Nine hands-on experiments take you from raw internet text to a thinking agent. Tokenize a sentence, explore meaning, play the sampler, train a model, run an agent. Built to be used, not just watched.
Modern AI learns from datasets like FineWeb — trillions of words scraped from the web. But raw web text is full of spam, duplicates, and junk. It must be cleaned first.
A model cannot read letters. Text is split into tokens — chunks of characters, each mapped to a number — turning language into a sequence of integers.
Each token becomes a vector placing it in space. Words with similar meaning land near each other — the model discovers this just by reading.
Inside a transformer, attention lets every word look at every other word and decide which ones matter. The word "it" learns to look back at what "it" refers to.
A model only ever predicts the next token. It assigns a probability to every word, then samples one. You do it now — build a sentence by picking the next word, one token at a time.
Each wrong guess produces an error that nudges the model's parameters. Repeat trillions of times and the loss falls — the model learns. Run it yourself.
With instruction tuning, RLHF, and tools, a base model becomes an agent: it reasons, takes an action, observes the result, and loops until done.
A model optimises for plausible, not true — it can hallucinate. We build systems around it to catch and fix errors before they reach you. See it in action.
Every AI travelled this river. Click any stage to revisit that experiment.