Andrej Karpathy
Slovak-Canadian AI researcher, born 1986. Co-founder of OpenAI (2015). Former Director of AI at Tesla (2017-2022). Stanford PhD under Fei-Fei Li. One of the small number of practitioners who consistently teaches the field by building from scratch and narrating the process.
Currently runs Eureka Labs (founded 2024), focused on AI-native education.
Why He Matters Here
Beyond his technical contributions, Karpathy has been unusually influential on how individuals + small teams should adopt LLM tooling. Three of his patterns are load-bearing in this very wiki:
- The LLM wiki pattern — this vault you’re reading. Markdown notes as an LLM-readable knowledge base. Query via
cd ~/wiki && claude. See concept-rag for how this relates to broader retrieval patterns. - The “vibe coding” framing — writing code by describing intent to an LLM and iterating; Karpathy named it in early 2025.
- The neural-net-from-scratch curriculum — his “Zero to Hero” YouTube series (micrograd, makemore, GPT, etc.) is the most-cited single ML-self-teaching path.
Public Threads Worth Tracking
- Software 2.0 (2017 essay) — frame for treating ML models as compiled artifacts, the next layer after Software 1.0 (handwritten code).
- GPT explained lectures — definitive intro to transformer-era LLMs at intuition level.
- NanoGPT (GitHub) — ~300-line GPT implementation that became the standard “minimal working GPT” reference.
- llm.c — C-only GPT-2 training reimplementation; reduced education-side complexity dramatically.
- Eureka Labs (LLM101n) — course in development as of 2024-2026, building toward AI-native pedagogy.
His Pattern: Build to Understand
The unifying methodology: he doesn’t explain by abstraction, he explains by reproducing the artifact end-to-end at minimum complexity. nanoGPT is GPT in 300 lines so you can read every line. The LLM wiki pattern is “your notes plus Claude Code” because that’s the minimum useful unit; vector DBs come later if you need them.
This is unusual in AI research, where most communication abstracts up. Karpathy’s bet: the foundations are simple, the moats are confusion. Removing confusion is the contribution.
What He’s Been Saying Recently (2025-2026)
- Wikipedia-but-for-everything framing — LLMs let small operators maintain personal canonical references in domains they care about, which the Karpathy LLM wiki pattern operationalizes.
- “Cursor + Claude Code era” — the IDE has split into IDE-as-text-editor and IDE-as-AI-collaborator; both will coexist for years.
- Skepticism on autonomous agents for production. He repeatedly argues humans in the loop is the right setting for current model capability.
Cross-Realm
- philosophy: His “build to understand” approach is implicit pragmatism — anti-abstraction, anti-jargon. There’s a connection to overview-vedanta’s emphasis on anubhava (direct experience over scholastic study).
- history: The figures he resembles in pedagogy are Feynman (physics) and Knuth (CS) — practitioners who taught by demonstrating the construction.
- computing / ai-computing: Operating system for this realm.