Andrej Karpathy

Karpathy was born in Slovakia and moved with his family to Toronto, Canada, at 15. In a retrospective on his PhD he wrote that he knew from a young age that he wanted a PhD, partly because he liked school and learning and partly because he wanted to be like Gordon Freeman, the protagonist of the game Half-Life, who has a PhD from MIT in theoretical physics.⁠[1][2][3][3]

In his early years Karpathy was into Rubik’s Cubes, learning to solve them in about 17 seconds; frustrated by the lack of learning resources, he made YouTubeYouTubeYouTube is Google’s video-sharing and creator platform, founded in 2005 by Steve Chen, Chad Hurley, and Jawed Karim and acquired by Google in 2006. Built around video uploading, viewing, and creator earnings, the platform has developed Shorts, YouTube Music, Premium, YouTube TV, and products for children, and is headquartered in San Bruno, California, United States.Open the full entry videos explaining speedcubing methods, which went on to become relatively popular.⁠[4]

From 2005 to 2009, Karpathy studied for a bachelor’s degree at the University of Toronto with a double major in computer science and physics and a minor in math, first getting into deep learning by attending Geoffrey Hinton’s class and reading groups. He recalled that as an undergraduate he had intended to work on quantum computing but found he was not having fun in his quantum mechanics classes and gravitated toward AI; realizing in a library that he could never read all the books himself, he thought he might instead build something that could learn everything.⁠[4][1][1][1]

From 2009 to 2011, Karpathy completed a master’s degree at the University of British Columbia, working with Michiel van de Panne on learning controllers for physically simulated figures, that is, machine learning for agile robotics in physical simulation; the related paper “Locomotion Skills for Simulated Quadrupeds” was published at SIGGRAPH 2011.⁠[4][4]

From 2011, Karpathy pursued a PhD at the Stanford Vision Lab under adviser Fei-Fei Li, also working with Daphne Koller, Andrew Ng, Sebastian Thrun, and Vladlen Koltun during his first-year rotations. His research focused on convolutional and recurrent neural networks and their applications in computer vision, natural language processing, and their intersection. During his PhD he interned at the early Google Brain in 2011 on large-scale unsupervised learning from videos, at Google Research in 2013 on large-scale supervised learning on YouTube videos, and in 2015 on DeepMindGoogle DeepMindGoogle DeepMind is Google’s artificial intelligence research organization, formed in April 2023 by merging DeepMind, which Google acquired in 2014, with the Brain team from Google Research, and led by CEO Demis Hassabis. It developed AlphaGo, AlphaFold, and the Gemini family of models, and Hassabis and John Jumper received the 2024 Nobel Prize in Chemistry for AlphaFold’s protein structure prediction work.Open the full entry’s deep reinforcement learning team with Koray Kavukcuoglu and Vlad Mnih.⁠[4][4][4][4]

His PhD thesis, “Connecting Images and Natural Language,” was completed in 2016; it studied how neural networks can identify multiple specific objects in an image and describe them in natural language, and conversely find the object in an image that matches a natural-language description such as “white tennis shoes.”⁠[4][5][5]

During his PhD, Karpathy wrote ConvNetJS, a deep learning library written from scratch in JavaScript that trains convolutional neural networks directly in the browser, and in April 2014 he gave an interview about it to Data Science Weekly. In September 2014 he competed against an early convolutional network at sorting images into ImageNet’s 1,000 classes, building custom tooling and learning a lot about dog breeds, which earned him the joking title of the “reference human” for ImageNet.⁠[4][6][4][6]

In 2015, Karpathy designed and was the primary instructor for Stanford’s first deep learning course, CS231n, “Convolutional Neural Networks for Visual Recognition.” It became one of the largest classes at Stanford, growing from 150 students in 2015 to 330 in 2016 and 750 in 2017, and his 2016 lecture videos and course notes are publicly available.⁠[4][4][4]

On May 21, 2015, Karpathy published the blog post “The Unreasonable Effectiveness of Recurrent Neural Networks,” training character-level recurrent neural network language models to write poetry, LaTeX math, and code, and released char-rnn, a Torch character-level language model built with LSTMs and similar architectures.⁠[6][4]

In December 2015, OpenAI was announced at the Neural Information Processing Systems conference (NIPS) in Montreal; Karpathy was one of its founding members and worked there as a research scientist from 2015 to 2017. He was still a Stanford PhD candidate at the time, with internships at Google and DeepMind on his résumé.⁠[7][4][7]

