in real time, but did not give a direct answer when he asked specifically whether such private sessions could have factored into training.
