@ElmoTheHomie @pointed_max @fentpot or approximations ie plausible abstractions work well most of the time but fail at extremes? might be useful for some things to have 80/20 understanding of what's going on, but not for guaranteeing behavior wrt edge cases
@ElmoTheHomie @pointed_max @fentpot like, I think anthropomorphizing LLMs is a useful mental model in many ways for prompting bec their training data stems from human-human interactions - but taken too literally ofc it is very wrong & misleading to think of a chatbot as a human mind.
More notes on nearby ideas
Reply · Jul 20, 2023I certainly agree LLMs are hitching a ride on top of the abstraction layer of human language... but I'm not sure how much this matters for "real…Tweet · May 22, 2023imo thinking abt AGI as "alien" or a "random distribution" is kinda misleading... LLMs are trained via the *human* abstraction layer of language, on…Reply · Sep 29, 2023@deepfates lol... I really DO think LLMs will become superhuman empaths (read: manipulators) given a bit more timeReply · Nov 3, 2023@techn0_sap1en I feel like it's one of those things that shouldn't be surprising, that any conversational pattern that is more frequent / functional…