OpenAI Says 10,000 Agents Solved a 90-Year-Old Math Problem. Two Mathematicians Are Asking If the Agents Read Their Homework First.

OpenAI announced that roughly 10,000 coordinating AI agents worked for 88 hours and produced a proof that a three-dimensional fluid governed by the Navier-Stokes equations can develop a singularity in finite time, a question mathematicians have chased for close to a century. That's a genuinely big claim on its own. What's actually dominating the conversation this week isn't the math, it's a fight over whether the swarm got there cleanly, and it's the kind of dispute that should make anyone doing serious, unpublished work inside a coding agent pay closer attention to where that work actually goes.

The proof, and the collision

A month before OpenAI's announcement, NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge had reportedly made their own breakthrough, on August 15, proving finite-time blowup with smooth forcing for the 3D Euler equations, the frictionless cousin of Navier-Stokes. That work wasn't published. OpenAI's announcement landed on top of it weeks later, covering the related but distinct Navier-Stokes case.

Buckmaster has said he never saw OpenAI's proof and doesn't know what OpenAI's models actually did or whether his or Alpöge's data fed into them. He's also said, separately, that information about their progress reached OpenAI before the announcement. Those two statements sitting next to each other are exactly why this is a story and not a footnote: an unpublished result reportedly became known to a competing lab, and shortly after, that lab published a related result using an approach nobody outside OpenAI can fully audit.

What OpenAI says, and what it doesn't rule out

OpenAI's position is that its researchers and its agents did not see Buckmaster and Alpöge's unpublished work before releasing the Navier-Stokes proof, and the company denies inspecting any specific user's private data. Where it gets uncomfortable is the next sentence: OpenAI concedes that de-identified platform usage data could plausibly have helped train its models. That's not an admission of wrongdoing, and it's a genuinely hard line to draw cleanly, "we didn't look at your session" and "our training data may include patterns from sessions like yours" are different claims. But for a researcher who ran unpublished proof attempts through OpenAI's own coding tools, the distinction between those two claims is exactly the thing they can't independently verify.

There's a second, messier thread here too: a proposed path in which Buckmaster might help present OpenAI's Navier-Stokes proof without Alpöge listed as an author, with Alpöge's employment at a rival lab (Anthropic) cited as the complication. OpenAI's Sebastien Bubeck disputed that characterization directly, saying he never asked for Alpöge to be dropped from authorship of his own work. Whatever actually happened in that conversation, the fact that authorship credit and competitive lab affiliation got tangled together at all tells you how much is riding on who gets named as the source of a breakthrough this visible.

Why this matters if you're not a mathematician

Strip out the equations and what's left is a data-boundary question that applies to anyone using a coding agent, chat assistant, or IDE integration for real, proprietary work: does the platform's training pipeline, even in de-identified form, learn from what you type into it, and if it does, what happens the day a competitor produces something that looks a lot like what you were building. Most terms of service have a clause covering this. Almost nobody reads it closely enough to know what "de-identified" actually promises them, and this dispute is the first time that gap has played out in public at a scale big enough to make headlines instead of a support forum complaint.

If you're a solo operator prototyping something you consider genuinely novel, a pricing model, a niche algorithm, a dataset you built by hand, inside Claude Code, Cursor, Copilot, or any agent-based tool, this is worth an honest five minutes with whatever privacy or training-data policy you agreed to when you signed up. Not because your side project is as high-stakes as a century-old open math problem, but because the mechanism (usage patterns training future model behavior) doesn't care how important your work is. It only cares that you typed it somewhere.

The honest take

I think it's likely both things are true at once: OpenAI's agents did not literally read Buckmaster and Alpöge's private files, and OpenAI's models still benefited, indirectly and diffusely, from being trained on a mountain of usage data that included patterns similar to what serious researchers were doing across competing platforms. That's not a smoking gun. It's just what training on aggregate usage data looks like when it works as designed, and it's uncomfortable precisely because nobody set out to steal anything, the system just doesn't have a clean boundary for "this specific person's unpublished breakthrough" versus "the general shape of how people solve hard problems."

Where I could be wrong: if OpenAI's account holds up completely, no training signal from that specific unpublished work reached the model in any meaningful way, then this is a story about optics and unfortunate timing, not data leakage, and the authorship dispute is a separate, more mundane conflict over academic credit that got dragged into a bigger story because the timing was suspicious. I don't think that's the way to bet, but it's not ruled out either.

What I'd actually do

If you use a coding agent for anything you'd call genuinely proprietary, check whether the vendor offers an enterprise or zero-retention tier that actually excludes your sessions from training, not just from human review. Most do, and most solo operators never turn it on because it costs more or requires a sales call. After this week, that tradeoff is worth re-pricing. The free tier's real cost was never the dollar amount, it was always this.

Author

Lukas

@lukcombinator

Sources

• OpenAI Navier-Stokes Proof Sparks Credit Dispute With NYU Mathematician
• OpenAI solves longstanding math problem with 10,000-agent swarm, but can't rule out benefitting from a researcher's private Codex data
• OpenAI's 10,000-Agent Navier-Stokes Claim Solves the Wrong Problem, and the Right One Has a Provenance Controversy