Ex-DeepMind researcher argues LLMs still lack auditable reasoning, calls for AlphaGo-style rethink
ThoreG · x · 2026-10-02
Former Google DeepMind researcher Thore Graepel marks the 10th anniversary of AlphaGo's Move 37 with an MIT Technology Review essay arguing that today's AI systems are missing something fundamental.
Key points:
- Move 37 wasn't intuition alone: AlphaGo could search possible futures, test its instincts, and reason about outcomes
- LLMs are remarkably capable, but longer chains of thought are not genuine reasoning — they typically lack an explicit, inspectable record of what they know, what remains uncertain, what evidence supports a conclusion, or whether real progress has been made
- This is why he recently left DeepMind, advocating a fresh approach to machine reasoning drawing on AlphaGo's architectural lessons
- For AI to produce trustworthy, genuinely novel insights in science and medicine, conclusions must arise from an auditable process of evidence, inference, and belief revision
More from AGI Musings
- A field guide to the confusing AI safety debate and its factions — marigo · 2026-10-02
- Artist on AI and Creativity: 'A Machine Doing All My Imagining Would Defeat the Point' — artBoffin · 2026-10-02
- Gary Marcus amplifies pointed quip on LLM consciousness: compute the next token by hand — GaryMarcus · 2026-10-02
- Miles Brundage: AI persuasion has a bad track record, reality does the convincing — Miles_Brundage · 2026-10-02
- Cloudflare launches Monetization Gateway to charge AI agents per request, settled in USDC on Base — kleffew94 · 2026-10-02
- Researcher: labs no longer sit on models for months before release — gleech · 2026-10-02