← Learn
Learn · AI

How LLMs work

Seven lessons on how large language models work and how to build with them well: tokens and the context window, prompting, getting JSON your code can use, tool calling, temperature, cost and speed, and testing the output with evals. No maths beyond arithmetic, and no particular product.

  1. Tokens and context windows

    What a language model actually does, what a token is, and why the context window limits everything.

    6 questions

  2. Prompting

    System prompts, being specific, examples, clear separators, and treating prompts like code.

    6 questions

  3. Structured JSON output

    Getting data your program can use: JSON with a fixed shape, validating it, and what to do when it's wrong.

    6 questions

  4. Tool calling

    Letting a model ask your code to look things up or act, and keeping your code in charge.

    6 questions

  5. Temperature and sampling

    How the next token is picked, what temperature and top-p change, and which settings suit which jobs.

    6 questions

  6. Cost and latency

    How LLM calls are priced, where the waiting time goes, and the levers that trade quality for speed and cost.

    6 questions

  7. Evaluating output quality

    Eval sets, code checks, model graders and people, the right metric, and testing every change.

    6 questions
More info

How it works

Each lesson is a short read with examples, then a few practice questions. Some ask you to type an answer (a number, an address, what a piece of code prints), some have you pick one, and some have you tap steps or lines into the right order. You can try again as often as you like; after two misses there’s a button to show the answer.

What do I need to know first?

Nothing for the ideas. A few examples are short Python programs (you only need to read them); the programming basics course covers everything they use.

The other AI and data engineering courses

Where is my progress saved? In this browser, like the rest of the site. Nothing to sign up for.

“Python” is a registered trademark of the Python Software Foundation. This course isn’t made by or affiliated with it.