brightlabInside the context machine
A language model reads one long sequence of numbers and predicts the next one. Chat, memory, search and agents are all built on that one idea. Let’s take it apart, with real models and real data.
The journey · about 70 minutes
- 01Your words become numbersTokens
- 02One token at a timePrediction
- 03What’s it trained on?Training data
- 04Training is next-word guessing, at scaleHow it learns
- 05A chat is one long sequenceChat templates
- 06The context you don’t seeHidden context
- 07Fluent is not the same as trueHallucination
- 08Numbers live on circlesInside the network
- 09Better context in, better predictions outTakeaways
Running it with a class? Put it on the big screen with P (present mode) and move with ← →. Students can follow along on their own devices.
Everything here is real: a real tokenizer, real training data and real model probabilities, recorded or running in your browser. Nothing students type is sent anywhere.