Lesson 5 of 7
The LLM as an operating system
6 min read
We keep bolting tools, memory, and files onto the model. Step back and squint: doesn't this start to look like a computer — with the model as its chip?
The model is the CPU
Here's a way to hold the whole picture (an idea popularized by Andrej Karpathy). The model is the CPU — the part that thinks and coordinates. It doesn't act alone; it drives the parts around it. The context window is its RAM: fast working memory, but small — only what's in it right now can be used, and it fills up.
The LLM OS reframes the model as a coordinator, not a lone brain: a CPU driving the memory and devices around it.
Memory, tools, and the web as devices
The rest falls into place. Long-term memory and your files are the disk — storage that outlasts a single chat. Tools and apps are peripherals — the keyboard, printer, and network the CPU reaches out to. And the live web is its network card. The model's job is to move the right things in and out of its small RAM at the right time.
Once you see it as an OS, the limits make sense: a job too big for RAM (the context window) has to be paged in from disk and tools — which is exactly what agents do.
It's an analogy, not literal hardware — but a useful one. It explains why context is precious (RAM is small) and why tools matter (a CPU alone can't do much).
The shape of it
- —The model is the CPU — it thinks and coordinates the parts.
- —The context window is RAM: fast, small working memory that fills up.
- —Memory is disk; tools and the web are peripherals it reaches out to.
In the "LLM OS" picture, what plays the role of RAM — the fast but limited working memory?
Continue in the app
Take the whole Agents & Tools course — tracked
Get your personalized path, progress and streaks in the app — this lesson and every next one, in order.