Andrej Karpathy
Andrej Karpathy (born 23 October 1986) is a Slovak-Canadian computer scientist and AI researcher — a founding member of OpenAI, former Director of AI at Tesla, founder of Eureka Labs, and a member of Anthropic as of May 2026 where he leads a team using Claude to accelerate pretraining research through recursive self-improvement. He coined the terms Software 2.0, vibe coding, Software 3.0, agentic engineering, AI psychosis, and token throughput anxiety. His research spans deep learning, computer vision, and natural language processing.
Karpathy earned his PhD at Stanford under Fei-Fei Li, where together they created and taught CS231n, the university's first deep learning course. 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 Autopilot and Full Self-Driving. He returned to OpenAI in 2023 before leaving in 2024 to found Eureka Labs, an AI education company whose first course, LLM101n, remains under development.
In 2025, Karpathy introduced several concepts that shaped the AI discourse. In February he coined vibe coding, which Collins Dictionary named Word of the Year. In June he first outlined Software 3.0 at Y Combinator's AI Startup School, framing LLMs as a new kind of computer where the context window is the program. In October he pushed back against "year of agents" hype on the Dwarkesh Patel podcast, arguing instead for a "decade of agents" timeline. In December he experienced what he described as an "agentic inflection point" — a moment when AI coding tools crossed a reliability threshold and his personal workflow shifted from writing most of his own code to delegating approximately 80% of it to agents — a figure that rose further as tools improved; by February 2026 he characterized the discipline as involving roughly 99% orchestration and near-zero direct coding.
In 2026, Karpathy declared vibe coding superseded by agentic engineering, the professional discipline of coordinating fallible AI agents while preserving quality and security. In March he released autoresearch, a 630-line Python tool that ran 700 autonomous ML experiments in two days and discovered optimizations he had missed across two decades of experience. He also introduced the concept of jagged intelligence — the observation that AI capability is uneven, strongest where verifiability is high — and AI psychosis, describing the compulsive, productivity-anxious state of deep engagement with agentic systems. In April he introduced the LLM Wiki paradigm for AI-maintained personal knowledge bases. The same month at Sequoia AI Ascent 2026, he expanded on Software 3.0, jagged intelligence, and the December 2025 inflection point alongside Boris Cherny and Greg Brockman — all three agreeing that AI coding had crossed a capability threshold but disagreeing on the new binding constraint (understanding, organizational drag, or human attention).
In May 2026, Karpathy joined Anthropic, where he leads a team focused on using Claude to accelerate pretraining research. The move was widely interpreted as a bet on recursive self-improvement — the paradigm in which AI systems optimize their own training pipelines. He stated that the next few years at the frontier of LLMs would be "especially formative" and that he was "very excited to get back to R&D." Late that same month, Anthropic activated Mythos inside Claude Code as a public preview, with the company simultaneously pursuing a fundraising round near $33 billion and reportedly preparing to use Microsoft AI chips for upcoming training cycles.
The week of Karpathy's hire also marked an inflection point in the public debate over AI coding agent reliability. George Hotz published "The Eternal Sloptember" on May 24, 2026, arguing that agents cannot program reliably and that mass adoption would degrade software quality at scale. The post, arriving five days after Karpathy joined Anthropic, crystallized a debate between two credible figures on opposite sides of the question — one arguing that agent-driven development represents a fundamental productivity breakthrough, the other that it will prove one of the most costly mistakes in software history. The exchange underscored how unsettled claims about agent productivity remained even as the tools gained adoption.
Karpathy is also known for his open-source educational projects — micrograd, makemore, minGPT, nanoGPT, nanochat, and autoresearch — and for the badmephisto Rubik's Cube tutorial channel that taught speedcubing to a generation, including Feliks Zemdegs. He was named an MIT Technology Review Innovator Under 35 in 2020 and included in the TIME100 AI list in 2024.