Thore Graepel Leaves DeepMind to Pursue AI Reasoning Beyond Move 37
Updated
Updated · MIT Technology Review · Oct 2
Thore Graepel Leaves DeepMind to Pursue AI Reasoning Beyond Move 37
1 articles · Updated · MIT Technology Review · Oct 2
Summary
Thore Graepel, a core AlphaGo architect, said he recently left Google DeepMind to build a new approach to machine reasoning modeled on the system behind AlphaGo’s famous move 37 in 2016.
His argument is that today’s large language models still operate as next-token predictors: chain-of-thought improves math and coding, but does not create a separate, auditable reasoning mechanism.
Graepel says current chatbots lack three essentials for scientific-grade reasoning—an explicit epistemic state, a clean split between knowledge and inference, and trustworthy traces of how conclusions were reached.
He proposes systems that maintain and update inspectable belief states, use tools and experiments to reduce uncertainty, and independently verify each reasoning step before revising beliefs.
The broader claim is that scaling intuition alone will not deliver trustworthy AI for medicine, science, climate or drug discovery; progress requires evidence-backed, inspectable reasoning more like AlphaGo’s search.