Former DeepMind Scientist Says LLMs Lack True Reasoning

Thore Graepel, a former core member of the AlphaGo team who recently left Google DeepMind, is calling for a fundamental shift in how AI systems are built. Writing about the famous March 2016 Go match where AlphaGo defeated Lee Sedol 4-1, Graepel highlights move 37 in game two of the five-game series. While commentators thought the move was a glitch, it was actually a product of AlphaGo's unique architecture. Unlike Deep Blue, which defeated Garry Kasparov in 1997 by evaluating 200 million positions per second and looking six to eight moves ahead, AlphaGo combined a policy network with a search tree. This search mechanism evaluated future consequences, selecting a move that its intuitive network estimated had only a 1 in 10,000 chance of being played by a human.

Graepel argues that today's large language models completely lack this deliberative reasoning capability. While techniques like chain of thought help chatbots decompose problems, they still rely entirely on next-token prediction. This process lacks an explicit, inspectable epistemic state to track hypotheses and evidence. Furthermore, LLMs do not separate their knowledge from how they manipulate it, and they often generate post-hoc explanations that do not reflect how they actually reached an answer.

To build trustworthy AI for high-stakes domains like medicine and engineering, Graepel suggests developers must move beyond simply scaling up next-token prediction. Instead, future systems should mimic AlphaGo by maintaining an explicit record of what they know, doubt, or have ruled out. By using an independent evaluator to update these beliefs only when backed by evidence, practitioners can create auditable, self-improving systems capable of genuine scientific discovery.