A tiny model is learning to write, live.
It starts with random settings, so its first writing is random letters. Then it practices: it guesses the next letter in a piece of text, checks how wrong it was, and adjusts itself a little. Repeat that thousands of times and the guesses become words, then phrases.
Nothing is fake or pre-written.
The page contains a small language model and a hand-written training engine. Every adjustment is computed here in your browser from real errors. The “math check” at startup compares the engine’s gradients against tiny numerical nudges and only starts if they agree.
What the panels mean.
How wrong its guesses are: a score of its mistakes. Lower is better. The dashed line is text it has not practiced on, so it shows real learning rather than memorising.
Which earlier letters it looks at: when choosing the next letter, the model can look back at earlier letters. Brighter squares mean it is paying more attention to that earlier letter.
Inside its tiny memory: a map of its letter ideas and small pictures of the numbers it is changing. They move as it learns.
Why it starts by itself.
This page is built to be watched silently. It autoplays a short run, shows the step-0 writing beside the latest writing, and captions the stages in plain English. Open Controls any time to pause, reset, or experiment.