Former DeepMind researcher says LLMs lack true reasoning ability

On 2 October 2026, MIT Technology Review reported that a former Google DeepMind researcher argued large language models (LLMs) do not possess true reasoning ability. He points out that LLMs generate answers by predicting tokens one at a time, and their chain-of-thought reasoning still arises from the same next-token process rather than a genuinely separate reasoning mechanism. Such models lack an inspectable, persistent cognitive state, their knowledge and reasoning are not separated, and chain-of-thought reasoning is often fabricated after the fact. He advocates drawing on AlphaGo's game-tree search architecture to introduce an explicit cognitive state into machine reasoning. The view has now been published, and discussion is ongoing.

Written by AI from the reports · updated 14 d ago