Haydn Belfield compared the pipeline's architecture to DeepMind's Co-Scientist.
Positive users like AI agent systems that generate mini-books on historical topics for creating tooltip-based explainers in technical domains, while negative users dismiss them as a debugging nightmare due to unchecked hallucinations.
Of course they are not books in the sense of carrying out a sustained argument, or building structural metaphors/analogies, over hundreds of pages. Models cannot really do that still ime, which is ironic given how much their analysis of writing is obsessed with structure.
Related use case I like: when I wanna learn about a technical domain have an LLM make me an explainer doc where every jargon term has a tooltip aimed at explaining it like-I'm-12, but without dumbing it down. Give me a concrete example of what an instance of The Jargon would look like:
@rustbeltjacobin I would but these things all exist within one app, and some of my research interests are things I keep to myself because I don’t like to publicly share artifacts of half-baked ideation
@deanwball Post the links to the books
@HaydnBelfield @rustbeltjacobin I want to try this with the new approach I'm using which is maybe 50x more token intensive
@deanwball @rustbeltjacobin Couldn't an agent pull out the text to a separate website pretty easily?
Loads of folks would be interested to have a gander at one of these minibooks
@deanwball Curious what type of website it resembles? Something like this https://notes.eecs245.org/linear-independence/linear-independence/?
@deanwball I'm doing something my similar to study languages. Here's the content part
"I'm a widely read PhD, know all type. Pls suggest 10 pieces on subject X that I'm likely to not know"
Then I select.
@deanwball Booo. Bean W Dall.
@deanwball Or you could read a short book about the place
@deanwball @rustbeltjacobin Cock tease
@deanwball debugging this must be a nightmare. what happens when research hallucinates and the fact-checker just agrees