The role of AI
What you hand in
How we grade
Eight small experiments: you train the machine, take it apart, and plan how students should work with it. Everything runs in your browser: no account, no key, no data leaves your device.
Label messages and watch the machine learn by counting words.
Try it → 02How does a language model chop up text? One merge at a time.
Try it → 03The “Be the language model” game on a computer: probabilities, a wheel of fortune, temperature.
Try it → 04Train a real neural network on English first names and watch gibberish turn into names.
Try it → 05Draw circuit symbols, train an image recognizer, and see what it says about a scribble.
Try it → 06From noise to picture: watch a diffusion model “develop” an image step by step, and see what memorization means.
Try it → 07One task in three setups: watch messages travel between the model, the harness and the tools.
Try it → 08Turn the AI level dial and watch the task, the deliverables and the grading change. With a ready worksheet and declaration.
Try it →Each experiment takes 5–10 minutes. If you are curious how AI works, start with 01–03; if you are planning how students should work with it, 07 and 08 are for you.
The machine does not understand the messages. It counts how often each word appears in spam and in normal messages, and computes a probability from that. Change the labels, add a message, and watch the verdict.
Pink = pulls towards spam, blue = towards a normal message; the bigger, the stronger. Crossed out = unknown word.
Question for the group: write a spam message that gets past the filter! That is exactly what real spammers do.
Language models see pieces of words: tokens. The tokenizer learns from the training text which letter pairs are worth merging: always the most frequent one. Move the slider: fewer and fewer, bigger and bigger pieces.
The ▁ mark shows the start of a word. Real models have 50–200 thousand merges; a common English word is a single token, while a rarer word in a smaller language often takes 3–4.
This model counts what followed what in the training text. At every step it computes probabilities, then spins the wheel of fortune. The temperature decides how closely it sticks to the most likely continuation.
Experiment: look-back 4, temperature 0.1 → the model repeats the training text word for word. Temperature 2 → nonsense, but “creative”. Big models do the same, with far cleverer arithmetic.
A real, small neural network (about 9 thousand parameters; big models can have a trillion) learns from English first names: it predicts the next letter from the previous three. Start it and watch the error drop and the names become more and more name-like.
The error measures how surprised the model is by the real next letter. The dashed line is the level of random guessing.
Solid colour: a new name, not on the training list. Faded: it repeated a learned name.
What takes half a minute on a laptop here takes months on thousands of supercomputers for a big model. The principle is the same: reducing the error, step by step.
Three classes, a few sample drawings each, and a small convolutional network learns from them. Draw new examples, retrain, then test: draw a symbol on the pad and look at the guess.
Not trained yet.
This is what the model sees: your drawing shrunk to 28×28 pixels.
Hallucination in pictures: draw a house or a little sun. The model will still pick one of the three classes, often with high “confidence”, because those are all it knows. Language models also always answer something.
The model sits in the middle, the harness ring around it, the tool planets outside. You can run the same task in three setups and follow step by step where the messages go.
“Make a 10-question test from the Transformer chapter on Drive and email it to the head of department.”
The point: the model only suggests what should be done (“call the gmail.send tool”). The harness decides whether to execute it, and before important steps you have to approve.
Image-generating AI (diffusion models) starts from pure noise and guesses step by step what picture might be hiding underneath. Here a tiny 32×32-pixel toy model shows how. The same mathematics, just with 160 training images instead of billions.
During training we add a random amount of noise to every picture, and the model learns how to remove it. If it does this well, it can “develop” a picture even from pure noise.
In a real model, a text-understanding network turns the prompt into numbers, and these steer the denoising. Here the button simply chooses which family of pictures to guess from.
Few steps: fast but a rougher picture. Many steps: slower, finer. Real models take 20–50 steps, in a few seconds on a graphics card.
This toy model only knows 160 pictures, so it almost always returns one of its training images. Big models mix new pictures from billions, but sometimes they too return learned images: that is where copyright disputes come from.
The same learning goal, five different tasks. As you turn the AI level, the task changes too: AI is first a tutor, then a solver that has to be checked, and finally the subject of investigation itself. At every level the student is responsible for what they hand in and must be able to explain it.
The levels follow the internationally used AI Assessment Scale (Perkins et al., 2024), simplified. As the role of AI grows, the weight of the finished product in the grade goes down, and the weight of understanding and checking goes up.