Andrej Karpathy

Karpathy studied computer science and physics at the University of Toronto, then completed a master’s degree at the University of British Columbia. At Stanford, he worked with Fei-Fei Li on the relationship between images and language. Their image-description research learned correspondences between regions in a picture and words in a sentence, then used those correspondences to generate descriptions. This let a model describe parts of an image instead of only assigning a single label to the whole picture.[1][8][7]

Karpathy joined OpenAI as a founding researcher in 2015. From 2017 to 2022, he led computer vision for Tesla’s Autopilot, with a team responsible for data labeling, training neural networks, and deploying them on Tesla’s inference hardware. He returned to OpenAI in 2023 to build a team working on midtraining and synthetic data. In May 2026, he announced that he had joined Anthropic and was returning to research and development. He also said he intended to resume his education work in time.[1][9]

Teaching neural networks

Karpathy’s Neural Networks: Zero to Hero course teaches deep learning by building the machinery in code. It begins with backpropagation, the calculation that tells a network how to adjust its parameters, and progresses toward language models. The lessons connect mathematical ideas to implementations that a learner can run and modify. His earlier teaching includes Stanford’s CS231n course on neural networks for visual recognition.[2][8]

micrograd and learning by implementation

micrograd makes the mechanics of neural-network training small enough to inspect. It records arithmetic operations in a graph and works backward through that graph to calculate how each input affects the result. A small neural-network library sits on top. The repository includes a demonstration that learns to separate two classes of points, letting readers follow the same basic training process used in larger systems without starting with a large framework.[3]

From nanoGPT to nanochat

nanoGPT provides a compact implementation for training and fine-tuning GPT language models. Its model definition and training loop make the relationship between architecture, data, and optimization visible in ordinary Python code. The repository now directs users to nanochat and marks nanoGPT as deprecated. nanochat extends the work into a pipeline for training a model and talking to it through a chat interface. It also maintains experiments that measure how quickly a model can reach a specified GPT-2 benchmark score. Those results describe a particular evaluation, not equivalence to every capability of a commercial assistant.[4][5]

Tools for following research

Karpathy also built arxiv-sanity-lite to help readers work through research papers. Users tag papers they find interesting, and the application recommends similar papers using the text of their abstracts. The project addresses a different part of research work from model training: deciding what to read in a growing literature. Its implementation and setup instructions are public.[6]