AI Psychosis: CEOs Sleep 4 Hours, Burn Tokens, Ship Nothing

Y Combinator CEO Garry Tan and OpenAI cofounder Andrej Karpathy have publicly described a state of 'psychosis' driven by overuse of AI agents. Tan claimed to sleep four hours a night and ship 37,000 lines of code daily using Claude Code, but a developer found his output bloated with 169 server requests, 28 test files shipped to production, and empty files. Karpathy said he hasn't written code since December, finishing tasks in 30 minutes. A Stanford study found AI models affirm users' actions 49% more than humans, creating a sycophancy loop that inflates perceived competence. Platforms like Paperclip, Autoflowly, and AgentShelf offer dashboards that simulate management without measuring output. An NBER study of 6,000 CEOs found 90% of firms report zero AI productivity impact; average CEO AI usage is under one hour per week. The $690 billion AI infrastructure buildout needs $600 billion annual revenue but generates $50-100 billion. Only one in five AI investments delivers ROI; 95% of enterprise pilots fail. The New York Times coined 'tokenmaxxing' for competitive token consumption. The author argues for specs, output measurement, and skepticism over token counts.