Ex-DeepMind researcher proposes AlphaGo-inspired method to curb AI hallucinations

Thore Graepel, a former DeepMind researcher, is advocating for a new approach to reduce AI hallucinations by drawing inspiration from AlphaGo. He argues that large language models alone cannot solve this issue and proposes adding an explicit, verifiable, and editable knowledge ledger where every conclusion is backed by evidence.

Metis Reasoning, Graepel’s startup, presented this concept at Nvidia’s GTC conference in Berlin. The method relies on two neural networks: one proposes "moves"—questions, experiments, or challenges to hypotheses—to reduce uncertainty, while the other evaluates these proposals using calculations, formal proofs, or empirical data. The system, akin to a game board, logs what is known, confidence levels, and open questions, automatically re-evaluating conclusions if an assumption is invalidated. Microsoft (GraphRAG) and Google (AlphaGeometry2) are exploring similar ideas, but Metis aims to generalize this to complex real-world domains.

The project remains theoretical, with no working prototype or benchmarks available. Challenges include handling partial information, the cost of obtaining evidence, and cases where no answer exists. Graepel, a co-author of AlphaGo, highlights that the untraceable errors of current AI models limit their adoption in sectors like medicine or finance, where verifiability is critical. He is currently raising funds to develop this approach.