OpenAI Solved Navier-Stokes. The Fallout Is Worse.

On September 8, OpenAI announced that 10,000 AI agents running on an unreleased model had solved the Navier-Stokes existence and smoothness problem — one of the seven Millennium Prize Problems, each worth $1 million from the Clay Mathematics Institute. The 166-page proof, formally verified in Lean, shows that three-dimensional Navier-Stokes equations can develop a singularity in finite time. If it holds up to peer review, it is the most significant mathematical result ever produced by an AI system.

📖 Read the full version with charts and embedded sources on ComputeLeap →

But within hours, the achievement was overshadowed by something uglier. NYU mathematician Tristan Buckmaster alleged that OpenAI had fought dirty — front-running his own unpublished work, pressuring him to drop a co-author who works at Anthropic, and leveraging rumor intelligence to mobilize massive compute against a problem two academics were quietly solving on their own. The controversy exposes a structural conflict that goes far beyond one math proof: AI labs are simultaneously selling tools to researchers and competing against them.

View full discussion on Hacker News →

The Timeline That Broke Trust

The sequence of events reads like a case study in platform risk.

Tristan Buckmaster and Harvard mathematician Levent Alpoge had been collaborating on fluid dynamics problems for nearly a year. They used OpenAI's Codex as a research scratchpad — typing in drafts, proofs, and mathematical logic. By August 15, they had achieved a breakthrough on the forced Euler equations, a closely related problem. Word leaked through the academic grapevine.

On September 1 — days after those rumors started circulating — OpenAI launched what it described as a massive internal effort. According to OpenAI's own account: "Our effort began on September 1st after hearing a rumor which we later realized was related to Levent Alpoge...and Tristan Buckmaster."

Five days later, OpenAI announced they had beaten the academics to the punch — not on the Euler equations Buckmaster and Alpoge were working on, but on the full Navier-Stokes problem, which extends their approach to viscous fluids.

As Nature reported, the proof took 88 hours of compute with 10,000 agents working concurrently. The Quanta Magazine coverage called it "by a significant margin, the most important mathematical proof to have been arrived at by an artificial-intelligence model to date."

The Codex Problem

Buckmaster's most damaging allegation centers on data access. He and Alpoge had been pouring their unpublished mathematical work into Codex sessions for months. The central question: could that data have informed OpenAI's own research effort?

OpenAI's response was notably hedged. As Simon Willison highlighted, the company stated: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models."

That single sentence should concern every researcher and developer who uses AI lab products for proprietary work. OpenAI is saying, in plain English, that they cannot guarantee your inputs don't train their models — even when you think you've opted out.

⚠️ The structural problem: OpenAI reserves the right to train models on user interactions. Users may opt out, but the "Help improve model" toggle has reportedly re-enabled itself. Even if the toggle works perfectly, de-identified aggregate data from Codex sessions could still shape model behavior on similar problems.

Buckmaster described the situation as "absolute academic malpractice." He raised a question that resonates beyond mathematics: if you use a lab's tool to develop a breakthrough, and the lab then solves a competing version of the same problem days later, how can anyone prove what the model did or didn't "know"?

View full discussion on Hacker News →

View original article on Substack →

The Anthropic Co-Author Pressure

The controversy deepened when Buckmaster alleged that OpenAI's math team lead, Sebastien Bubeck, pressured him to remove Alpoge's name from any joint announcement — because Alpoge works at Anthropic, OpenAI's direct competitor.

According to TechCrunch's reporting, Bubeck told Buckmaster: "Why would you ruin your career?" followed by "If you don't want me to be nice, then I don't have to be nice."

Alpoge himself disputed OpenAI's framing, telling reporters: "I also like the idea of the labs cooperating...It's a shame!"

OpenAI denied any misconduct. Bubeck called the allegations "false and inflammatory." But the damage was done — the story was no longer about mathematics.

The Community Verdict

The Hacker News thread "More questions about whether researchers can trust OpenAI with unpublished math" hit 740 points with 682 comments — an unusually intense discussion even by HN standards. The community's verdict was harsh and structural.

One commenter framed the core issue: "OpenAI is directly competing against their own customers." Another drew the analogy that had been forming across the tech world: this is the Amazon marketplace problem, but for science. The platform operator uses its privileged position — access to customer data, knowledge of market demand, superior resources — to compete against the very people who depend on it.

View full discussion on Hacker News →

A commenter on the related thread "OpenAI fought dirty on career-making math problem" cut deeper: "Nobody actually cares about the proof itself. The tragedy is that labs solve problems for marketing rather than mathematical advancement."

