Reference

Built-in Explainers

Eight ideas that are much easier to see than to read about. Each one steps through in four or five frames. Press play, or step through at your own pace.

These are not videos. They are drawn in the page, so they are instant to load, readable on a phone, and you can hold a frame for as long as you need rather than scrubbing a timeline. Each one also appears inline on the track where it belongs.

Tokens and the context window

Where almost every confusing failure in your first month comes from. Appears on the beginner track.

How a model writes a sentence

One token at a time, with no plan. Worth watching before you argue with a model about a fact. Appears on the beginner track.

What attention actually does

The mechanism the whole field is built on, in five frames. Appears on the intermediate track.

RAG in four moves

Question in, vectors, nearest chunks, answer. Appears on the beginner track.

The agent loop

What an agent actually is, and where the harness sits around it. Appears on the intermediate track.

Same model, two harnesses

Why two teams with identical models get very different results. Appears on the advanced track.

Vector search against a knowledge graph

A two-hop question that similarity search cannot answer. Appears on the intermediate track.

Recursive language models

What you do when the input is a hundred times larger than the window. Appears on the advanced track.

About these

Each explainer is one SVG whose parts are revealed step by step, with the captions written to carry the explanation on their own. They are deliberately simplified. Where a simplification would mislead, the track page around it says so in the text.

To add one, open assets/explainers.js, copy an existing definition, and drop <div class="xp" data-xp="your-id"></div> wherever it belongs. There is no build step.

Maintained by the AI team. Corrections and additions are welcome as pull requests.