AI Primer

Under the hood of the AI systems you use every day.

You've been using ChatGPT, Claude, or Gemini long enough to have intuitions — what works, what doesn't, when the model seems brilliant and when it falls apart. But the system behind the text box is still a mystery.

This primer opens it up. Nine short chapters, each building on the last: how a text predictor becomes a chatbot. How a chatbot gains tools. How tools enable autonomous agents. And the finite context window that governs them all.

The Plain LLM — What an LLM actually is: a stateless probability machine that predicts the next word.
Multimodality — Images, PDFs, and audio all get compressed into tokens — with trade-offs in detail, cost, and reliability.
The Chatbot — The LLM has no memory; the application fakes it by resending the full conversation every time.
The System Prompt — A hidden instruction at the start of every conversation that steers how the model behaves.
Structured Output — Making the model respond in a fixed structure instead of freeform text, so software can act on it.
Tool Use — The model can call external tools — search, calculate, look things up — but the application is what actually runs them.
The Agentic Loop — Instead of one tool call, the model plans, acts, observes, and repeats until done.
Multi-Agent — When one agent isn't enough, an orchestrator delegates to specialized subagents.
Context Engineering — Everything competes for the same finite context window. This is the discipline of filling it well.

This grew out of years of working with these systems and explaining them to the people around me — and never finding quite the right guide to point them to. It deliberately simplifies — the goal is a useful mental model, not a textbook.

Written by Myke Näf — computer scientist turned VC (Übermorgen Ventures) who still can't leave the terminal alone. With considerable help from the machines described herein.

Found an error? Have a better way to explain something? Contribute on GitHub — contributions welcome.