Andrej Karpathy

Who he is

Andrej Karpathy is an AI researcher and teacher whose career has run through several of the institutions that shaped modern machine learning. He was born on 23 October 1986 in Bratislava, then part of Czechoslovakia, and his family moved to Toronto when he was 15^[1]. He is known both for research and engineering work at OpenAI and Tesla and for courses and lectures that many people have used to learn deep learning^[1][4].

Education

He studied at the University of Toronto, earning bachelor's degrees in computer science and physics in 2009; his personal site lists the degree as running from 2005 to 2009, with a minor in mathematics^[1][3]. He then did a master's at the University of British Columbia from 2009 to 2011, working on learning controllers for physically simulated figures^[1][3]. That problem is about training a model to produce the control signals that make a simulated body move, instead of hand-designing those signals.

He moved to Stanford for a PhD, completed in 2015 under Fei-Fei Li, focusing on convolutional and recurrent neural networks^[1][3]. Convolutional networks are the architecture that made image recognition work well, and recurrent networks were then the standard tool for sequences such as text.

Teaching

At Stanford he was the primary instructor of CS231n, the deep-learning course, from 2015 to 2017, and enrolment grew from 150 to 750 students over that period^[1]. He later became known for the online course "Neural Networks: Zero to Hero" and for a YouTube channel of lectures on large language models and AI^[4]. This teaching work runs alongside his research career rather than behind it: the same material that appears in a lecture is often the material he is working through himself.

OpenAI and Tesla

He was a founding member of OpenAI in 2015 and worked there as a research scientist until 2017^[1][2][3]. In June 2017 he became Tesla's director of AI, reporting to Elon Musk^[1]. In that role, which he held from 2017 to 2022, he led the computer vision team for Autopilot and, briefly, Tesla Optimus^[3]. He announced his departure in July 2022^[1].

Back to OpenAI, then Eureka Labs

He returned to OpenAI on 9 February 2023 to build a team working on midtraining and synthetic data generation, and left again in February 2024^[1][3]. Midtraining refers to training stages that sit between large-scale pre-training and the final fine-tuning of a model, and synthetic data is data produced by models rather than collected from people.

On 16 July 2024 he announced Eureka Labs, an AI education company whose first product was the course LLM101n^[1]. Eureka Labs was started in 2024 to apply AI assistants to education, and TechCrunch noted that it was unclear whether he would continue with it^[4]. In February 2025 he coined the term "vibe coding"^[1].

Joining Anthropic

On 19 May 2026 he announced that he was joining Anthropic to work on pre-training^[1][4]. TechCrunch reported that he announced the move on X and would start the week of the announcement^[4]. He works on Anthropic's pre-training team under team lead Nick Joseph^[4], and an Anthropic spokesperson said he will start a team focused on using Claude to speed up pre-training research^[4]. Pre-training is the long, expensive first stage in which a language model learns from large amounts of text before any task-specific tuning.

He wrote that "the next few years at the frontier of LLMs will be especially formative"^[4]. He also said he remains deeply passionate about education and plans to return to that work in time^[4].

Where he sits in the field

Karpathy's path is unusual in how often it crosses between building systems and explaining them. He was present at the founding of OpenAI, spent five years on the vision stack of a car company's driver-assistance system, started a company aimed at teaching with AI, and has now moved to the pre-training side of a major model developer^[1][3][4]. The common thread is training: how a network is taught from data, whether the data is images from a car, a simulated body learning to move, or the text a large model reads before it is tuned.