Google DeepMind AlphaProof Nexus solves 9 Erdős problems

AlphaProof Nexus is one of the clearest examples yet of frontier AI models moving from benchmark demos into tools that can do real, verifiable scientific work. By chaining a large language model with the Lean proof assistant, DeepMind isn’t just spitballing solutions – it’s generating proofs that survive formal checking and resolve long‑standing Erdős problems and dozens of conjectures. That’s qualitatively different from earlier “math‑capable” models that produced pretty but unreliable arguments.

For the race to AGI, this matters on two fronts. First, it’s a concrete step toward AI systems that can explore highly structured problem spaces with tight feedback loops, a pattern that generalises well beyond pure math into software verification, hardware design and safety‑critical reasoning. Second, DeepMind is explicitly framing the result as “still not AGI,” signalling a deliberate effort to keep expectations grounded while still showcasing a major capability jump. That positions Alphabet as a serious contender in the emerging ‘reasoning plus verification’ arms race alongside OpenAI and others, and underscores that formal methods may be a key ingredient in safe, scalable advanced AI.