Andrej Karpathy

AI researcher and educator: founding member of OpenAI, Director of AI at Tesla, founder of Eureka Labs, author of the LLM wiki pattern.

Website ↗X ↗GitHub ↗YouTube ↗karpathy.bearblog.dev ↗Medium ↗karpathy.github.io ↗

Work

Andrej Karpathy is an AI researcher and educator.1 He is Slovak-Canadian and was born in 1986 in Bratislava. In 2026 he joined the pretraining team at Anthropic.2

Career

He studied at the University of Toronto (BSc) and the University of British Columbia (MSc). From 2011 to 2015 he did his PhD at Stanford with Fei-Fei Li. There he designed and taught CS 231n, which he calls the first deep learning course at Stanford. From 2015 to 2017 he was a research scientist and founding member of OpenAI. From 2017 to 2022 he was Director of AI at Tesla and led the computer vision team for Autopilot. From 2023 to 2024 he was back at OpenAI, with a new team for midtraining and synthetic data.1

Teaching

In 2024 he founded Eureka Labs, an AI education platform, and in February 2025 he coined the term vibe coding.2 Since 2024 he has published videos on AI on YouTube, among them the series Zero to Hero.1

LLM wiki

On 4 April 2026 he published LLM Wiki Idea File, a text for an LLM agent that describes a knowledge base the agent builds and maintains.3 This knowledge base follows that pattern.

More on this

Made
  • LLM Wiki Idea File Andrej Karpathy's idea file from April 2026: an LLM agent builds and maintains a wiki of linked Markdown files from raw sources.

    A GitHub gist by Andrej Karpathy, created on 4 April 2026, with the single file llm-wiki.md. It calls itself an idea file: a text to paste into an LLM agent, which then works out the details with its user.

    Content

    • The idea: instead of searching raw documents again for every question, an LLM builds and keeps a wiki of linked Markdown files. The gist calls it "a persistent, compounding artifact".
    • Three layers: raw sources that the user curates and nobody changes, the wiki that the LLM owns, and a schema file such as CLAUDE.md or AGENTS.md with the conventions.
    • Operations: ingest reads a new source, updates the pages it touches and logs the change. Query answers from the wiki and can file good answers back as pages. Lint looks for contradictions, stale claims, orphan pages and missing links.
    • Navigation: index.md lists the content, and log.md records every change in order.
    • Background: it ties the idea to Vannevar Bush's Memex.

    This knowledge base follows the pattern; How an LLM Wiki Works shows how.

See also
  • Anthropic American AI company, a public benefit corporation in San Francisco, that develops the Claude language models.
  • OpenAI American AI company in San Francisco, a public benefit corporation, that develops ChatGPT, Codex and the GPT models.
Related to their work

Show in the graph view →