The million dollar question hiding in your coffee cup

Pour cream into coffee and don't stir. Just watch it for a few seconds. You get these pale curls that fold over on themselves, stretch, thin out, and then give up and go beige. Somewhere inside those fifteen seconds is a problem worth a million dollars, and this month it may or may not have been solved. The "may or may not" is the good part, so stay with me.

It's just F = ma, wearing a trench coat

The Navier-Stokes equations get described as some of the most feared equations in mathematics, which does them a disservice. At heart they are Newton's second law applied to fluid. In physics class you write F = ma for a block sliding down a ramp. For a fluid you write the same idea for every tiny parcel of water or air, all at once, and you keep track of three things pushing on it: pressure from the neighbouring fluid, gravity, and friction from the fluid sliding against itself. That last one is viscosity. Honey has a lot. Air has a little.

Claude-Louis Navier and George Gabriel Stokes put this together between the 1820s and 1840s. No computers. No proof that atoms existed. Just the stubborn idea that a river and a puff of smoke ought to obey the same rules as a falling apple.

And they do. That's what I want you to sit with for a second. Every weather forecast you have ever checked came from running these equations forward in time. Every aircraft wing and Formula One diffuser was shaped by them. Blood flow through a stent, ventilation in a hospital room, how a wildfire jumps a road, the ocean in an animated film, the cooling fan in the laptop you're reading this on. Engineers solve Navier-Stokes numerically thousands of times a day and it works.

Nobody can prove it always will.

The actual gap

Here is the precise question. Start with a fluid that is perfectly smooth, with no sharp corners and a sensible finite amount of energy. Let the equations run. Do the solutions stay smooth forever, or can the math reach a moment where the velocity at some point rockets off to infinity in a finite amount of time?

That runaway is called a blowup, or a singularity. In two dimensions we know it can't happen. In three dimensions, the ones we live in, nobody knew. For two centuries.

It sounds like bookkeeping. It isn't. If the equations can manufacture an infinity out of nothing, then either our description of fluids is quietly incomplete, or something genuinely violent can happen in a fluid that we have never properly understood. Either answer would be worth knowing, and every simulation in every wind tunnel is currently trusting a thing nobody has checked. Mathematics is not usually comfortable living there.

Seven problems, a million dollars each

In 2000 the Clay Mathematics Institute chose seven problems it considered the most important open questions in mathematics and put one million US dollars on each. Navier-Stokes is one of them. A quick tour of three others, jargon removed:

• P vs NP. If a computer can quickly check that an answer is correct, can it also quickly find that answer? Sudoku is the kitchen-table version: confirming a finished grid is right takes a moment, filling in a nasty one takes an afternoon. Is that gap permanent, or are we just not clever enough yet? Essentially all online security is betting that the gap is real.
• The Riemann Hypothesis. Primes look scattered, but their long-run rhythm seems to be controlled by where a certain function equals zero. In 1859 Riemann guessed that all the interesting zeros sit on a single line. Still unproven, and a startling amount of modern number theory is written in the form "assuming Riemann is true, then...".
• The Poincaré Conjecture. The only one that has fallen. Loosely, if a three-dimensional shape has no holes in it, does that force it to be a sphere in disguise? Grigori Perelman proved it, posted the work online instead of submitting it to a journal, and then turned down both the million dollars and the Fields Medal. He thought the credit was being carved up unfairly and he walked away from research mathematics entirely.

The remaining three live in quantum field theory, number theory and geometry, and I'd be lying if I said I could make them bite-sized without mangling them.

Why this one touches everything

Of the seven, Navier-Stokes is the one whose answer would ripple furthest outside mathematics, and I think that's what makes it such a good story for students.

For mathematicians it's a test case for a whole family of equations where small effects feed on each other. A swirl stretches into a thinner, faster swirl, which stretches again. Whether that self-feeding process runs away forever or eventually gets damped out is the question, and the techniques built to attack it get reused all over analysis.

For physicists it is tangled up with turbulence, which Richard Feynman is supposed to have called the most important unsolved problem in classical physics. Turbulence is why you can predict the weather for a week and not a month. Small uncertainties get stretched and amplified until they swamp the forecast.

For engineers the practical version already bites. Simulations have finite grid resolution, and when a real flow starts concentrating energy into smaller and smaller structures, the grid stops being able to see it. Knowing whether nature can push that all the way to infinity, or whether something always steps in first, tells you what your simulation is allowed to miss.

