Andrej Karpathy co-founded OpenAI, led Tesla's Autopilot team, coined vibe coding, and founded Eureka Labs. A profile of the AI researcher who made deep learning accessible to millions.
On a February night in 2026, Andrej Karpathy went to sleep while his computer stayed up. By morning, an AI agent he had built called AutoResearch had run 700 experiments overnight on a neural network model Karpathy had spent years tuning by hand. The agent found optimizations he had missed, committed the improvements, and kept going.
No human was in the loop.
That scene captures something essential about Karpathy: he does not just predict where AI is going. He builds himself into the proof.
Andrej Karpathy was born in Bratislava in 1986 and moved to Toronto with his family at 15. He first showed a gift not for algorithms but for explanation.
In 2006, long before he had a PhD or a job title, he posted Rubik's cube tutorials on YouTube under the channel badmephisto. The videos went viral among the speedcubing community and have since collected over 9 million views.
That instinct to teach before being asked would become the defining thread of his career.
Karpathy completed his BSc in Computer Science and Physics at the University of Toronto in 2009, studying under Geoffrey Hinton, the researcher who helped make deep learning practical. He earned his MSc at the University of British Columbia in 2011, then joined Stanford University for his PhD under Fei-Fei Li, focusing on the intersection of computer vision and natural language processing. His dissertation, "Connecting Images and Natural Language," was completed in 2015.
Before finishing his PhD, Karpathy interned at Google Brain in 2011, Google Research in 2013, and DeepMind in 2015. By the time he graduated, he had already built a reputation in three of the four major AI labs of the era.
In 2015, Karpathy designed and taught the first deep learning course at Stanford: CS231n: Convolutional Neural Networks for Visual Recognition. The course started with 150 enrolled students and grew to 750 by 2017. Video recordings have been watched over 800,000 times on YouTube.
Thousands of today's AI researchers and engineers trace their foundations to that course.
That same year, he joined the founding team at OpenAI, working as a research scientist on foundational deep learning and reinforcement learning. He was one of seven original researchers at the lab.
In 2017, Tesla hired him directly to lead their Autopilot vision team, reporting to Elon Musk. Over five years, his team moved Tesla's self-driving stack from hand-coded rule systems to end-to-end neural networks trained on real-world fleet data, handling all in-house data labeling, training, and deployment on Tesla's custom inference chip. The work culminated in the 2021 Tesla AI Day, one of the most-watched technical presentations in the automotive industry.
He left Tesla in July 2022 and immediately turned to YouTube. His series Neural Networks: Zero to Hero walks viewers through building GPT-style models from scratch. His channel has reached 1.34 million subscribers and over 27 million views across just 17 videos.
In December 2022, he released nanoGPT, a ~300-line Python implementation that reproduces GPT-2 (124M parameters). The repo collected 57,090 stars on GitHub.
He followed it in 2025 with nanochat, a full-stack ChatGPT-style pipeline designed to run on a single 8xH100 node. NanoChat has 52,409 stars.
After a brief return to OpenAI from February 2023 to February 2024, where he led a new team focused on midtraining and synthetic data, Karpathy made his next move public in July 2024.
"Eureka Labs is the culmination of my passion in both AI and education over about 2 decades."
– Andrej Karpathy (Reuters)
Eureka Labs is built on a single premise: the best teachers in every subject are extremely rare, and AI can extend their reach to anyone who wants to learn. Human experts design the course material; an AI Teaching Assistant guides students through it.
The startup's first product is LLM101n, an undergraduate-level course that teaches students to train their own AI from scratch. Course content is free and permissively licensed; revenue comes from running digital and physical cohorts.
In February 2025, Karpathy posted a few sentences on X that named something millions of developers were already doing.
"There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists."
– Andrej Karpathy (X, February 2025)
The term went global within days. Merriam-Webster listed it as slang in March 2025. Collins English Dictionary named it Word of the Year for 2025.
The underlying idea, that natural language is now the dominant programming interface, was something Karpathy had flagged two years earlier with a tweet: "The hottest new programming language is English."
That moment crystallized a pattern running through his entire career: Karpathy consistently names what the field has not yet named, and does so with enough precision that the name sticks.
His influence extends across multiple layers. As an educator, his open-source repositories function as university syllabi for self-taught engineers around the world. You can see that pattern reflected in the AI statistics shaping how practitioners think about the field today.
As a practitioner, he shaped how Tesla approached autonomous driving, which informed how the entire industry thought about neural network pipelines for safety-critical systems.
"His biggest impact on the world, however, may come not from his research but from his role as one of the world's foremost educators on neural networks."
– TIME100 AI 2024
Karpathy was named one of MIT Technology Review's Innovators Under 35 in 2020 and one of TIME's 100 Most Influential People in AI in 2024. Those awards track reputation. What is harder to quantify is the number of researchers who built their first neural network by following one of his tutorials.
"I'm a little bit obsessed with coming to the core of things."
– Andrej Karpathy (TIME100 AI 2024)
He credits that instinct to his physics education, which trained him to find the simplest explanation for a complex system, strip away everything that is not structural, and make the mechanism visible.
In December 2025, Karpathy described what he called a "hard flip." Since that month, he has written essentially no code by hand. He now orchestrates AI agents, spending roughly 16 hours a day directing them rather than doing the work directly. He frames the shift not as a loss of craft but as a change in job description: from writing code to managing the systems that write code.
His project AutoResearch is the sharpest expression of where he sees AI going. In March 2026, AutoResearch ran autonomously overnight and completed 700 experiments on a nanochat model, finding improvements Karpathy had missed after years of hand-tuning. The repository has attracted 77,718 stars, making it his most-starred project.
"Programming is becoming unrecognizable. This is nowhere near 'business as usual' time in software."
– Andrej Karpathy (Business Insider, 2026)
He is building Eureka Labs into a curriculum platform where the model of one expert teaching millions is the default rather than the exception. He has said clearly that he does not want to gatekeep educational content. Course material is free; the business runs on cohort fees.
Whether he can make that model financially durable is one of the more interesting questions in AI education over the next several years. His approach prefigures the zero-human company model where AI extends human reach rather than replacing human judgment.
"People are obviously pre-money if they're trying to learn a lot of stuff. So I get paid in people thanking me."
– Andrej Karpathy (TIME100 AI 2024)
The through-line from 2006 Rubik's cube tutorials to 2026 AutoResearch is not hard to see. Karpathy has always been drawn to the part of a complex system that most people cannot see clearly, and to making that part visible. Now he is applying that same instinct to AI itself, building tools designed to help researchers understand and improve models without needing to sit inside the loop at every step.
What makes him unusual is that he keeps doing this publicly, with code anyone can run and explanations anyone can follow. The best AI harness tools being built today trace intellectual lineage back to his work on AutoResearch and nanoGPT.
That is not an accident. It is how he works.
The most current artificial intelligence statistics for 2026, covering market size, investment, adoption, workforce impact, productivity, healthcare, and public trust. 65 sourced data points from McKinsey, OECD, Goldman Sachs, Pew Research, and more.
