@ctjlewis
actually learn?
4 year olds will modify their behaviour in real time based on feedback, and eventually become 5 year olds
LLMs are frozen in time; they can’t genuinely learn things yet
@mitchazj
the chat models are obv constrained in that each time step is a new message, so it isn’t very high resolution input, but it is literally “real time”
and it can only age from a 0-chat context to an N-chat context, but it does in fact update on feedback in that time
@ctjlewis
> and it can only age from a 0-chat context to an N-chat context, but it does in fact update on feedback in that time
Do you mean the slightly longer input with each Nth message, up to the context window limit? That's not really learning new capabilities; just prompting up to…
@LiamPaulGotch
@mitchazj
@ctjlewis
These aren’t new capabilities, it’s rote memorization. If you gave a four year old a blue crayon, they would draw blue pictures, but they didn’t learn anything about how color works.