Skip to content

Andrej Karpathy

Category: entity Last updated: 2026-04-03 Status: draft

Summary

Andrej Karpathy is a Slovak-Canadian AI researcher and engineer, formerly Director of AI at Tesla and a founding member of OpenAI. He coined the term "vibe coding" (building software by describing intent to an AI without reading the resulting code), which became a foundational concept in the discourse around AI-assisted development. As of 2026, he is exploring LLM knowledge bases — using AI to incrementally compile personal research wikis from raw source documents — and sharing learnings publicly on X.

Details

Background

  • Formerly Director of AI at Tesla (Autopilot)
  • Founding member of OpenAI
  • Known for educational content on deep learning (YouTube, X)
  • Based in the San Francisco Bay Area

Key Contributions to AI Discourse

Vibe Coding — coined the term for building software by describing desired behavior to an AI without reading or understanding the generated code. Originally framed as appropriate for personal projects and prototyping where bugs only affect yourself; the term was later expanded colloquially to cover all AI-assisted coding, which entities/simon-willison argues dilutes its usefulness.

LLM Knowledge Bases — described his workflow for using LLMs to compile personal research wikis: source documents in raw/ → LLM incrementally compiles a wiki of .md files with summaries, backlinks, and concept articles → Q&A via index-based navigation → outputs as Markdown/slides/visualizations filed back into the wiki. The canonical description is his GitHub gist "LLM Wiki" (Apr 4, 2026). He cites Vannevar Bush's Memex (1945) as the historical antecedent. See concepts/llm-knowledge-base and entities/vannevar-bush.

Ephemeral Wiki Vision — proposed that every query to a frontier LLM could spawn a team of LLMs to construct an ephemeral wiki, lint it in a loop, and write a full report — going far beyond a single text response.

Tools & Stack (as of early 2026)

  • Obsidian as the IDE/frontend for viewing raw data, wiki, and visualizations
  • Obsidian Web Clipper for converting web articles to Markdown
  • Marp (Obsidian plugin) for slide output
  • Custom vibe-coded search engine over the wiki (used directly and via CLI as an LLM tool)
  • Matplotlib for chart outputs

Key Claims & Data Points

  • Coined "vibe coding" — [source: wc8FBhQtdsA] (Simon Willison reference)
  • At ~100 articles / ~400K words, no RAG needed; index-based LLM navigation suffices — [source: karpathy_thread]
  • "I think there is room here for an incredible new product instead of a hacky collection of scripts" — [source: karpathy_thread]

Open Questions

  • What specific topic domains has Karpathy built LLM knowledge bases for? (raised by: entities/andrej-karpathy, 2026-04-03)

Sources