LLM Knowledge Bases post by Andrej Karpathy
Andrej Karpathy recently published this post on X about building and managing personal knowledge bases using LLMs, and there's some interesting information in the post that I am going to collect below.
- Lately, Andrej is spending more tokens manipulating knowledge than manipulating code.
- He uses Obsidian for storing all the knowledge (articles, papers, repos, datasets, etc.) in Markdown and image formats. And uses Obsidian's Web Clipper to grab info from web articles.
- He puts everything in a
/rawfolder and then uses an LLM to incrementally "compile" a wiki with visualizations. He specifically mentions that the LLM writes and maintains all of the data of the wiki and he doesn't manually edit/add anything. - He has ~100 articles on several topics that he can ask LLMs to fetch answers to complex questions against the wiki.
- When chatting with the LLM against the wiki, he prefers generating new Markdown files, slideshows (Marp format), or matplotlib images. And then even files the outputs back into the wiki to further enhance it.
There's much more info in the post that you can read.
I love this approach and I guess I will also be using something similar very soon, as managing lots of stuff is becoming difficult over time. And if I do, I will write about it.
Webmentions