Research at scale
From ImageNet and video understanding at Google and DeepMind to Autopilot vision at Tesla — training and deploying neural networks on real-world data at production scale.
I like to train deep neural nets on large datasets. This page collects my career timeline, teaching, writing, open-source projects, talks, and publications — from Stanford CS231n and Tesla Autopilot to LLM education on YouTube.
Andrej Karpathy is an AI researcher and educator known for bridging research, production systems, and clear explanations of how modern neural networks work.
From ImageNet and video understanding at Google and DeepMind to Autopilot vision at Tesla — training and deploying neural networks on real-world data at production scale.
Architect and lead instructor of Stanford CS231n. Today, educational videos on LLMs and neural networks reach learners through technical and general-audience tracks.
Open-source tools like micrograd and char-rnn, plus essays such as Software 2.0 and A Recipe for Training Neural Networks, shaped how many developers learn deep learning.
A chronological view of roles, research, and milestones — aligned with the biography on karpathy.ai.
I create educational videos on AI on my YouTube channel. The videos come in two parallel tracks: a technical track and a general audience track.
For all the latest, I spend most of my time on X/Twitter or GitHub.
I came back to OpenAI where I built a new team working on midtraining and synthetic data generation.
I was the Director of AI at Tesla, where I led the computer vision team of Tesla Autopilot and (very briefly) Tesla Optimus. My team handled in-house data labeling, neural network training, and deployment on Tesla's custom inference chip. See Tesla AI Day 2021 for more.
I was a research scientist and a founding member at OpenAI.
My PhD focused on convolutional/recurrent neural networks in computer vision, NLP, and their intersection. Adviser: Fei-Fei Li. I designed and was the primary instructor for CS 231n: Convolutional Neural Networks for Visual Recognition — growing from 150 students in 2015 to 750 in 2017. Internships at Google Brain (2011), Google Research (2013), and DeepMind (2015).
Worked with Michiel van de Panne on learning controllers for physically-simulated figures — machine learning for agile robotics in simulation.
Double major in computer science and physics with a minor in math. This is where I first got into deep learning, attending Geoff Hinton's class and reading groups.
Selected appearances on LLMs, autonomous driving, PyTorch at Tesla, and the Software 2.0 stack.
From the first Stanford deep learning class to today's YouTube lectures on LLMs and neural networks from scratch.
Technical YouTube playlist building micrograd, makemore, and GPT-style models from scratch. Full series page.
Convolutional Neural Networks for Visual Recognition — course notes, syllabus, and 2016 lecture videos.
Deep dives into how ChatGPT-style systems work and practical guides for using LLMs in everyday workflows.
Popular posts on training neural networks, Software 2.0, RNNs, PhD advice, and more — across GitHub, Medium, and Bear blogs.
Educational and research code that helped many developers understand autograd, RNNs, arXiv discovery, and in-browser deep learning.
A tiny scalar-valued autograd engine implementing reverse-mode autodiff over a dynamically built DAG, with a small neural net library and PyTorch-like API.
Torch character-level language model built from LSTMs/GRUs/RNNs — paired with the famous blog post on RNN effectiveness.
Discover relevant arXiv papers, search by similarity, and get recommendations — a from-scratch rewrite of the original arxiv-sanity project.
Early image captioning in (lua)Torch, extended later with Justin Johnson to dense captioning.
Deep learning in JavaScript — train CNNs entirely in the browser with many interactive demos.
Privacy-first productivity tracker — an alternative to cloud-based time tracking for developers.
Sometimes jokingly called the reference human for ImageNet after competing against an early ConvNet on 1,000-way classification. See the blog post and Wired article.
Vision-language models, video classification, PixelCNN++, and web agents — full list on Google Scholar.
Quick answers on background, teaching, and where to follow new work.
Andrej Karpathy is an AI researcher and educator. He was a founding member of OpenAI and later Director of AI at Tesla, where he led the computer vision team for Autopilot. At Stanford he architected CS231n, one of the university's most popular deep learning courses.
He creates educational videos on AI via YouTube — including the Zero to Hero technical series and general-audience explainers on large language models. Updates are most active on X/Twitter and GitHub.
Beginners interested in how LLMs work can start with intro and deep-dive videos on YouTube. Developers building intuition from scratch should follow the Zero to Hero playlist and micrograd repository.
Software 2.0 is Karpathy's framing where neural networks become the primary program — optimization finds weights instead of humans hand-writing all logic. The essay remains influential in ML engineering culture.
As Director of AI, he led Autopilot vision: in-house labeling, training, and deployment on Tesla's custom inference hardware, with the long-term goal of scalable full self-driving.
This site is a lightweight static tribute page. The official home page remains karpathy.ai — built with pure HTML and CSS, no heavy frameworks.