Andrej Karpathy doesn't use LLMs the way most people do.

He doesn't ask random questions. He builds a knowledge base.

You dump raw sources into a folder. Articles, papers, GitHub repos. Your LLM compiles them into a markdown wiki with summaries, backlinks, and concept maps. All auto-generated.

Once you hit 100+ articles, you query it like a search engine. Find connections between topics. Spot inconsistencies. No fancy RAG setup needed.

The part making this compounding: every answer gets filed back into the wiki.

Most people treat AI like a search bar. The people getting real value treat it like a research partner.

Build the wiki. Let it grow.