logoalt Hacker News

glassresearch • today at 3:43 AM • 1 reply • view on HN

This is true for a large subset of problems. LLMs are bad at basically all games, for instance. LLMs are even bad at chess, which has a huge amount of publicly available data. LLMs are really bad at games like backgammon, which has only 10s of millions of tokens of public data.


Replies

MichaelZuo • today at 3:44 AM

If it’s thought of as extremely extremely sophisticated auto-complete, then that wouldn’t be surprising at all.

➕ show 1 reply