ℹ️ Terence Tao's warning deserves special attention. The Fields Medal winner noted: "Even the rumor of someone working on a problem can trigger massive AI-powered effort to flatten it before research reaches full potential." This describes a new failure mode for academic research — one where merely being close to a breakthrough makes you a target.

What the Money Says

Prediction markets moved fast. On Polymarket, the "AI lab announces another Millennium Prize solution" market spiked 27.5% in a single day — the biggest mover on the entire AI board — reaching 85% probability on the near-term contract with $61K in volume.

That 27-point single-day jump doesn't happen on vibes. The market is pricing in that this won't be the last time a lab throws thousands of agents at a famous unsolved problem. The Millennium Prize isn't just a math competition anymore — it's a marketing opportunity for labs that need to justify billion-dollar compute investments.

Meanwhile, the broader "best AI model" market on Polymarket shows Anthropic at 88% for September, with OpenAI at just 9%. The math controversy hasn't boosted OpenAI's standing — if anything, it has reinforced the perception that OpenAI prioritizes spectacle over substance.

View full discussion on Hacker News →

The Platform Risk Pattern

If this story feels familiar, it should. The tech industry has a long history of platforms competing with their customers:

• Amazon uses marketplace seller data to identify hot products, then launches competing Amazon Basics versions.
• Google promotes its own products in search results over competitors who depend on search traffic.
• Apple copies successful App Store apps into iOS features, a practice developers call "Sherlocking."
• AWS has repeatedly launched managed services that compete directly with startups building on its infrastructure.

The AI lab version is more insidious because the "product" being competed on is knowledge itself. When Amazon copies a bestselling garlic press, the original seller still has their design. When an AI lab potentially absorbs your unpublished mathematical breakthrough through tool usage data, you have nothing left.

View original article on The Decoder →

⚠️ Contrarian Corner: It is entirely possible OpenAI is telling the truth — that they independently arrived at their proof without accessing Buckmaster's Codex data. The mathematical approach (attacking Navier-Stokes through forced Euler equations) was not secret; it was a known viable strategy. But that is precisely the point: even if OpenAI is completely innocent in this specific case, the architecture of the relationship makes trust impossible. A lab that sells research tools AND conducts competing research faces the same structural conflict as a stock exchange that also trades stocks. You cannot audit your way out of a conflict of interest — you have to eliminate it.

What This Means for You

The implications extend far beyond mathematics.

If you are a developer building on AI APIs: Your prompts, code, and interactions may be training the next model that competes with your product. Read the terms of service carefully. Opt out of training data contribution where possible. But understand that "de-identified aggregate data" is a category broad enough to drive a truck through.

If you are a researcher: The Buckmaster case is a warning shot. Any unpublished work you feed into a commercial AI tool is potentially accessible — directly or indirectly — to the lab that operates it. Consider using open-source models for sensitive pre-publication work, even if they are less capable.

If you are a startup founder: Platform risk from AI labs is real and growing. OpenAI's API cutoff of Cursor showed that labs will protect their interests over customer relationships. The Navier-Stokes controversy shows they may also compete directly with the work customers produce using their tools. Build with exit strategies. Use multiple providers. Do not make a single lab your sole dependency — we have written about this pattern before.

If you are evaluating the lab landscape: The Anthropic vs. OpenAI rivalry now has a new dimension. Anthropic's employee was the co-author OpenAI allegedly tried to erase. The competitive dynamics between labs are bleeding into academic research, and researchers are becoming collateral damage.

The Proof Might Be Real. The Trust Is Gone.

The mathematics may well be correct. Nature, Quanta Magazine, and independent mathematicians have noted that the Lean-verified proof is available for scrutiny, and early assessments are cautiously positive. Science magazine's deep dive documented the competing claims in detail. If confirmed through peer review, it would represent a genuine milestone — not just for AI, but for mathematics itself.

But the achievement is now permanently stained by the process. As Axios noted, "the controversy strikes at a core trust question for AI-assisted science: whether researchers can safely use frontier labs' tools to work on unpublished discoveries."

The answer, after this week, is clearly no — not without structural safeguards that do not yet exist. The Clay Mathematics Institute requires two years of community validation before awarding a Millennium Prize. OpenAI may have the proof. But the community's trust will take far longer to verify.

---

The Navier-Stokes Millennium Prize Problem asks whether smooth solutions to the equations describing fluid motion always exist, or whether they can "blow up" into singularities. OpenAI's proof claims the latter — that finite-time blowup is possible. The Clay Mathematics Institute has not yet commented on the claim.

Originally published at ComputeLeap