On September 8, OpenAI announced that its experimental AI found a solution to the Navier-Stokes problem — one of math's famous "Millennium Problems." The company published the proof. But now people aren't just discussing the math — they're also discussing whether it borrowed someone else's ideas.
What are these equations, and why do they matter?
The Navier-Stokes equations describe how fluids and gases move: how water flows, how air moves, how vortices spin up.
These are the mathematical rules of motion. But knowing the rules isn't the same as understanding everything that can happen under them. That's exactly where the problem hides.
What was the million-dollar prize for?
The Clay Institute offered a million dollars for an answer to one question: can the mathematical description of fluid motion "break," even if everything started out fine?
The terms of the problem allow for two paths:
1. Prove that, given admissible initial data and no external forcing, the solution always stays smooth — no mathematical discontinuities or infinities.
2. Find at least one admissible scenario where a mathematical "blowup" occurs in finite time. Here the rules allow for a specially chosen but smooth external force.
Important: "smooth" doesn't mean "calm, no whirlpools." It's about mathematical properties, not a pretty water surface.
What exactly did the AI find?
Based on the published paper, the system found the second option.
In the model, the fluid starts out at rest. Then a specially chosen external force acts on it — no jumps, no infinities.
A vortex forms. The region of most intense motion shrinks while the velocity inside it grows without bound. And it happens in finite time — not "eventually, given eternity." The total energy stays finite throughout.
This is what's called a *blowup* — a mathematical "explosion." And it's not just a pretty computer simulation: OpenAI presented a mathematical proof of this scenario.
Real water isn't about to learn how to move at infinite speed because of this. The result shows the limit of a mathematical model — it doesn't hand you a way to blow up the universe in your kitchen sink.
How did the AI get there?
Not out of nowhere. Mathematicians have studied scenarios like this for a while. Terence Tao (mathematician, Fields Medalist), for instance, proved a "blowup" was possible for a modified version of the equations back in 2014. But that still wasn't a solution to the original problem.
According to OpenAI, the group that found the solution ran about 10,000 AI agents — programs that explored the problem in parallel and traded results.
From the first agents launching to the solution landing took about 88 hours. This wasn't "ask regular ChatGPT a question and get an answer" — it was a massive computational project.
So where does the theft story come from?
Two mathematicians — Tristan Buckmaster and Levent Alpöge — had been working on related fluid-motion problems. They had results of their own, but not the same Navier-Stokes proof OpenAI presented.
Buckmaster voiced his suspicion: could their unpublished work — including material fed into Codex — have helped the company along?
OpenAI denies that its employees or agents accessed this work before it was published. But the company doesn't rule out that anonymized data from product usage could have played a role in improving its models.
The suspicion is serious. But it's not yet proof of a "peeked at the drafts and stole the discovery" scheme.
So where does that leave us?
The proof has been published, including a version for computer verification. But publication, independent verification, and official recognition are three different things.
Under the Clay Prize rules, you need publication in a suitable scientific journal, a minimum two-year wait after that, and recognition from the mathematical community. OpenAI itself isn't claiming the million.
And the most uncomfortable question here is this: how comfortable can you really be discussing a discovery with an AI when the company that owns the tool is itself racing in the same scientific competition?