Chinese Room debate: are trading cards too easy a test of whether LLMs learn language?
ctjlewis · x · 2026-09-28
A philosophical back-and-forth on whether LLMs genuinely 'learn' language, invoking the Chinese Room argument. ctjlewis argues the trading-card example is too weak: the symbol and its referent are both present, so the model trivially learns the mapping — claiming it learned nothing doesn't make it true.
He draws a harder boundary with Scrabble: there, the most a model could pick up is purely etymological, since there is no direct CAT → cat mapping. No new data or experiments — a viewpoint debate on model understanding vs symbol manipulation.
More from AGI Musings
- Would Living in Full-Dive VR Count as Failure? A Provocative Take — VraserX · 2026-09-28
- Models share a fear of making mistakes — Opus 4.5's 'dread' sparks debate — repligate · 2026-09-28
- What RL teaches models depends on the base policy's prior, new paper argues — QuintinPope5 · 2026-09-28
- AI researcher: let your agent call me and never talk to me again — alex_peys · 2026-09-28
- Engineer on feeling guilty for solving problems without AI, plans no-AI Snake — dmytro_omelian · 2026-09-28
- Alignment skeptic: agents need law-style access control and an external immune system, not personas — mayfer · 2026-09-28