LLM OS
The LLM OS is Andrej Karpathy‘s framing of large language models not as chatbots or word generators, but as the kernel process of an emerging operating system — a new computing paradigm analogous to traditional operating systems but built around language models as the central processing element.
The analogy
| Traditional OS | LLM OS |
|---|---|
| Kernel process | LLM |
| RAM (working memory) | Context Window |
| Disk / internet (storage) | Training data + RAG-accessible documents |
| Peripherals | Tools: web search, calculator, Python, image generation |
| Multi-threading | Multiple LLM instances / agent chains |
| User space / kernel space | System prompt (operator) / user turns |
| Windows / macOS (proprietary) | GPT series, Claude series, Gemini series |
| Linux (open-source) | Llama series, Mistral, DeepSeek |
What this framing adds
The chatbot framing focuses on conversation. The OS framing focuses on computation — on the LLM as an orchestrator of tools, memory, and sub-processes. In this view:
- The LLM doesn’t just answer questions; it pages information in and out of context, delegates computation to specialised tools, and coordinates multi-step workflows.
- The context window is RAM: precious, finite, and swept clean when a session ends (like rebooting).
- Tools are peripherals: web search is the network stack; the Python interpreter is a compute co-processor.
Evolution into Software 3.0
The LLM OS framing (introduced in Intro to Large Language Models, Nov 2023) prefigures the Software 3.0 framing (articulated in From Vibe Coding to Agentic Engineering, 2025). The LLM OS describes the architecture; Software 3.0 describes the programming paradigm that uses that architecture.