Thank you, Gwern. The post from today that you link is fascinating. It makes it obvious that the design I’m toying with is not radical enough. Fascinating to see how the web browsing model uses the internet as a sort of knowledge graph. I would have thought it too messy.
I'm pretty sure you could do repeated prompting + RL finetuning to expand outlines into books, the inverse of https://openai.com/blog/summarizing-books/ . Or consider today's post, finding a lot of snippets on the Internet and collating them and then writing a larger summary/argument based on them: https://openai.com/blog/improving-factual-accuracy/
Thank you, Gwern. The post from today that you link is fascinating. It makes it obvious that the design I’m toying with is not radical enough. Fascinating to see how the web browsing model uses the internet as a sort of knowledge graph. I would have thought it too messy.