brightlabNew · Guided tours for every level
Take it apart
A chatbot reads one long line of tokens and guesses the next one. These tours take that machine apart, scene by scene, with a real tokenizer, real training pages and real model probabilities.
Run a tour on the big screen and talk it through, or let students explore on their own devices.
How does a talking computer work?
Words become blocks, the computer guesses the next word, and it only knows what we tell it.
- Words become blocks
- It read lots and lots of stories
- Guess the next word
- Turn up the silly dial
- It only knows what you tell it
- Can it make things up?
How a chatbot writes
Tokens, next-word guessing, training, and the hidden words that shape every answer.
- Your words become numbers
- One token at a time
- What did it read?
- How does it learn?
- A chat is one long tape
- The words you don’t see
- When it doesn’t know
Inside the context machine
From tokens and training to chat templates, hidden context, retrieval and the geometry inside a network.
- Your words become numbers
- One token at a time
- What’s it trained on?
- Training is next-word guessing, at scale
- A chat is one long sequence
- The context you don’t see
- Fluent is not the same as true
- Numbers live on circles
Real, and labelled
Every interactive says where its numbers come from: recorded from SmolLM2 1.7B, live in your browser, or drawn from an open training set. No pretend models.
Private by design
Nothing students type leaves the device. The tokenizer and the optional live model run in the browser. Recorded answers are static files.
Made to be talked through
One idea per scene, a short explanation and a “Try this” prompt for the class. Press P for present mode and use the arrow keys.