Ex-DeepMind researcher: LLMs don't reason—future AI needs AlphaGo-style deliberation
MIT Tech Review AI · rss · 2026-10-02
Writing in MIT Tech Review, a researcher who recently left Google DeepMind argues that today's LLMs don't truly reason, using AlphaGo's famous move 37 as the lens.
Core argument
- Move 37 wasn't machine intuition: AlphaGo's policy network gave it 1-in-10,000 odds of an expert playing it; what chose it was explicit tree search weighing thousands of future branches—Kahneman's System 2 deliberation, whereas next-token prediction is pure System 1.
- Chain of thought yields real gains in math and coding, but the intermediate steps are still produced by the same next-token process—no separate reasoning mechanism.
Three reasons LLM output isn't reasoning
- No explicit, persistent, inspectable epistemic state (hypotheses, confidence, open questions).
- Knowledge and reasoning are inseparably woven into the same weights.
- Research shows models often fabricate reasoning chains after the fact.
Proposed direction
The author calls for AlphaGo-style architecture: a system that maintains an explicit epistemic state (what's settled, doubted, ruled out, open) and reasons via moves that reduce uncertainty, with an independent component accepting only evidence-backed belief updates—a "scientific method on steroids." Scale alone won't make System 1 trustworthy; only such reasoning systems can deliver novel insights in drug discovery, materials, climate, and diagnosis.
More from AGI Musings
- AI systems lack executive function, not scale: four manager-layer strategies to fix it — Designer-Material633 · 2026-11-19
- LeCun refuses doomer faction letter, enjoys doomers infighting — ylecun · 2026-11-19
- AI governance scholar Luiza Jarovsky calls for a global ban on RSI — LuizaJarovsky · 2026-11-19
- Stanford AI100 Releases Third Report on the Scaling Era and Rise of Generative AI — StanfordHAI · 2026-11-19
- Interpretability solved—then what? Najoung Kim proposes filtering for recurrent constructs — najoungkim · 2026-11-19
- Elicit on agent swarms: pharma winners of 2036 will robustly deploy them without going rogue — jungofthewon · 2026-11-19