Essay says verifiers let AI search hard math problems depth-first
Original title: AI Is Getting Good at Math. Here’s Why.
The author contrasts months-long human search with models that can test weak hints in minutes if a checker rejects false proofs.
A Latent Space essay argues that recent OpenAI and Anthropic math results come from cheap test-time search plus automated verification, not from human-style insight. It cites an internal OpenAI counterexample to Erdős's unit distance conjecture, later claimed solutions to other open problems, and Anthropic work on Fermat's Last Theorem and the Riemann hypothesis. The author uses Timothy Gowers's hint-sequence idea to explain why models can pursue implausible directions that humans abandon.
Unconfirmed: The solved-problem claims are the author's account; the excerpt does not independently verify the proofs.
- The unit-distance counterexample is described as found in May by an internal OpenAI model.
- Gowers's proposed complexity is written as 2^k + t, with k the shortest hint sequence.
- The essay says chain-of-thought search is greedy and depth-first without a verifier.
- A checker can only judge finished arguments, not half-developed ideas.
From the original article
15 min read at latent.space
AI models are solving more mathematical problems. A pattern is emerging in their successes that could help us predict which problems they’ll solve next.
By Ankit Maloo · latent.space
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.