We absolutely need public sources of truth at the very least until we can build systems that actually reason based on a combination of first principles and experience, and even then we need sources of truth for experience.
You simply cannot create solutions to new problems if your data gets too old to encompass the new subject matter. We have so systems which can adequately determine fact from fiction, and new human experiences will always need to be documented for machines to understand them.
and to be clear - as a layman, (in almost every field) I've recognized that llm's weren't up to the challenge of disavowing that notion from my phd friends up until o1 and never even tried, even though I've 'believed' since gpt 2
Any time I tried some of the more complex prompts I was working through something with Sonnet or 4o, o1 would totally miss important points and ignore a lot of the instructions while going really deep trying to figure out some relatively basic portions of the prompt.
Seems fine for reasonably simple prompts, but gets caught up when things get deeper.
But it takes so long and so much prompting skill to get to that point that the use cases seem limited.
The way I am connecting external content to the epub is through an anchoring system -- sequences of words can be hashed to form unique ids that are referenced by the enhancement. Doing this lets me index the enhancement content in such a way that is format-independent and doesn't require modifying the underlying epub.
The specific task o1 helped me with was determining what the text is visible at any given point in time. This text is then turned into hashes to pull the relevant enhancement content.
Getting the current words on the page doesn't seem all that complex, but the epub.js API is pretty confusing. There are abstractions like "Rendition", "Location", "Contents", "Range", and it's not always intuitive which of these will provide the appropriate methods. I'm sure I would have figured it out eventually by looking at the API docs, but GPT probably saved me and hour or two of trial-and-error.
My mom and grandma also read books with a physical notebook handy, and seems like modern technology should make that unnecessary.
I was laid off as part of the massive layoffs in 2022 and made a first pass at the project. I didn't use epubjs and instead wasted a lot of time making a custom rendering engine -- this was dumb. Eventually I got another job and paused the project. I started again in earnest a month or so ago using epubjs as the base.
hmm having a sort of mini-forum-like experience tied to particular pages in a book seems like a fascinating idea! being able to discuss plot twists and such only once you've already gotten to that point?
wow this seems like an amazing idea actually! any names yet? I'd love to check it out once it's done!
It makes me wonder if the AI successfully found and digested obscure sources on the internet or was just better at making sense of the esoteric documentation than me. If the latter, perhaps the need for public samples will diminish.