Andrej Karpathy is a Slovak-Canadian computer scientist known for his research contributions, his leadership of Tesla's Autopilot program, and his influential public AI education work. He completed his PhD at Stanford under Fei-Fei Li, focusing on connections between Computer vision and Natural language processing, and taught Stanford's CS231n convolutional neural networks course, whose lecture notes and materials became widely used self-study resources for a generation of Deep learning practitioners.
OpenAI and Tesla
Karpathy was a founding member of OpenAI in 2015, working on early Reinforcement learning and Deep learning research. In 2017 he left to join Tesla as senior director of AI, where he led the Autopilot computer vision and neural network team, restructuring Tesla's self-driving software around large neural networks trained on fleet-collected data rather than heavily hand-engineered rules, part of the broader industry arc around self-driving cars. He left Tesla in 2022 and briefly returned to OpenAI in 2023 before departing again in 2024 to focus on independent education projects.
Education and public writing
Throughout his career, Karpathy has maintained an unusually high public profile as an educator, publishing the widely read 2015 blog post "The Unreasonable Effectiveness of Recurrent Neural Networks," building the "Neural Networks: Zero to Hero" video course series, and creating minimal, readable reference implementations of core Deep learning and language model training code intended to demystify how systems like GPT-2 work internally. In 2024, he founded Eureka Labs, an AI-native education startup aiming to combine expert-created course content with AI tutoring.
Coining "vibe coding"
In February 2025, Karpathy popularized the term "vibe coding" in a widely shared social media post, describing a style of software development in which a programmer describes what they want in natural language, accepts and iterates on code generated by an AI coding assistant or agent, and does not closely read or fully understand the generated code, "forgetting that the code even exists." The term was quickly adopted across the industry to describe a broader shift in how developers used tools such as Claude Code, Cursor, and GitHub Copilot, while also drawing criticism from some engineers who argued it normalized poor code review and technical debt; the debate over its costs and benefits has continued as agentic coding tools became more capable through 2025 and 2026.