Andrej Karpathy
Former OpenAI and Tesla AI researcher whose viral April 2026 X post on building personal LLM knowledge bases with flat markdown files and Claude Code spawned a new practice pattern. His LLM Wiki approach — no vector DB, no embeddings, just well-indexed markdown that the LLM reads directly — became a dominant alternative to RAG for sub-million-document knowledge bases.
In a Sequoia Capital interview, Karpathy expands that worldview into Software 3.0, Agentic Engineering, Verifiability Frontier, and the idea that human understanding remains the bottleneck even when agents can outsource large amounts of thinking and implementation [src-055].
Key facts
- Published viral April 2026 X post on building LLM knowledge bases
- Runs ~100 articles / ~500K words through his own wiki
- Advocates flat folder structures over heavy subfolder organisation
- Recommends Obsidian as the IDE for browsing markdown wikis
- Runs LLM health-check lints to find inconsistencies and gap-fill articles
- Explicitly left his pattern vague so others could customise it
- Frames Software 3.0 as prompting/context over an LLM interpreter, extending his earlier Software 1.0 / Software 2.0 distinction [src-055]
- Distinguishes vibe coding from Agentic Engineering: vibe coding raises the floor, while agentic engineering preserves professional quality while coordinating agents [src-055]
- Argues current LLMs automate what can be verified, producing jagged peaks in code/math and weaker performance outside trained/verifiable circuits [src-055]
- Says LLM knowledge bases are useful because they re-project information into structures that improve human understanding [src-055]
- Howell cites Karpathy’s introductory LLM material and Neural Networks: Zero to Hero as practical deep-learning/LLM learning resources, and closes with Karpathy’s advice to learn through concrete projects and self-explanation [src-075]
Related
- See also: Sequoia Capital, Claude Code, OpenClaw, Gemini, Imagen 3 (Nano Banana 2)
- Concepts: Software 3.0, Agentic Engineering, Verifiability Frontier, Jagged Intelligence, Agent-Native Infrastructure, Understanding Bottleneck, Project-Based AI Learning, AI Learning Roadmap
Source references
- [src-004] Nate Herk cluster — Nate Herk — Claude Code cluster (21 videos)
– Videos referenced: sboNwYmH3AY
- [src-013] Nate Herk — “Build & Sell Claude Code Operating Systems (2+ Hour Course)” (2026-05-01)
– Nate reads Karpathy’s original tweet and gist verbatim, builds the raw/ → wiki/ architecture from the prompt Karpathy published, and demos it live with the AI2027 article as the first source ingested.