Andrej Karpathy

December 2025 is the moment «Andrej Karpathy» identifies as the switch. During a break he noticed that the code chunks the new models produced came out right on the first try, so he stopped correcting them and let the agent keep going. In that same interview the co-founder of «OpenAI» and former architect of «Tesla Autopilot» says he has never felt more behind as a programmer.

He coined «vibe coding» in February 2025, and the term quickly stopped being his. Kate Holterhoff built an account of the summer of 2025 around it; Addy Osmani's Google whitepaper places it at one end of a spectrum whose other end is agentic engineering. Karpathy draws a similar line: «vibe coding» raises the floor, agentic engineering preserves the quality bar.

His recurring argument concerns jaggedness. Labs train with reinforcement learning on verifiable domains, which creates peaks of capability and holes elsewhere: Opus 4.7 refactors 100k lines and then advises walking 50m to the car wash. Agents, in his description, are interns with excellent recall and no taste. He rejects the animal metaphor for LLMs and calls them ghosts, statistical simulation circuits.

Other people turn his observations into artifacts. The «andrej-karpathy-skills» repository compiles his notes on model failure modes into four «Claude Code» guidelines and drew over 3,400 stars in days. Ryan Sarver cites Karpathy's LLM wiki as the model for a readable markdown agent memory. His own formula: "You can outsource your thinking but you can't outsource your understanding."