Karpathy: LLM-compiled personal wikis beat RAG at small scale

Andrej Karpathy describes a workflow where LLMs compile source documents into a personal wiki of markdown files, then answer complex queries against it — all without manual editing. He finds that at ~100 articles/400K words, a simple LLM agent with auto-maintained indexes outperforms fancy RAG. The setup uses Obsidian as frontend, with LLM handling ingest, Q&A, linting, and even slide generation. Karpathy hints this hacky script collection could become a new product.

• #karpathy
• #llm
• #knowledge-base
• #obsidian
• #rag

Andrej Karpathy@karpathy