Google's AI solves 56-year-old math problems autonomously but DeepMind CEO says this is still not AGI

Google DeepMind’s AI system AlphaProof Nexus has autonomously solved nine open Erdős problems, some dating back 56 years, according to researchers. The system also proved 44 open OEIS conjectures, resolved a 15-year-old question in algebraic geometry, and discovered a new optimization parameter previously unknown to humans. Unlike earlier AI proof systems, AlphaProof Nexus operates independently, generating and verifying proofs using Lean, a formal verification tool, to eliminate logical errors. The breakthrough addresses a critical issue in AI-generated mathematics: hallucinations. Researchers noted that AI systems often produce flawed proofs by inventing false lemmas or avoiding core difficulties, which can evade human review. Google’s approach combines AI reasoning—powered by Gemini 3.1 Pro—with Lean’s strict verification, ensuring every step adheres to mathematical rules. This eliminates unsupported claims and reduces reliance on human validation. Human mathematicians collaborating with the system reported that even failed proof attempts improved their understanding of problems. Lean’s formal verification allows experts to focus on unresolved parts rather than rechecking entire arguments. The system’s efficiency is notable, with each proof costing only a few hundred dollars in computing resources. Despite these achievements, Google DeepMind CEO Demis Hassabis dismissed the idea that this progress signals artificial general intelligence (AGI). In an interview, Hassabis stated that current AI systems remain ‘nowhere near’ true AGI, emphasizing the gap between specialized problem-solving and broad cognitive capabilities. The work highlights both the potential and limitations of AI in advancing mathematical research.