Raw LLM Responses

​@lawschuelkeyeah so this is where smart minds can definitely disagree. I think of this from the writing/art perspective. The LSAT example is great. That's something these models can do and will definitely be able to improve on in the near future. These models can read so many LSAT questions that they have internalized all the internal patterns behind the questions. They can use these patterns to critique your questions or write new ones. It's definitely more than fancy auto complete but it's not thinking for itself either. However, with things like writing a novel or creating art, the limitations are clearer to me. It can easily write really flowing text "in the style of" Stephen King and it can even write new material that blends Stephen King and George RR Martin, but it will not be able to come up with the next big trend. Stephen King can look out at everything he's made before and think, what have I never written? What has no one ever written? And write something groundbreaking. I don't think we can LLM hard enough to get that. Right now, LLMs can write really decent code and really good paragraphs of fiction, but if there's not a human involved it falls apart really quickly when you try to scale it up. That's because as the text gets bigger the patterns get exponentially harder. I do think there are ways around this problem and we will be able to see huge advances in things like coding but I don't think it will be because the LLM got better, I think it will be because of the rules humans place around the LLM. Hopefully that all makes sense. I work in the field but I'm no PhD or anything, so keep that in mind.