In an essay for MIT Technology Review, a researcher who helped build AlphaGo argues that the program's famous "creative" move against Lee Sedol was not a flash of intuition. It was the product of a search mechanism that explicitly constructed and evaluated a tree of possible future moves. That architecture, the author contends, is fundamentally different from how today's large language models work.
LLMs generate text by predicting the next token over and over, which the author likens to fast, associative System 1 thinking. Chain-of-thought prompting adds intermediate steps, but those steps are still produced by the same next-token process, not by a separate deliberative mechanism. The author identifies three shortcomings: no explicit, inspectable epistemic state; no separation between knowledge and the manipulation of that knowledge; and chains of thought that are often concocted after the fact.
The author argues that for high-stakes fields like medicine and science, we need systems whose reasoning can be audited—where we can see what the system believes, what it doubts, and how it reached a conclusion. That is why they left Google DeepMind to pursue a fresh approach based on AlphaGo's game-tree search, where the system maintains an explicit record of what it knows and what remains uncertain.