Andrej Karpathy
Andrej Karpathy is a Slovak-Canadian computer scientist and AI researcher who has worked at OpenAI, Tesla, and Anthropic. See [[people/andrej-karpathy]] for the full biography.[^c1][^c2][^c3] He is one of the most widely recognized figures in artificial intelligence, known for pioneering deep learning education, coining influential concepts that shaped the AI discourse, and moving between frontier AI labs at critical moments in the field's development.
Karpathy earned his PhD at Stanford University under Fei-Fei Li, where together they created CS231n, Stanford's first deep learning course.[^c1] He interned at Google Brain, Google Research, and DeepMind before joining OpenAI as a founding member in 2015. In 2017, Elon Musk recruited him to Tesla as Director of AI, where he led the computer vision team behind Tesla's autonomous driving systems. He returned to OpenAI in 2023 before departing in 2024 to found Eureka Labs, an AI education company.
In 2025, Karpathy introduced a series of concepts that defined the year's AI discourse. He coined "vibe coding" in February, describing a programming style where developers describe intent in natural language and accept AI-generated output without thorough review — a term that Collins Dictionary named Word of the Year. In June he formally outlined the Software 3.0 paradigm at Y Combinator's AI Startup School, framing large language models as a new kind of computer. In October he pushed back against industry hype on the Dwarkesh Patel podcast, arguing for a "decade of agents" rather than a "year of agents" timeline grounded in the reliability engineering lessons learned from autonomous driving.
The pace accelerated dramatically in 2026. In February, Karpathy released microGPT, a 200-line dependency-free GPT implementation, and declared vibe coding superseded by [[concepts/agentic-engineering|agentic engineering]], the professional discipline of coordinating fallible AI agents at scale.[^c4] In March, he appeared on the No Priors podcast and stated he had not typed a line of code since December 2025, describing English as his new programming language and AI agents as his new compiler — a statement that reverberated across the software industry.[^c8] He introduced the cognitive core concept — the prediction that a lean reasoning engine of roughly one billion parameters could achieve strong reasoning when stripped of memorized knowledge and paired with external tools, a thesis fulfilled ahead of schedule by dense small models such as Qwen 3.5 4B.[^c5][^c7] He released Auto Research, a 630-line Python tool that ran hundreds of autonomous machine learning experiments. He introduced the LLM Wiki paradigm, AI psychosis, token throughput anxiety, and jagged intelligence. In May, a benchmark of microGPT on an FPGA showed a single M4 Max MacBook P-core running a C implementation achieving 71x the throughput of a custom Cyclone V FPGA implementation, demonstrating the importance of competent commodity-hardware baselines.[^c11]
In May 2026, Karpathy joined Anthropic's pretraining team under Nicholas Joseph, where he leads a sub-team using Claude to accelerate pretraining research — a concrete implementation of recursive self-improvement.[^c2][^c3] In June, he reviewed Anthropic's Claude Fable 5, calling it a "major version upgrade" comparable to Claude 4.5.[^c6] Days later, the US Commerce Department issued an export control directive requiring Anthropic to suspend access to Fable 5 and Mythos 5 by any foreign national, including foreign national Anthropic employees. As a Slovak-born Canadian citizen, Karpathy was named among key Anthropic personnel who could be affected by the order, raising questions about whether he could access the models he was hired to improve and highlighting the broader tension between US export control policy and the AI industry's dependence on foreign-born talent.[^c10]
Anthropic itself experienced explosive growth in 2026, reporting a Q2 revenue projection of $10.9 billion and a $47 billion run rate, raising $65 billion in Series H funding at a $965 billion post-money valuation. Across the industry, governance benchmarks showed widening gaps: the Stanford AI Index documented the Foundation Model Transparency Index dropping from 58 to 40, AI incidents rising 55 percent to 362, and 80 of 95 notable models shipping without training code, while the FLI AI Safety Index graded no company above C+ overall and no company above D in existential safety planning.
Karpathy's framing of the Software 1.0, 2.0, and 3.0 eras, and his distinction between vibe coding (which "raises the floor") and agentic engineering (which "raises the ceiling"), provided a widely adopted vocabulary for understanding the transformation of software engineering in the age of AI agents.[^c4][^c9] By June 2026, the [[projects/andrej-karpathy-skills]] CLAUDE.md file — four rules distilled from his observations of LLM coding pitfalls — had become one of the fastest-growing repositories in GitHub history, exceeding 169,000 stars.
Karpathy maintains numerous open-source educational projects — including micrograd, makemore, nanoGPT, nanochat, and the Zero to Hero video series — and is a small angel investor in Decart, a real-time video and world models lab. He was named an MIT Technology Review Innovator Under 35 and included in the TIME100 AI list.