Don't be fooled–LLMs don't reason
4 hours ago
- AlphaGo's creative move 37 in the 2016 match against Lee Sedol was a product of reasoning through game tree search, not mere intuition or probability.
- Current large language models (LLMs) rely on fast, associative pattern completion (System 1) and lack genuine deliberation, as chain-of-thought still uses the same next-token prediction process.
- Three key shortcomings prevent LLMs from achieving true reasoning: no explicit, inspectable epistemic state; no separation between knowledge and manipulation; and tendency for post-hoc reasoning that does not reflect actual deliberation.
- Future AI systems need architecture inspired by AlphaGo, with an independent epistemic state and evaluation mechanism to produce trustworthy, auditable conclusions in high-stakes fields like medicine and science.
- Scale alone cannot turn intuition into reasoning; systems must explicitly represent beliefs, weigh evidence, and revise knowledge through a sequence of deliberative moves.