Thore Graepel, chair of machine learning at University College London and a former core member of the AlphaGo team at DeepMind, argues that current large language models (LLMs) lack the genuine reasoning capabilities required for high-stakes applications like medicine and scientific research.
LLMs Lack Genuine Reasoning Capabilities Compared to AlphaGo, Expert Argues
While many perceive LLMs as deliberative due to techniques like chain of thought, Graepel contends they remain primarily "System 1" thinkers—fast, associative, and focused on next-token prediction. Unlike AlphaGo, which utilized a search machinery to weigh future consequences and build a game tree, LLMs do not maintain an explicit, inspectable epistemic state. Instead, they often concoct reasoning steps after the fact to justify a pre-determined answer.
Graepel suggests that true machine reasoning requires a system that can maintain a record of its knowledge, evaluate how new information reduces uncertainty, and update its beliefs based on evidence. He posits that simply scaling current models will not bridge this gap, as larger scales merely sharpen pattern recognition rather than introducing deliberative reasoning.
Sources
- Don't be fooled–LLMs don't reason （Hacker News Frontpage, 2026-10-02）