Andrej Karpathy

The simplest, fastest repository for training/finetuning medium-sized GPTs. Arguably the most influential educational ML codebase, with 55k+ GitHub stars. Made transformer training accessible to individual developers.

Full-stack ChatGPT clone: pretraining, SFT, RLHF, and inference in ~8,000 lines of PyTorch. Trainable end-to-end on a single GPU for under $100. 49k+ stars.

AI agents that autonomously run ML experiments on a single GPU overnight. 630 lines of Python, built on the nanochat training core. 43k+ stars.

LLM training in simple, raw C/CUDA with no PyTorch or Python dependency. Achieved multi-GPU bfloat16 training with flash attention 7% faster than PyTorch nightly. 29k+ stars.

Eureka Labs' first course: 'Let's Build a Storyteller.' An undergraduate-level guide to training your own AI from scratch, designed to be guided by an AI Teaching Assistant. 36k+ stars for the course repo alone.

A tiny scalar-valued autograd engine and neural net library with PyTorch-like API. The canonical educational resource for understanding backpropagation. 15k+ stars.

YouTube lecture series covering backpropagation, language modeling, tokenization, attention, and GPT training from scratch. Over 1 million subscribers. Required viewing for aspiring ML engineers.

Co-designed and was primary instructor for Stanford's first deep learning course, Convolutional Neural Networks for Visual Recognition. Grew from 150 to 750 students and became one of the most popular CS courses at Stanford.

As Director of AI at Tesla (2017-2022), led the computer vision team responsible for Autopilot's vision-based self-driving pipeline, moving Tesla from radar+camera to a vision-only approach.

Coined 'Software 2.0' (neural networks as code written by optimization) and later 'Software 3.0' (natural language as the programming interface for LLMs), reshaping how the industry thinks about software development paradigms.