A Very Unlikely Chess Game

Scott plays (and narrowly wins) a sloppy chess game against GPT-2 — a text-prediction model with no concept of chess, trained on PGN notation, that learned passable openings but loses track of board state ~move 13. The point: the same model Gwern taught to write poetry and (via ABC notation) folk-song music learned to play chess without being designed for any of it — supporting his earlier 'GPT-2 as a general pattern-recognizer with text I/O' thesis. Presciently (Jan 2020) gestures at LLM generality; ends 'How impressed should we be that the same AI can write poems, compose music, and play chess? I still don't know.'

Quality 67 · Strong. Strong: a fun, notable, prescient early demonstration of LLM generality (the generality-from-text-prediction point aged extremely well), but a light, reportorial writeup (game narration + quoted tweets) rather than a developed essay. 67.

Claude’s paradigm shift 50 · Moderate. An early, prescient demonstration of LLM generality; the thesis is from his prior GPT-2 post. B50.

Real-world impact 2 · Minor. A notable early GPT-2-generality demonstration; no material reach. RWI2.