Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI

Andrej Karpathy discusses his experience with AI coding agents and the dramatic shift in software development workflows since December, describing a transition from writing code manually to delegating almost entirely to agents. He shares his 'Dobby' home automation claw project built in a few prompts, explores the concept of AutoResearch (autonomous hyperparameter optimization loops that outperformed his own two decades of tuning experience), and discusses parallelizing multiple agents across repos. He reflects on LLM jaggedness — models excelling at verifiable tasks but stagnating on unverifiable ones like jokes — and speculates on model speciation, distributed untrusted compute pools for collaborative AI research (analogous to Folding@Home), and the future of software engineering jobs via Jevons paradox. He also shares his perspective on why he works outside frontier labs rather than inside them.