Post

"That only plausibly looks like a chess move" I sneer as Stockfish absolutely destroys me

Bram Spookmund@brambrockmund.bsky.social

Not to dunk, but this is anti-anthropomorphising ad absurdum. LLMs don't output "something that plausibly looks like a sentence", they output sentences. Torturous phrasing doesn't serve clarity.

Maybe this is a circular argument, but I think "looking like a chess move" is sufficient for actually being a chess move more than the same for sentences. (Though sometimes "looking like a sentence" is adequate for the task at hand)

Short answer: chess is code Chess is not, fundamentally, expressive. A player may have a certain style as with anything else, but the primary thing determining a chess move is analysis of the position, and producing coherent language is more than that

I agree there is more expressiveness in human language than in chess. But I'd say that is separable? That is, a chess move is just syntax, and a sentence is just syntax as well. On top of that syntax, there can be meaning and interpretation, and certainly a lot more in the case of human language.

To deny that LLMs even emit *sentences*, as in that quote, just strikes me as bizarre. They do emit sentences, and even grammatically valid ones.