I said that there twitter.com/arjie/status/16356939270713… some hour and a half after his post which is about when I saw it. I didn't realize others also did. They felt isomorphic to me. If the AI generated content is regurgitation of previous content, novel human-generated content will stand out and human-generated content will become valuable. i.e. the balancing mechanism is that novelty will beat dumb replication in the same way that an active manager who has true alpha can beat a passive fund. Now, personally, I think that multimodal LLM-groups plus feedback plus many more sensors is not far from where we are, so I mentioned that if AIs do become creative it's not a problem - they will be us or better! I, personally, believe that's where we'll go. Anyway, I'm not too inclined to discuss in this forum (though we can chat if you're in SF) since it's easy to misunderstand and jump into argument spirals, but I just wanted to make it clear that it's not a "gotcha" or a dunk or anything like that. Nothing is being disparaged. It's just a conversation like a normal conversation where you bring up what you think and mention ideas you have that others might find interesting. If many others have brought it up, about the best I can say is that in this respect it appears I am not particularly inventive - perhaps the best evidence that LLMs are already human-like ;)

Generative AI is just the new "bottom" in terms of quality. All that you have to do to compete against it is to be a little better than it. The question to me really is whether the quality of this new "bottom" is adequate. Sometimes it is for some people and for some applications, and sometimes it is not. I do not use it myself because I am a researcher and I often ask questions that don't have a lot of "training data" yet. And even if an area is well covered in terms of "training data", often there is a lot of "know how" that really isn't written down in an easily digestible form. It is passed verbally or through examples in person. So the idea that the "training data" is complete is also not true in general. Many other people in this thread have already covered that books are much more structured and organized than any answer generative AI gives you. Let me discuss another reason why books still matter. Books can give you a wider view than the "consensus" that something like ChatGPT gives you. I know a lot of books in my field that derive results in different ways, and I often find value in these different approaches. Moreover, suppose that only one book answers the question that you want answered but others gloss over that subject. Generative AI likely will not know precisely what one random book said on the subject, but if you were searching through multiple books on the subject yourself, you likely would pick up on this difference. Relevant Paul Graham quote [1]: > We can't all use AI. Someone has to generate the training data. [1] x.com/paulg/status/1635672262903750662