AI is solving old math problems, and a pattern explains why

An OpenAI model disproved a math conjecture that had been open since 1946. That is one of several recent results, and a writer at Latent Space argues there is a pattern behind them. AI does best where a fast, automatic check can tell right answers from wrong ones.

What happened

In May, an internal OpenAI model found a counterexample to Erdős's unit distance conjecture. Many experts had tried to prove it true. The model showed it was false instead. The article says OpenAI then published a proof of the Cycle Double Cover problem a month later, and solutions to ten more open problems by July. It also reports a Navier Stokes proof in September. Anthropic is credited with a formalization of Fermat's Last Theorem and advanced bounds on the Riemann hypothesis.

Why humans skip ideas

The article describes solving a hard problem as a chain of insights, with long stretches of work between them. Developing one idea can take a researcher months. So humans only commit to ideas that show early promise, and they drop most others. Ideas that look unlikely never get tried.

What AI does differently

A model can run many attempts at once. Left alone, that produces plenty of convincing but false proofs. The fix is an automated checker that tests each finished attempt and keeps only the ones that hold up. The article says you can learn within minutes whether an idea works.

That changes the math. When a test takes minutes, an idea no longer has to look promising first. It can be ordinary, or even implausible, and still be worth a try. That may be why a model could find a counterexample that experts did not think likely.

The catch

The checker only judges finished arguments. It cannot say whether a half-built idea is heading anywhere. The pattern is also one writer's explanation, not a proven rule for predicting what AI will solve.

What it means for a business

The same logic applies outside math. Tasks with a quick, automatic way to confirm an answer are where AI is likely to improve first. Tasks where judging the result is slow, vague or a matter of opinion are harder to hand over. Asking how quickly you could check an AI's work is a useful way to pick where to try it.