>No, that's not right. Nobody programmed the LLM how to respond, it is just based on training data. It is emergent behavior.

while you're right, i do think it's a matter of clarifying and discretely organizing training data. there's a reason data management has been an emerging tech juggernaut in the last decade. there may be a plateau there somewhere but i don't think we've reached it yet.

my guess is we'll soon have different "modes" of translation and interaction plus a suite of micro-genre, very specified neural networks like a purely medical one for example. making data segregation easier with the added bonus that some of them are varied in when they need retraining. a subsciption program with small micro-transactions to access various genre of neural networks would be the tech-bro's wet dream.