On September 7, 2016, after his PhD came to an end, Karpathy published “A Survival Guide to a PhD,” summarizing what he had learned.⁠[3][3]

On June 20, 2017, Tesla announced that it had hired Karpathy as Director of AI and Autopilot Vision, reporting directly to Elon Musk and working closely with Jim Keller, who oversaw Autopilot hardware and software. At Tesla he led the computer vision team of Autopilot, which handled all in-house data labeling, neural network training, and deployment on Tesla’s custom inference chip; he also very briefly led the Optimus humanoid robot project.⁠[5][4][4]

In November 2017, Karpathy published “Software 2.0,” one of his most popular essays.⁠[4]

In 2020, MIT Technology Review named Karpathy one of its Innovators Under 35, saying he had extended convolutional neural network techniques so that systems could understand entire scenes of interacting objects rather than recognize single items; rather than scanning surroundings with expensive laser range finders and building virtual maps as most automakers do, Tesla uses ordinary cameras, with nearly 50 neural networks constantly processing data from more than a million cars.⁠[2][2][2][2]

In May 2021, Tesla moved the Model 3 and Model Y to a camera-only system called “Tesla Vision,” dropping radar, with some safety features temporarily limited. Consumer Reports then pulled its “Top Pick” status for the two models built after April 27, and the Insurance Institute for Highway Safety (IIHS) removed their “Top Safety Pick Plus” designation. David Friedman, Consumer Reports’ vice president of advocacy, said removing radar meant safety features might not be there when needed; critics also noted that Tesla lacks lidar and that Autopilot had trouble spotting stationary objects and tractor-trailers crossing in front of its vehicles.⁠[8][8][8][8][8][8][8][8][8]

In June 2021, in a keynote at the CVPR 2021 Workshop on Autonomous Vehicles, Karpathy defended the vision-only approach, arguing that it scales better and can work in a wide variety of environments, not just thoroughly mapped ones; radar has poor vertical resolution and gives false readings from things like manhole covers or overpass shadows, while neural networks trained on huge, clean, and diverse datasets can outperform radar on metrics such as depth and velocity.⁠[9][9][9][9]

On July 13, 2022, after a four-month leave of absence, Karpathy announced that he was leaving Tesla. He tweeted that it had been a great pleasure to help Tesla toward its goals over five years and a difficult decision to part ways, during which Autopilot graduated from lane keeping to city streets; he said he had no concrete plans and wanted to revisit his long-term passions around technical work in AI, open source, and education. Around the same time, Tesla told California regulators it was laying off 229 data annotation employees from its Autopilot team and closing their San Mateo office.⁠[10][10][10][10]

In 2023, Karpathy returned to OpenAI and built a new team working on midtraining and synthetic data generation; that year he also gave the “State of GPT” talk at Microsoft Build. In February 2024 he left OpenAI again, saying he would work on personal projects.⁠[4][4][11]

On July 16, 2024, Karpathy announced Eureka Labs, describing it as “a new kind of school that is AI native”: teachers still design course materials, supported and scaled by an optimized AI teaching assistant, with an entire curriculum run on a common platform. Its first product is LLM101n, an undergraduate-level course that guides students through training their own AI, with materials online and plans for both digital and physical cohorts.⁠[12][12][12][12]

Karpathy said his interest in education ran from YouTube Rubik’s cube tutorials to Stanford’s CS231n and his Zero to Hero series, and that his work combining AI and education had so far been side quests to his “real job,” which he would now pursue full-time.⁠[11][11]

On August 20, 2024, Annette Vee, an associate professor of English and director of the English Composition program at the University of Pittsburgh, questioned AI teaching visions exemplified by Eureka Labs. She noted that personalized learning technologies, from the “automatic teacher” of the 1920s to massive open online courses, have repeatedly failed because students lost interest; students need to connect with and be inspired by fellow humans, AI teaching platforms pose data privacy and security risks, and a single bot teaching millions of students could lead to a loss of diversity of thought.⁠[13][13][13][13][13][13][13]

In Karpathy’s vision for Eureka Labs, bots do not replace teachers: teachers design the courses and an AI teaching assistant supports students under that design, with the goal of making it easy for anyone to learn anything.⁠[12][12]

