On thing I think I’m starting to learn about LLM’s is like: “Assuming it has been trained in a lot of code, it should be one of the best judges of what is good

On thing I think I’m starting to learn about LLM’s is like: “Assuming it has been trained in a lot of code, it should be one of the best judges of what is good/rare/valuable or not.” It feels like the search space for a lot of LLM’s is relatively narrow. Maybe that’s what deep research is for? But like, I was talking to an LLM about nostalgia earlier today, and I noticed rather than trying to combine sort of a number of different references to come up with an answer, it was really limited to one source I think that I gave it, and just gave bad answers that I had to correct and steer. (This was Claude). So I think to really combine info in a sophisticated way, would be a relatively difficult task for an LLM, at least, in normal chat. So like, generating code is a fairly simple task because I think the weights it’s writing down are organized naturally in this way from training on lots of code, but then combining that with sort of weights on the books and blog posts that talk about how to write code? I don’t think it actually does that.