I think it's important to occasionally state this, but LLMs are purely static and immutable sets of numbers and do not "learn" or change with use.
9
Replies
0
Boosts
@thomasfuchs yes, and a lot of what people see as them "learning" is just appending more crap to the input.
@thomasfuchs ...unless retrained, and that costs a heck of a lot of processing power
@csolisr That's a new LLM. ¯\_(ツ)_/¯
Yeah, basically! And you can't believe how much work has been wasted in building a LLM able to be partially retrained without starting from scratch.
Closest thing that's been made so far is to stick a new partial model to the current base one, but the base and the appendix must be specifically built to allow for it. Low-Rank Adaptation it's called (LoRA, not to be confused with the radio frequency one)
@csolisr it’s still not learning or changing with use. It’s more like you’re occasionally getting a new chapter for your textbook.
Closest thing that's been made so far is to stick a new partial model to the current base one, but the base and the appendix must be specifically built to allow for it. Low-Rank Adaptation it's called (LoRA, not to be confused with the radio frequency one)
@thomasfuchs I saw someone talking (with actual practical authority) about usability testing methodologies and they quickly pivoted to, "This is kind of like how LLMs reason... we can't quite see it." Like, no, those programs aren't "reasoning" at all. Similar shades of the "born" language... these are still very very much not people. Words matter, still.
@thomasfuchs this is so important to understand. and so painfully nuanced because all the harness vendors work SO hard to obscure this fact
