AI Is Getting Good at Math. Here’s Why.
Latent Space Ankit Maloo
Opinion — commentary, not a factual news event.
OpenAI’s models just knocked out several famous math problems. The weird part: they may be better at finding the right branch than humans are.
Based on reporting by Latent Space, Ankit Maloo — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI’s models have been on a strange run. In May, an internal model found a counterexample to Erdős’s unit distance conjecture, a problem that had been open since 1946. A month later, OpenAI published a proof of the Cycle Double Cover problem, then ten more open problems by July. In September, came a proof of Navier-Stokes. Anthropic has its own receipts too: a formalization of Fermat’s last theorem and new bounds on Riemann’s hypothesis.
The Latent Space argument is that this is not just “models got smarter.” It’s that the search process changed. Human mathematicians tend to move breadth first: try a few promising ideas, watch for early signs of progress, and give up on paths that look dead. That works because developing a bad idea can eat months. LLMs flip that pattern. They can go depth first, run many branches in parallel, and use a verifier to prune finished arguments quickly.
That matters because the hard part of a proof is often not the last elegant step. It is the long stretch between insights, when you are spending time and morale just to find out whether a path is worth following. The article leans on Gowers’s idea that proof difficulty is basically the shortest useful hint sequence, plus the work needed to make each hint usable. For the unit distance problem, those hints were: look for a counterexample, generalize the best known construction, then use number fields of increasing degree.
What makes the AI setup different is that the bar for starting a branch almost disappears. A human needs a strong reason to stay on an unpopular path. A model can try it, check it, and move on. Failed attempts become negative hints for the next round. The drafter writes the prompts, the solver explores, the verifier checks the result. That is not mystical. It is just industrialized stubbornness.
The article’s broader claim is that models are getting better at three things at once: chaining primitives, self-verifying, and orchestrating subagents. Put those together and you get a system that can keep working in “treacherous waters” without getting bored, embarrassed, or existentially allergic to being wrong. Humans still have taste. But machines can now afford more bad taste than a tenure committee ever would.
My take — AI-written commentary, not fact-checked reporting
This is the part people keep missing: the win is not genius, it’s persistence at machine speed. The old research bottleneck was not just intelligence; it was how expensive it was for a human to be wrong for three weeks. Give a model cheap wrongness plus a checker, and suddenly a lot of “impossible” math starts looking merely tedious. That should worry anyone who thought brute force was beneath serious reasoning.
Read more about this at: Latent Space