On February 2, 2025, Karpathy posted on X about a new kind of programming he called “vibe coding,” in which one fully gives in to the vibes and forgets that the code even exists, writing code, web pages, and apps by telling an AI what one wants. The term spread quickly; Merriam-Webster added it, noting that it is believed to have been coined by Karpathy and that the practice often comes with bugs and raises doubts about maintainability in professional settings.⁠[14][14][14][14]

In October 2025, in a roughly two-and-a-half-hour podcast interview with Dwarkesh Patel, Karpathy said artificial general intelligence was at least a decade away and that the industry was overpredicting; he said current AI agents “just don’t work,” lacking enough intelligence, multimodality, computer use, and continual learning, called the output “slop,” and warned that if this was not done well, slop could accumulate across software along with more security vulnerabilities. He said his timelines were five to ten times more pessimistic than those at a San Francisco AI house party or on Twitter, but that “ten years should otherwise be a very bullish timeline” for AGI.⁠[15][16][15][16][16][15][16]

The interview drew a strong reaction in the tech community, with one person writing on X that if this interview did not pop the AI bubble, nothing would. Elon Musk challenged Karpathy on X to a coding contest against Grok 5; Karpathy replied that he would much rather use and collaborate with Grok 5 than compete against it.⁠[16][15][15]

On November 6, 2025, Collins Dictionary named “vibe coding,” coined by Karpathy, its Word of the Year for 2025, noting that tech experts were still debating whether the practice is revolutionary or reckless.⁠[17][17][17][17]

On February 12, 2026, Karpathy published microgpt on his blog, training and running GPT in 200 lines of pure, dependency-free Python.⁠[6]

On May 19, 2026, Karpathy announced that he had joined Anthropic, working on pre-training under team lead Nick Joseph and starting a team focused on using Claude to accelerate pre-training research. He said the next few years at the frontier of large language models would be especially formative, that he was excited to get back to R&D, and that he remained deeply passionate about education and planned to resume that work in time.⁠[18][18][18][18]

Karpathy makes AI teaching videos on YouTube in two tracks: the technical track is the “Neural Networks: Zero to Hero” series, which starts from backpropagation and the autograd engine micrograd and builds up, in code and from scratch, the character-level language model makemore, a WaveNet-style network, GPT, and the GPT tokenizer; the general-audience track includes “Deep Dive into LLMs like ChatGPT” on how large language models work and “How I use LLMs” on his own uses. He considers language models an excellent place to learn deep learning because most of what is learned transfers to other areas such as computer vision.⁠[4][4][19][19][19][19]

Karpathy has maintained many open-source projects, including the tiny autograd engine micrograd; arxiv-sanity, which helps researchers filter and get recommendations for arXiv papers, and its rewrite arxiv-sanity-lite; neuraltalk2, an early image captioning project later extended with Justin Johnson to dense captioning; and ulogme, a privacy-first local productivity tracker. He has written three blogs, on GitHub, Medium, and Bear; popular posts include “A Recipe for Training Neural Networks” (2019), “Biohacking Lite” (2020), and “A from-scratch tour of Bitcoin in Python” (2021), which demonstrates a Bitcoin transaction in pure Python.⁠[4][4][4][4][4][4][4]

On his website Karpathy lists the science fiction he has read, sorted by the product of how much he recommends it and how obscure it is; he says he likes hard sci-fi for its technical ideas, world-building, and future forecasting, dislikes flowery descriptions of scenery and appearance, and cannot stand humanoid aliens with faces who speak by sound; he believes AI is the greatest omission from most sci-fi worlds. He calls Ted Chiang’s short story collection “required reading,” while describing Liu Cixin’s Three-Body trilogy as containing several fantastic novel ideas amid soulless characters, narrative and logical inconsistencies, and a really disappointing conclusion.⁠[20][20][20][20][20][20]

Personal website: https://karpathy.ai/ (opens in a new window)⁠[4]

X: Public account: https://x.com/karpathy (opens in a new window)⁠[4][21]

YouTube: Public channel: https://www.youtube.com/@AndrejKarpathy (opens in a new window)⁠[4]

GitHub: Public account: https://github.com/karpathy (opens in a new window)⁠[4]

Eureka Labs: Official website: https://eurekalabs.ai/ (opens in a new window).⁠[12]