LLMs are becoming maintainers of structured knowledge bases

Andrej Karpathy describes a workflow where raw sources are collected into folders, then an LLM compiles them into a Markdown wiki with summaries, backlinks, indexes, outputs, and maintenance passes. Why it matters: This reframes LLM knowledge work as a compounding system instead of disposable chat. The model becomes a maintainer of an inspectable corpus: ingest raw sources, compile structured pages, answer questions from the wiki, generate artifacts, and lint the system over time.

Try this: For one research topic, create raw, wiki, outputs, and indexes folders. Add 5-10 sources, ask an LLM to compile concept pages and indexes, then require each complex answer to be saved back as a Markdown artifact.