LLM OS
Andrej Karpathy's mental model for dropping the 'LLM as chatbot' view and seeing the model as the processor of a new kind of computer. In the analogy the model is the CPU, the context window is RAM, tools are peripherals (calculator, browser, code execution), external memory is the hard drive, and the agent is the orchestrator running the whole thing. Karpathy calls the shift 'Software 3.0': you no longer program in lines of code, you drive a system whose core is a language model. It is not a product to install, it is the lens that explains why context, memory and tools matter more than the perfect prompt.
Strengths
- Gives a clear lens: CPU, RAM, disk, peripherals, orchestrator
- Moves the effort from the perfect prompt to what really matters: context, memory, tools
- Ages well: the analogy stays true even as the models change
Limitations
- The 'OS' label is hijacked by the hype to sell plain API wrappers
- An analogy is not an architecture: no turnkey LLM OS exists, you assemble it yourself
Best for
- A PM or designer who wants to grasp where the effort really goes when moving from assistant to system
- A tech lead designing a durable AI setup who wants an architecture compass