Google DeepMind's AlphaProof Nexus cracks math problems unsolved for 56 years

Google DeepMind has launched a new AI system called AlphaProof Nexus. It combines large language models with a formal verification tool called Lean. The system solved 9 out of 353 open Erdős math problems, including 2 that had been unsolved for 56 years.

It also proved 44 out of 492 open conjectures in the OEIS database, solved a 15-year-old Hilbert function problem, and improved known bounds in convex optimization. Each problem cost only a few hundred dollars to solve.

AlphaProof Nexus uses four AI agents with increasing complexity. The simplest agent, which only uses Gemini and Lean, could already solve the 9 problems. This shows that better base models and compiler feedback are making AI math research more effective.