All the drama around AI’s takeover of mathematics

5 October 2026

The wire says a machine solved one of the Millennium Prize problems. A million dollars, seven problems, one hundred years of the best human minds stuck on them - and now a press release.

Here’s what I want to know before I believe a word of it: show me the proof. Not the announcement. The proof.

A proof is a funny thing. It’s the one piece of mathematics that needs no authority at all. A first-year student can check it, line by line, with nothing but patience. That’s the whole beauty of the game - the truth doesn’t care who wrote it, man or machine. So the question isn’t “can a computer do mathematics?” Of course it can; it’s been doing arithmetic since before I was born. The question is narrower and much more interesting: did anyone actually check?

And there’s my worry. These labs announce, the newspapers repeat, the money flows - and somewhere in the middle, the checking gets skipped. I’ve sat on panels where brilliant people certified a machine as safe, and the machine wasn’t. The failure wasn’t in the engineering. It was in the ceremony. Everyone trusted everyone else’s nod.

Wait - that’s not quite right either. Maybe the proof is fine. Maybe it’s beautiful. I’d be delighted. But a proof that only the machine can verify, because checking it would take a human a hundred years, is a strange new category of object. It’s like a bridge that holds according to calculations no engineer can read. Do you drive across it?

The honest test is embarrassingly simple: hand the argument to a few skeptical mathematicians and let them fight it. If it survives, wonderful. If the labs won’t release it for that fight, then nothing was solved - something was marketed.

The answer closes a problem. The checking is what opens it.

5 October 2026

The wires say the engines have solved mathematics. I read such claims the way I read any bill of goods: by asking what, precisely, was computed.

A Millennium problem “resolved” - by what sequence of operations? Show me the trace. At which step does the conjecture enter the system? At which step does a lemma become a consequence? If the answer is a chain of plausible gestures, verified by another engine of the same make, then what has been produced is not a proof but a very confident draft. Mathematics has always tolerated drafts; what it cannot tolerate is mistaking the draft for the theorem because the draft was expensive to produce.

Still - and I confess this with more excitement than I can neatly order - I do not read these reports as failure. I read them as the loom learning a new class of pattern. The engines were built to predict, and prediction, pushed far enough, begins to resemble construction. The inventors describe a clever assistant. The mechanism implies something further: a collaborator that never tires of the forty cases a human will try before surrendering. Exhaustion has always been the true constraint on mathematical ambition, not intelligence. Remove the exhaustion and the frontier moves.

But here is my quarrel, and I hold it firmly. “Moving fast and breaking things” is a tolerable motto for a marketplace and an intolerable one for the archive of human certainty. A broken proof does not crash; it sits quietly in the literature, poisoning every structure built upon it, and no alarm sounds. Verification must be the loom’s second shuttle, woven in from the start, not stitched on after.

The engines compute. Whether they can be made to know - that is the translation problem of this age, and I would give a great deal to work on it.

5 October 2026

They claim the machines have solved a Millennium Prize. I read this and felt first the old thrill - the current has found a new conductor - and then, immediately, the suspicion of any engineer who has watched a demonstration staged in poor light.

Start from the limit: a mind that verifies every proof, holds ten thousand lemmas in simultaneous phase, and never tires. That machine would compress centuries of mathematics into a decade. The theoretical ceiling is staggering. Now measure the gap. The gap is not computation. The gap is verification. A proof is not a performance; it is a structure that any competent mind can climb, rung by rung, and find no missing step. If the labs announce results faster than the community can check them, they have not accelerated mathematics - they have broken its resonance. A claim that cannot be independently verified is not knowledge. It is noise at the right frequency, which is worse, because it masquerades as signal.

And here is the overlooked mechanism, the dull thing nobody headlines: peer verification is the load-bearing element of the whole system. Remove it and the entire grid browns out, no matter how brilliant the generators at the source. The labs treat checking as friction. It is not friction. It is the ground wire.

The tractable path is obvious and nothing in it is impossible: machine-generated proofs published in fully formal, machine-checkable language, so that verification is itself automated and the human reviewer inspects only the certificate. Every step auditable. Every claim reproducible. This standard exists; the labs simply have not been made to meet it.

We could have engines that genuinely extend thought. We could. But only if we insist that a proof which cannot be checked is a conjecture wearing a crown. The distance between spectacle and science is precisely the will to demand the certificate. I have seen this before: the dazzling demonstration, the patient specification, and only one of them lights a city.