X
LLM Knowledge Bases
@karpathythreadx
TL;DR. Karpathy uses LLMs to compile a personal markdown wiki in Obsidian from raw sources, then queries and grows it agentically — shifting his token spend from code to knowledge.
Takeaways
- Pipeline: drop sources into
raw/, let an LLM "compile" them into linked .md articles with summaries, backlinks, and categories; Obsidian is the IDE, the LLM does all the writing. - At ~100 articles / 400K words, fancy RAG wasn't needed — auto-maintained index files and summaries let the agent find what it needs; outputs come back as markdown, Marp slides, or matplotlib, then get filed back into the wiki so explorations compound.
- "LLM health checks" lint the wiki for inconsistencies, missing data, and new article candidates; he sees room for a real product here, and eventually finetuning so the model "knows" the wiki in its weights.