AlphaGo co-creator Graepel argues LLMs don't reason and leaves DeepMind to build it

机器之心 · wechat · 2026-10-04

AlphaGo co-creator Thore Graepel left Google DeepMind to start a company on machine reasoning, arguing in MIT Technology Review that LLMs don't reason. AlphaGo's famous move 37 came from explicit search over a game tree, not intuition — while LLMs are a hyper-trained System 1: chain-of-thought is still next-token prediction, with no inspectable cognitive state, no separation of knowledge from reasoning, and explanations often fabricated post hoc. His alternative: an explicit, auditable cognitive state where reasoning is actions that update beliefs, scored by how much uncertainty each step resolves — "the scientific method on steroids." In follow-up debate, David Duvenaud noted humans fail the same three criteria; Graepel said the trick is structuring the process, prompting the retort that this is exactly his startup's plan. He added frontier labs are all-in on scaling, and auditable reasoning won't emerge from scale alone.

Original post →

AlphaGo co-creator Graepel argues LLMs don't reason and leaves DeepMind to build it

More from AGI Musings