And then there's the far end of the scale, which is my favourite bit. Gas spiralling into a black hole has a problem: it's in orbit, and orbiting things don't just fall in. To spiral inward it has to shed angular momentum, and it has to give that momentum to something. The answer turns out to be friction and turbulence inside the disk, which is a fluid dynamics question sitting in the most extreme gravitational environment we know of. The standard model of these accretion disks, from Shakura and Sunyaev in 1973, essentially bundles all that mess into one effective viscosity. In 1991 Balbus and Hawley identified a mechanism, the magnetorotational instability, where magnetic fields in the orbiting gas stir up exactly the turbulence required. When the Event Horizon Telescope produced those orange doughnut images of M87 and then our own galaxy's black hole, the way the pictures were interpreted was by comparing them against enormous fluid simulations, Navier-Stokes with magnetic fields and general relativity bolted on. So the same equations describing cream in coffee are how we read photographs of a black hole. I find that genuinely hard to get over.

So, this month

On 8 September 2026, OpenAI announced that an internal system had produced a counterexample to Navier-Stokes. Roughly ten thousand AI agents working in parallel for eighty-eight hours, after a warm-up of fifty hours on a related problem, producing a 166-page argument. And crucially, the whole thing was formalised in Lean.

Lean is worth explaining, because it's the part I'd tell a student to pay attention to. It's a proof assistant: you write your argument in a language so strict that a computer can check every single logical step, with no hand-waving and no "it is clear that". The Lean files behind this work run to millions of lines and add no extra assumptions. That means the internal logic genuinely is airtight, and anyone with a laptop and enough patience can verify it themselves. Machine-checked proof is a real shift, and it's arriving fast.

Here's where the headline and the mathematics part company.

The Clay problem is written with four alternatives, labelled A through D. A and B ask you to prove that solutions always stay smooth. C and D let you break them instead. The catch is that C and D permit you to apply an external force to the fluid, and the counterexample uses one: the flow starts at rest and is driven by a smooth force that is switched on in a limited region for a limited time.

Which is a completely legitimate reading of the written problem, and also not the question most mathematicians actually care about. The celebrated version is whether a fluid, left entirely alone, can tear itself apart using nothing but its own motion. Right now there is still a hand in the tank, stirring. OpenAI has said it does not intend to claim the million dollars, and Clay still lists Navier-Stokes as unsolved, though it has acknowledged the problem could have been settled and has opened its verification process. Their rules are slow on purpose: publication, two years of elapsed time, and general acceptance by the mathematical community.

There's a human tangle too. About twelve hours before the announcement, Tristan Buckmaster at NYU and Levent Alpöge at Anthropic published their own Lean-verified blowup results for forced versions of some closely related fluid equations, work they had finished in August. Both efforts build on a program developed by Diego Córdoba and Luis Martínez-Zoroa, who worked out how to make different length scales in a fluid feed off each other. Buckmaster has said publicly that he thinks Martínez-Zoroa deserves a Fields Medal for it, and has also raised pointed questions about whether OpenAI went down this road only after hearing about his unpublished work. OpenAI denies it. That argument is not settled and I'm not going to pretend I can settle it here.

Terence Tao, who has spent years on this problem, called the Buckmaster and Alpöge work a remarkable achievement, and has noted that nothing in principle seems to stop these methods from eventually reaching full Navier-Stokes. He also made the point I keep coming back to: mathematical understanding matters more than the scoreboard. A proof that no human comprehends is a strange kind of victory.

What I'd take from it

Not "AI solved maths". Something more useful than that.

The machines did something real here, and something narrow. They were enormously good at a grinding, technical construction where the path was roughly known and the work was brutal. They were pointed at that path by humans, standing on a program built by humans over a decade. Then a piece of software checked the logic, which is a thing we could not do at all a generation ago.

And the bit the headline flattened, the difference between a fluid that is pushed and a fluid left alone, is exactly the sort of distinction that decides whether you understand something or just recognise it. That's the same muscle as knowing why the quotient rule works instead of only knowing where it goes. Noticing that gap is the whole skill, in a first-year physics class and apparently also in a million dollar problem.

The coffee, meanwhile, is still doing something nobody can fully explain. Which I think is a lovely thing for two hundred year old equations to be up to.

If you want to poke at the mathematics underneath any of this, the rates of change in Calculus and the forces in Physics are where it starts, and there are free practice sets for both in our library.