Andrej Karpathy is an AI researcher and educator who was a founding member of OpenAI and later served as senior director of AI at Tesla, where he led the Autopilot vision team. He is widely known for his deep learning teaching, including the Stanford CS231n course and his “zero to hero” neural network video series, and for coining the term “vibe coding.” In July 2024 he founded Eureka Labs, an AI-native education startup. As of May 2026 he joined Anthropic to build a team using Claude to accelerate model pre-training research.
Recent
-
2026-08-02 Showed Claude Opus building a procedural 3D Lord of the Rings world from one paragraph – roughly 5,500 lines of Three.js, about two hours of agentic runtime, playable at karpathy.ai/lotr-movie. Framed it as evidence AI has moved past simple prompts into sustained creative construction. Benzinga
-
2026-07-26 Removed Anthropic affiliation from X bio, sparking rapid speculation of departure. Denied it within hours (“weird misinformation”), attributing the change to a routine profile update. Incident coincided with Anthropic’s public refusal to sign the 50-company open-weights coalition letter, creating visible tension given his history as an open-source advocate. BigGo Finance · EU 36Kr
-
2026-06-28 An updated CLAUDE.md attributed to Karpathy began circulating today, expanding the original 4 rules to 10 with a new agent self-check protocol for coding loops. He has not confirmed it publicly. The original 4-rule file already hit 100K GitHub stars. TechTimes · Raytar on X
-
2026-06-24 Called Claude Tag for Slack “the 3rd major redesign of LLM UIUX”: web chatbot was first, standalone app was second, the third is a persistent async agent embedded org-wide. X · Benzinga
-
2026-06-24 Posted that X’s algorithmic feed is “actively making the platform worse” and compared it to Reddit; elonmusk publicly agreed, replying “We need a complete overhaul of the algorithm.” American Bazaar
-
2026-06-22 Congratulated simonw on 23 years of continuous blogging, called his LLM coverage “really excellent,” and nudged followers to sponsor him on GitHub. X
-
2026-06-17 autoresearch repo is trending on GitHub with 226 new stars today — renewed attention on the March 2026 project that runs ML experiments autonomously overnight on a single GPU. Reported an 11% improvement in nanochat “Time to GPT-2” in his original thread. GitHub
-
2026-06-10 Called Claude Fable 5 “a major-version-bump-deserving step change forward (imo of the same order as Claude 4.5 was in November)” on launch day, noting it handles very ambitious long-horizon tasks without hand-holding. (tweet)
-
2026-06-06 Published writeup from his Sequoia Ascent 2026 fireside chat. Pushed the “Software 3.0” framing: LLMs don’t just speed up existing work, they enable product categories that couldn’t exist before. Drew a distinction between “vibe coding” (lower floor, anyone can build) and “agentic engineering” (professionals staying accountable for output). Put December 2025 as the tipping point when agents started reliably producing correct code. (blog, thread)
-
2026-06-06 Separately posted that he’s “never felt this much behind as a programmer” — went viral with millions of views. (tweet)
-
2026-06-06 Joined Anthropic’s pre-training team (May 19). Working on using Claude to accelerate pretraining research. TechCrunch · Fortune