Andrej Karpathy: The engineer who named the paradigm shifts

Andrej Karpathy was born in Bratislava in 1986 and moved to Toronto as a teenager. He studied computer science and physics at the University of Toronto, took a master's at the University of British Columbia, and earned a PhD at Stanford in 2016 under Fei-Fei Li, working at the intersection of computer vision and natural language, including some of the early convolutional and recurrent networks for image captioning.

The doctorate produced something arguably more consequential than the thesis: CS231n, Stanford's first deep learning course, which Karpathy designed and taught from 2015, putting the lectures online for free. The course became the standard on-ramp for an entire generation of researchers. The pattern was set early: do frontier work, then explain it so clearly that the explanation itself becomes infrastructure.

Karpathy was a founding member of OpenAI in 2015, working as a research scientist on deep learning and generative models. In June 2017 he left for Tesla to direct the Autopilot vision team, spending five years turning a camera-only perception stack into production code inside millions of cars. It was the most demanding applied deep learning job in the world at the time, and it shaped everything he wrote afterward.

In November 2017 he published a short essay, Software 2.0, arguing that a growing share of software would no longer be written as explicit instructions but assembled as datasets and compiled by optimizers into neural network weights. At Tesla he watched the prediction come true inside his own codebase: as the neural networks improved, handwritten C++ was deleted and functionality migrated into the trained stack. Software 2.0 stopped being a metaphor and became a description of an architecture he had personally overseen.

After leaving Tesla in July 2022, Karpathy returned to OpenAI in early 2023, but his most influential work in this period happened in public. His Neural Networks: Zero to Hero video series walked viewers through building backpropagation and language models from scratch, and in January 2023 he released nanoGPT, a rewrite of his earlier minGPT: about six hundred lines of readable PyTorch that could reproduce a small GPT-2.

nanoGPT became the canonical reference implementation for understanding transformers. At the exact moment the field's frontier models were becoming closed, capital-intensive, and opaque, Karpathy published a version anyone could read in an evening. The pedagogical wager was deliberate: if the technology is going to reshape software, the most leveraged thing one person can do is grow the population that actually understands it.

In February 2025 Karpathy coined the term vibe coding for the practice of describing software in natural language and letting an LLM write the code. Within months the phrase had escaped into the general culture; Collins Dictionary named it Word of the Year for 2025. Then, in June 2025, his keynote at Y Combinator's AI Startup School completed the trilogy he had started in 2017: Software 3.0 is programming the model directly, in English, with prompts as programs. He described LLMs as a new kind of computer, part utility, part fab, part operating system, circa the 1960s of its own computing history.

The same talk carried his characteristic correction to the hype cycle. Against declarations that 2025 would be the year of agents, he countered that this is the decade of agents, pointing to his own Tesla experience: a perfect self-driving demo in 2013 did not mean self-driving was solved in 2013. In an October 2025 interview with Dwarkesh Patel he extended the argument, calling current agent capabilities unimpressive relative to the marketing and criticizing reinforcement learning as a wasteful way of extracting supervision. The field's best-known enthusiast for the technology is also one of its most credible sources of timeline skepticism.

In July 2024 Karpathy founded Eureka Labs, an AI-native education company, with an initial course called LLM101n aimed at teaching students to train their own models. It is CS231n's logic applied to the LLM era: the course and the company are one artifact, with AI assistants built into the teaching itself.

In October 2025 he released nanochat, a minimal end-to-end implementation of a ChatGPT-style pipeline, from tokenizer through pretraining and fine-tuning to a working chat interface, in roughly eight thousand lines, trainable for on the order of a hundred dollars of compute. Where nanoGPT showed how a transformer works, nanochat shows how the whole product works. The pedagogy has scaled up with the technology, and the answer to closed labs remains the same: here is the entire thing, read it.

Karpathy's framings have an unusual track record. Software 2.0, dismissed by some in 2017 as a catchy label, accurately described the industrial stack that followed, and Software 3.0 gave engineers a working mental model within a year of the agent era beginning. His educational artifacts have held up the same way: CS231n and nanoGPT are still where serious people start. His Tesla tenure is the honest counterweight: Autopilot improved enormously under his leadership, but full autonomy took far longer than the company promised, and he now cites that experience as the source of his caution about agents.

The contested questions are current ones. His skepticism about near-term agents and reinforcement learning puts him at odds with the labs' claims, and it is too early to know whether the decade of agents framing is wisdom or a conservative bias formed by one hard problem. What is not contested is the method: build the thing, name what you learn, publish the explanation. In a field drowning in secondhand takes, Karpathy remains a primary source.