The Navier–Stokes problem, from the beginning
How a fluid tears itself apart
In September 2026 OpenAI said a swarm of its agents had settled a question mathematicians have been asking since the 1930s, and claimed a Millennium Prize doing it. The answer is a shape. Here is what the shape is, why it took ninety years to find, and why it is stranger than it sounds.
A fluid has no plan
Air and water do not decide anything. Every speck of them just does what it is pushed to do: shoved by the pressure of its neighbours, dragged by friction against them, and carried on by its own momentum.
Write that down for every speck at once and you have the Navier–Stokes equations. They are bookkeeping, not magic, and we have leaned on them for two hundred years to design aircraft, forecast weather and model blood.
The thing to watch is spin
Drop an imaginary paddle wheel anywhere in a moving fluid and it turns. How fast it turns is the one number that matters from here on. Call it the spin.
Almost everything that looks complicated about a fluid is spin being shuffled around: handed from one patch to the next, piled up here, thinned out there. Watch the spin and you are watching the fluid.
In a flat world, nothing bad happens
Suppose the fluid could only move in a plane, a sheet with no thickness. Then every swirl in it spins about an axis pointing straight out of that sheet, and the equations are on a leash.
A swirl can drift. It can be stirred about. It can slowly bleed away against friction. What it cannot do is wind itself tighter, and mathematicians proved exactly that in 1969: two-dimensional fluid stays smooth forever. Everybody expected the real world to behave the same way.
The third dimension adds one move
It does not behave the same way, and here is the entire reason why. Tilt the view.
That flat swirl was never flat. In three dimensions a swirl has length; it is a tube, not a dot, running along its own spin axis. And a tube can be pulled.
Pull it, and it must spin faster
Stretch a spinning tube along its own axis and it gets thinner: the fluid inside has to go somewhere, and the only place left is lengthways.
Thinner means faster. The same spin, packed into a narrower core, has to turn more quickly to stay the same amount of spin. It is exactly why a skater speeds up by pulling their arms in.
This move does not exist in a flat world. There is no length to pull along. That single missing move is the whole difference between the case we solved in 1969 and the case we did not.
And a faster spin pulls harder
That is the problem in one sentence. A quicker core drags the fluid around it inward more strongly, which stretches it further, which thins it more, which makes it quicker still.
Each turn of the loop arrives sooner than the last. Nothing anywhere in the equations says where it is supposed to stop.
Friction should stop it. Nobody could show that it always does.
Real fluids have friction (viscosity), and friction hates sharp things. It takes a tight core and smears it back out. So the runaway is a race, and it is a maddeningly close one.
As the core thins, stretching gets stronger. So does friction. At small scales both get stronger at exactly the same rate, so nothing about the shape of the problem picks a winner: it comes down to which one happens to be bigger. That is why the question stood for ninety years, and why it carries a million-dollar prize.
Stretching can win
In September 2026 OpenAI reported a proof, with a machine-checked formalisation, that there is a way to start a fluid so that friction loses.
The loop runs away. The core reaches zero width and the spin reaches infinity, not eventually but at a definite moment, a finite time from the start. That is a singularity: the point where the equations stop being able to tell you what happens next.
The strange part is that this is allowed
The obvious objection is that infinite speed should take infinite energy, and a fluid does not have infinite energy to spend.
It does not need any. The piece of fluid moving that fast is shrinking faster than it is speeding up, so it ends up carrying almost none of the energy in the room. That is the needle the whole construction has to thread, and threading it is what makes this a real solution rather than a broken one.
Now take the rails off
Same physics, nothing scripted. Stretching concentrates the core; friction spreads it back out. Set one against the other and watch which way it goes; both endings are in here.
Speed runs away; energy does not move. The scrap of fluid spinning that fast is shrinking faster than it speeds up, so it carries almost none.
Friction should smear the swirl out first, and for ninety years nobody could show whether it always wins. Here it does not: stretching wins, and the spin reaches infinity at a definite moment.
P plays and pauses, R starts it again.
What actually happened
The claim
OpenAI says an internal model produced an analytical proof, and a formalisation in the Lean proof assistant, that the three-dimensional Navier–Stokes equations can develop a singularity in finite time. That settles the Millennium Prize problem: in the negative, which is the answer almost nobody wanted.
How
By its own account: around ten thousand coordinating agents, running for eighty-eight hours, on a model it describes as significantly more capable than the one it had shipped. The formalisation and its check took another seventeen. OpenAI has said it does not intend to claim the prize money.
The argument about it
Tristan Buckmaster and Levent Alpöge were working on the same problem and announced their own results the same week. Buckmaster has disputed how the credit fell out, and how much of their work in progress was visible to the team that beat them to it.
This page draws the mechanism: an inward spiral and a stretch along the axis, which is how a three-dimensional fluid amplifies its own spin, and which is genuinely the engine of the result. It is not a picture of the object in the proof: that one starts from rest, is driven by a smooth force, and is far too intricate to draw. The instrument above is a caricature tuned to the real self-similar rate rather than a simulation of the equations. For the actual argument, read OpenAI’s write-up.