Refer back while you build
One page per concept: the idea, a mental model, what it changes when you build, and the gotchas. Each links to the moment it's explained.
The context window
Everything the model can see for one prediction. It has no memory between calls — if it isn't in the context, it doesn't exist.
Next-token prediction
A language model predicts the next piece of text from everything it can see, one piece at a time.
Non-determinism
The model outputs probabilities and samples from them, so the same input can produce different outputs. Design for the distribution, not the demo.
Tokens
Models read and write in tokens — chunks of text about ¾ of an English word. Cost, speed, and limits are all measured in tokens.
Training vs. inference
Training is when the model learns (at the lab, for months). Inference is using it — and during inference the model doesn't learn anything.
More concepts arrive with each episode. Browse the full glossary →