LLM as Operating System

Karpathy's framing: a large language model behaves like the kernel of a new kind of OS, not just a chatbot. The model is the CPU; its context window is RAM (scarce, must be paged in/out); files on disk are persistent state; external tools and APIs are peripherals/drivers; and other model calls are subprocesses.

This reframes a knowledge vault. The markdown is the disk. Claude Code is the kernel that reads and writes it. Slash-commands are installed programs. A daily note is the session log. Dashboards are the process monitor. Background jobs (cron, hooks) would be daemons.

The payoff is composability: instead of one-off automations, you get a small set of reliable "programs" operating over a shared filesystem under a consistent contract (the kernel config in CLAUDE.md).

---

Source: Andrej Karpathy — "LLM OS" talks/posts See also: Maps of Content, Atomic Notes