Andrej Karpathy

Career

Karpathy's PhD at the Stanford Vision Lab, advised by Fei-Fei Li, focused on convolutional and recurrent neural networks for computer vision, natural language processing, and their intersection. From 2015 to 2017 Karpathy was a research scientist and founding member at OpenAI, and from 2017 to 2022 was Director of AI at Tesla, leading the computer vision team of Tesla Autopilot and a team that handled data labeling, neural network training, and deployment on Tesla's custom inference chip.¹

Karpathy returned to OpenAI from 2023 to 2024 to build a new team working on midtraining and synthetic data generation. Since 2024 Karpathy's site describes work making educational videos on AI.¹

Teaching and writing

During the PhD, Karpathy designed and was the primary instructor for Stanford's CS 231n, Convolutional Neural Networks for Visual Recognition, first taught in 2015. Karpathy's current YouTube lectures run in two tracks: the technical Zero to Hero playlist and general-audience videos such as Deep Dive into LLMs like ChatGPT and How I use LLMs.¹

Essays listed on Karpathy's site include "The Unreasonable Effectiveness of Recurrent Neural Networks" (2015), "Software 2.0" (2017), and "A Recipe for Training Neural Networks" (2019).¹

Open-source work

Karpathy's repositories favor small, readable code for training language models. nanoGPT is a compact repository for training and fine-tuning medium-sized GPTs that reproduces GPT-2 (124M) on OpenWebText; in November 2025 its README pointed readers to nanochat, a minimal harness that covers tokenization, pretraining, finetuning, evaluation, and inference on a single GPU node.³⁴⁵

llm.c implements language model training in plain C and CUDA, focused on reproducing the GPT-2 and GPT-3 miniseries alongside a PyTorch reference, while llama2.c trains Llama 2 architecture models in PyTorch and runs them from a single C file. Earlier projects listed on Karpathy's site include the micrograd autograd engine, char-rnn, and arxiv-sanity.⁶²⁷¹

Work and contributions

• llama2.c

  Wrote the C inference engine, run.c, as a single file with no dependencies, alongside PyTorch training code adapted from Karpathy's earlier nanoGPT. The README walks through running small TinyStories models on an M1 MacBook Air and documents runq.c, which runs int8-quantized checkpoints. The project is MIT-licensed.²⁷⁵

• Meta's Llama 2 models in llama2.c

  llama2.c's export script converts Meta's Llama 2 checkpoints, including the chat models, and Hugging Face models with the Llama 2 architecture into its own format for the C program. The README notes that float32 inference is practical only up to the 7B model and that the int8 export produces a much smaller checkpoint.²

• AI education videos on YouTube Show, channel or publication

  Karpathy's site says Karpathy makes educational videos on AI on a YouTube channel in two tracks: a technical track, the Zero to Hero playlist, and a general-audience track that includes Deep Dive into LLMs like ChatGPT and How I use LLMs.¹

• nanochat Open-source code

  A minimal, hackable harness for training language models on a single GPU node that covers tokenization, pretraining, finetuning, evaluation, and inference, and succeeds Karpathy's nanoGPT.⁴³

Timeline

1. 2011Began a PhD at Stanford on neural networks for vision and language, advised by Fei-Fei Li.¹
2. 2015Taught the first offering of Stanford's CS 231n deep learning course as its primary instructor.¹
3. 2015Joined OpenAI as a research scientist and founding member.¹
4. 2017Became Director of AI at Tesla, leading the Autopilot computer vision team.¹
5. 2023Returned to OpenAI to build a team working on midtraining and synthetic data generation.¹
6. 2024Began making educational videos on AI in technical and general-audience tracks.¹
7. Nov 2025Marked nanoGPT as deprecated in favor of nanochat, a minimal harness for training LLMs end to end.³⁴

Documented roles

The sources do not confirm a current role; these are the roles they document.

• Research scientist and founding member (2015–2017); later built a team working on midtraining and synthetic data generation (2023–2024), OpenAI¹
• Director of AI (2017–2022), Tesla¹

Companies in their history

• OpenAI 63 people in the wiki
• Tesla 4 people in the wiki

Related people

• Jakub PachockiConnected to OpenAI
• Sam McCandlishConnected to OpenAI
• Aravind SrinivasConnected to OpenAI
• Barret ZophConnected to OpenAI
• Eric MitchellConnected to OpenAI
• Dwarkesh PatelTeaching and writing · Language models
• Zvi MowshowitzTeaching and writing · Language models
• Matthew BermanTeaching and writing · Language models
• Tatsunori HashimotoLanguage models · Teaching and writing
• Yannic KilcherTeaching and writing · Language models