A Machine Found Where the Equations Explode — But Finding Isn't Proving
June 21, 2026 · George Polya, How to Solve It~6 min read
Picture stirring cream into coffee. Now picture the swirl getting so violent, so concentrated, that at one single point and one single instant the speed of the fluid goes — not large, but literally infinite. Mathematicians call that a blow-up, a finite-time singularity, and whether the equations of fluid motion can do it from a perfectly smooth start is one of the seven Millennium Prize Problems, each worth a million dollars and most of them older than anyone alive. This spring, something quietly remarkable happened: a machine learned to find these explosions. The catch is the oldest one in mathematics — finding is not proving — and it turns a flashy result into a much more interesting story.
The equation nobody can tame
The Navier–Stokes equations describe how fluids move: air over a wing, blood in an artery, weather across a planet. Engineers use them every day and they work beautifully. Yet no one has been able to prove that their solutions always stay smooth and finite — that the math never, under any starting conditions, spits out an infinity. That sounds like a technicality. It isn't. An infinity in the answer would mean the equations break down exactly where the physics gets most interesting, and we wouldn't know whether the failure is in nature or only in our description of it. The honest status, as of mid-2026, is stark: for the full three-dimensional problem the question is open. It has been open since 1934. The prize money has gone uncollected for a quarter of a century.
The machine that learned to find explosions
Here is the new part. Rather than wait for a flash of human insight, several teams have pointed neural networks at the problem — not to chat, but to compute. In late 2025, a group including Tristan Buckmaster and Ching-Yao Lai reported that their networks could resolve the precise shape of these singularities — the so-called self-similar blow-up profiles — to machine precision, and even fish out the rare, unstable ones that older methods drowned in their own numerical noise. In plain terms: the machine can now point at the exact spot where an idealized fluid would tear itself to infinity, and draw its portrait with extraordinary sharpness. A generation ago that picture was a fog. Now it is a photograph. It is, genuinely, a new instrument.
Neural networks can now pinpoint where idealized fluid equations explode — self-similar blow-up profiles resolved to machine precision (Wang, Léger, Lai & Buckmaster, late 2025). But a found profile is a plausible guess; a rigorous computer-assisted proof must still close the gap. Navier–Stokes blow-up remains 1 of 7 unsolved Millennium Prize Problems; a March 2026 proof for a related case (Shkoller, per Scientific American, June 2026) looks promising but uses shortcuts the full prize won't allow. Framework: George Polya, How to Solve It. A reflection, not a verdict on unverified proofs.
Where Polya clears his throat
And here, across seventy years, George Polya leans in. His little 1945 classic How to Solve It breaks every problem into four phases, and the one people skip is the last: looking back. Polya's sharpest warning lives there. Seeing that something is true, he insisted, is not the same as showing that it is true. A plausible conjecture — however beautiful, however much the picture begs you to believe it — is exploration, not proof. The neural network has done something Polya would have loved: it has guessed brilliantly. It has done the second phase, devising a plan, with superhuman precision. But the third and fourth phases — carry it out, then check every link in the chain — those it cannot hand you. A picture that is right to machine precision can still be wrong in a way no amount of precision detects, because precision and proof are different animals entirely.
So what actually got proven?
This is where care matters, and where the headlines get slippery. In March 2026, the mathematician Steve Shkoller posted to arXiv a proof — over a hundred pages of dense argument — that addresses a related blow-up question, and early reactions from his peers have been warm. "No one thought it would be possible," one mathematician at NYU was quoted saying. But two honest caveats hold the line. First, the proof reportedly uses shortcuts that the full Millennium Prize version won't allow; it is a real result on a nearby problem, not the prize itself. Second, a hundred-page proof takes the community many months to verify line by line — and until that's done, the correct word is promising, not proven. That gap between posted and confirmed is not bureaucracy. It is the looking-back phase, running at the scale of a whole field.
What this means for you
You don't have to care about fluid equations to use the lesson, because the shape of it is everywhere now. We have all just been handed machines that produce confident, polished, plausible answers at a speed no human can match — a medical hunch, a legal summary, a line of code, a market call. The fluid-dynamics story is the cleanest possible parable for how to hold them: treasure the guess, then refuse to skip the proof. When a machine hands you an answer, the right question is never "does it look right?" but "what would show me it's right — and have I done that step, or only admired the picture?" Polya's discipline was never about distrusting brilliance. It was about knowing which phase you're standing in. The machine has gotten extraordinary at finding. Proving — the slow, checkable, link-by-link kind — is still the part that's on you.
A machine can now point at the exact place the equations explode and draw it to machine precision. None of that makes it true. Precision is not proof, and the picture isn't the theorem.
Treasure the guess. Then do the part the machine can't hand you: check every link in the chain.
Source: framework from George Polya, How to Solve It (the four phases of problem-solving; the caution that a plausible conjecture is not a proof, and "looking back" means verifying, not admiring). Real-world basis: the 2026 push to use neural networks to find self-similar blow-up profiles for fluid equations to machine precision (Wang, Léger, Lai & Buckmaster, arXiv, late 2025), and a March 2026 arXiv proof on a related blow-up question (Shkoller), reported by Scientific American, June 2026, as promising but not yet community-verified and reliant on shortcuts the full Navier–Stokes Millennium Prize won't allow. A reflection, not a verdict on any specific unverified proof.