Karpathy ai Personal Site

# Karpathy — karpathy.ai personal site (extracted 2026-05-27)

Source: https://karpathy.ai/

## Self-description

AI researcher and educator. Founding member of OpenAI. Former Director of AI at Tesla, leading the computer vision team for Autopilot. During his Stanford PhD, he designed and instructed the first deep learning course there (CS231n), which became one of the university's most popular classes.

## Past affiliations (full)

| Years | Role | Organization |
|---|---|---|
| 2024–present | Educational content creator | YouTube / Eureka Labs |
| 2023–2024 | Midtraining + synthetic data generation | OpenAI |
| 2017–2022 | Director of AI, Autopilot computer vision | Tesla |
| 2015–2017 | Research scientist, founding member | OpenAI |
| 2011–2015 | PhD, neural networks and computer vision | Stanford |
| 2009–2011 | MSc | University of British Columbia |
| 2005–2009 | BSc, double major in CS and physics | University of Toronto |

(Note: Wikipedia adds Anthropic, May 2026, leading pretraining team — newer than his personal site update.)

## Selected blog posts

- "A Recipe for Training Neural Networks" (2019) — canonical debugging-first checklist.
- "Software 2.0" (2017) — defining essay.
- "The Unreasonable Effectiveness of Recurrent Neural Networks" (2015) — pre-LLM RNN evangelism.
- "A Survival Guide to a PhD" (2016) — career advice.

## Major repositories

- **micrograd** — autograd engine in 200 lines + neural net library.
- **char-rnn** — character-level LSTM language model.
- **arxiv-sanity** — research paper discovery + recommendation.
- **neuraltalk2** — early image captioning system.
- **ConvNetJS** — browser-based deep learning library.
- **nanoGPT** — minimal GPT pretraining (~300 lines).
- **nanochat** (Oct 2025) — full ChatGPT-style pipeline in ~8,000 lines, capstone for LLM101n.

## Signature pattern

Every repo above proves a point by being short, readable, and self-contained. The pedagogical contract: "if you can't fit it in 200–8,000 lines, you don't understand it yet."