Andrej Karpathy

About this vault

The complete Karpathy knowledge base and personal AI second brain: vibe coding, autoresearch, agentic engineering, LLM wiki, and LLM deep dives — organized by topic, not just timeline. Covers his April 2026 LLM knowledge base + LLM wiki methodology — the 'compiler analogy' approach that bypasses RAG (16M+ tweet views), his three-folder system (raw → wiki → outputs) for building a self-maintaining personal AI knowledge base, the vibe coding manifesto he coined (now Collins Dictionary's word of the year), his autoresearch framework (21K+ GitHub stars), his 3.5-hour LLM deep dive lectures, and the post-code AI workflow shift from writing code to orchestrating knowledge. Each piece has an AI summary so you can decide what to read in seconds. Auto-synced — new Karpathy content appears here automatically.

Karpathy's Three-Folder LLM Wiki System

Andrej Karpathy's personal AI knowledge base uses three folders that process information in a pipeline. The system bypasses traditional RAG and works with any capable LLM.

1. 1. Raw — Unfiltered inputs. Articles, transcripts, X threads, research papers, podcast notes. Everything goes in as-is, timestamped. This is the append-only firehose. No organization, no tags, no pre-interpretation.
2. 2. Wiki — Synthesized topic pages. Concepts get their own living file, continuously edited as new raw material lands. The wiki is the map, not the territory — a self-updating mental model per topic, written for future-you.
3. 3. Outputs — Finished artifacts published downstream. Essays, tweets, lectures, code, talks. Outputs are what the wiki eventually becomes when an external deadline demands it — the wiki compresses into a public product.

Karpathy describes this as the “post-code AI workflow” — the shift from writing code as the primary creative act to orchestrating knowledge. The canonical pieces on the pattern are in the reading list below.