Andrej Karpathy is a Slovak-Canadian AI researcher who has had an unusual double career: he has held some of the most consequential engineering jobs in the field — founding member of OpenAI, director of AI for Tesla Autopilot — while becoming, in parallel, its most-watched teacher. Both halves run on the same conviction, stated in his 2016 essay "Yes you should understand backprop": the important systems are leaky abstractions, and the only durable advantage is understanding what is underneath.

## Formation

Born October 23, 1986 in Bratislava, then Czechoslovakia, he moved to Toronto at fifteen. A double bachelor's in computer science and physics at the University of Toronto (2009) was followed by a master's at UBC (2011) on physically simulated figures under Michiel van de Panne, then a Stanford PhD (2016) advised by Fei-Fei Li. His dissertation, "Connecting Images and Natural Language," and the CVPR 2015 paper "Deep Visual-Semantic Alignments" with Li helped establish image captioning and dense captioning as working research areas. Along the way he squeezed in two Google internships on YouTube-scale feature learning and a 2015 DeepMind internship on deep reinforcement learning.

The education thread starts here too. With Li he designed CS231n — Stanford's first deep learning course — and was its primary instructor as enrollment grew from 150 in 2015 to 750 in 2017, per his Stanford bio. He also wrote the browser-based deep learning libraries (ConvNetJS, RecurrentJS, REINFORCEjs, t-sneJS) that powered the course's live demos, built arxiv-sanity, and in May 2015 published "The Unreasonable Effectiveness of Recurrent Neural Networks" — the post, with its char-rnn code release, that made character-level language models a rite of passage.

## OpenAI, Tesla, OpenAI again

OpenAI's December 11, 2015 launch post lists him among the founding research engineers and scientists. There he led the World of Bits browser-agent benchmark inside the Universe platform (December 2016). In June 2017 Tesla hired him as Director of AI and Autopilot Vision, reporting directly to Elon Musk. His own summary of the five years that followed: he led the computer vision team of Tesla Autopilot — and very briefly Tesla Optimus — running in-house data labeling, neural network training, and deployment onto Tesla's custom inference chip. The public record of that work is mostly his own stagecraft: the Autonomy Day 2019 talk on fleet learning and the "data engine," the CVPR'20 keynote on multi-task networks, and the opening deep dive at Tesla AI Day 2021, where he explained how eight camera streams get fused into a 3D "vector space."

Three months into that same tenure he had also written the essay he is most cited for: "Software 2.0" (November 11, 2017) argued that neural networks are not another ML tool but a new way to write software — the programmer supplies goals and datasets, and optimization compiles them into weights.

He left Tesla on July 13, 2022 after a sabbatical, saying only that he wanted to revisit "technical work in AI, open source and education." What followed made that concrete: the Zero to Hero video course (starting with the micrograd lecture in August 2022), nanoGPT (December 2022), then a return to OpenAI in February 2023 — where, by his own account, he built a team on midtraining and synthetic data — a "State of GPT" talk at Microsoft Build, and the one-hour "Intro to Large Language Models" video. He left OpenAI again in February 2024, insisting there was "no drama," and that July announced Eureka Labs, "a new kind of school that is AI native," built on a teacher + AI-assistant symbiosis with the undergraduate course LLM101n as its announced first product.

## The vocabulary maker

Karpathy keeps naming the thing before everyone else has a word for it. "Software 2.0" in 2017. "The hottest new programming language is English" in January 2023. "Vibe coding" on February 2, 2025 — a tossed-off description of accepting all diffs and forgetting the code exists that MIT Technology Review traced into a movement and Collins Dictionary named its 2025 Word of the Year. His June 2025 YC keynote extended the lineage: "Software 3.0," in which LLMs are a new kind of computer programmed in natural language, and the practical art is partial autonomy — fast human verification loops, GUIs, an "autonomy slider," AI kept on a leash.

The caution is as consistent as the coinage. On Dwarkesh's podcast in October 2025 he called this "the decade of agents," not the year — pushing back on timeline hype while describing models as "ghosts, not animals": jagged stochastic simulations with superhuman recall and real deficits. Five years on Autopilot supply his reference lesson: a Waymo demo impressed him around 2014, and it took a decade to become a payable product.

## The work itself

His public artifacts share a recognizable aesthetic — the smallest code that honestly does the thing: micrograd (~100-line autograd engine), minGPT (~300-line GPT model), nanoGPT (GPT-2 in a ~300-line training loop), llm.c (the same in raw C/CUDA), and nanochat (October 2025, "the best ChatGPT that $100 can buy," covering tokenization through RL and inference). Each ships with a lecture, a walkthrough, or an essay attached. TIME's 2024 TIME100 AI entry framed him as the field's educator first.

## What's unsettled

The seams are visible: the reasons for both his Tesla and second OpenAI exits were never disclosed; the announced Eureka Labs course remains "under development" with its repo archived; and on May 19, 2026 he joined Anthropic's pretraining team — saying he would get back to research and resume education work "in time" — which leaves the school's roadmap open.

*This index was compiled from public sources and does not imply the subject's endorsement. Citations live in the packet's source catalog.*