Former Deepmind researcher thinks AlphaGo's playbook can cure AI hallucinations

Why it matters

Thore Graepel, a former Deepmind researcher who worked on AlphaGo, AlphaZero, and MuZero, has founded Metis Reasoning to tackle AI reliability. Presented at Nvidia's GTC conference in Berlin, the proposal adapts AlphaGo's search-based architecture to general reasoning: an external knowledge ledger records what a system knows, its confidence, and open questions, while a separate verification module checks each assumption against calculations, proofs, or empirical evidence. When an assumption fails, dependent conclusions are re-evaluated. Metis argues this could reduce hallucinations, produce source-cited explanations, and allow knowledge to be edited without retraining, potentially suiting smaller models in robots and edge devices. The company contends that scaling alone will not fix fabrication or unverifiable reasoning chains, which it blames for limited enterprise trust in domains like medicine, finance, and chip design. However, the framework remains theoretical, with no working system or benchmarks, and the team concedes real-world messiness, costly evidence, and the historical failure of hand-built symbolic systems as obstacles. Graepel joins other Deepmind alumni, including David Silver and Tim Rocktäschel, in founding ventures outside Google.