DeepMind proof agent solves 9 Erdős problems

The important shift here is not “AI beats mathematicians,” but “verification turns probabilistic reasoning into something that can actually survive contact with formal math.”

• –Solving 9 of 353 open Erdős problems is still a small hit rate, but every win is formally checked, which is a much higher bar than most benchmark math claims
• –The paper says a basic LLM-plus-Lean loop replicated the Erdős results, which suggests the search/verification scaffold matters as much as model intelligence
• –44 solved OEIS conjectures broadens the signal beyond one math niche and points to a reusable theorem-search pattern
• –The cost profile matters: a few hundred dollars per problem is the kind of number that makes this usable as a research assistant, not just a demo
• –For developers, the takeaway is broader than math: any domain with a strong validator could benefit from the same agent loop