The Lab
Play with how AI works.
Reading about AI is one thing. Poking it is another. These toys run entirely in your browser — nothing is sent anywhere. Break them, replay them, and the ideas from the line will click.
AI doesn't read letters the way you do. It chops text into tokens — whole words or fragments — and works with those. Type anything and watch it split.
A token is roughly ¾ of a word on average. This is a simplified splitter — real models use a learned vocabulary — but the idea is identical: text becomes tokens before anything else happens.
A language model writes by guessing the next word, over and over. Start a sentence, then keep picking from what it thinks comes next — you're doing exactly what a model does.
Notice you're never certain — you pick from likely options. That uncertainty is why the same prompt can give different answers, and why a model can confidently pick a wrong-but-likely word. That's a hallucination.
An AI learns from examples. If the examples lean one way, so does the AI. Drag the slider to change what the training data looks like, and watch the model's "belief" follow it.
All group BBalancedAll group A
The training data
What the model now "believes"
Nobody told the model to be unfair. It just reflected its data. That's why the data you train on — and who's missing from it — is a safety issue, not just a technical one.