OpenAI says one of its general-purpose models disproved a conjecture tied to Erdős's unit-distance problem, a question mathematicians had circled for decades. Tim Gowers, a Fields medalist, called it a real result. Hacker News spent the day arguing about whether this is the moment math changed. It isn't, quite — but it's worth being precise about what actually happened.
The model didn't prove a theorem. It found a counterexample. That distinction is the whole story, and it cuts against the hype, not for it. A counterexample is a single object that breaks a universal claim. You don't have to trust the search that produced it — you check the object, and the conjecture falls.
Disproof is cheap to verify, and that is exactly why we can believe this one.
Proof is the hard direction. It demands an argument that holds for every case, and the argument is the part we'd actually want a machine to understand. Falsification is the searchable direction. The space of these configurations is enormous, far past what a person can hold in their head, and a model that can move through it at scale and surface the one arrangement that violates the bound is doing something genuinely useful. Useful is not the same as insight. It located a needle. It did not explain the haystack.
The serious objection: this wasn't brute enumeration. The space is too large to grind through, so the model had to work it with something that resembles mathematical taste, and Gowers — who does not hand out praise lightly — was impressed by the path, not only the result. Granted. The path it took is real, and it is new. But taste in service of search is still search, and the reason we trust the output is the cheap check at the end, not the model's reasoning along the way.
That's the practical lesson, and it beats the headline. Point these systems where the answer is hard to find and easy to verify — counterexamples, bugs, edge cases — and they earn their keep. Ask them to be right where no one can check, and you're back to faith. The machine found the needle. Someone still has to certify it's iron.