Interactive work / A Litil project

How LLMs Learn

Watch a tiny language model make a guess, get it wrong, and change. Then take over and train it yourself.

Start the explainer Back to work

A 70-second film / then a live model in your browser

The whole brain

One small model. Nothing hidden.

The explainer runs a real next-word model in your browser. Its forward pass, loss, backward pass, and training steps are computed live—not animated placeholders.

251
numbers in the model
8 → 9 → 9 → 8
network shape
6
context words to teach
0.01
final loss in the film

The experience

Watch. Then take over.

  1. 01

    Watch the guess.

    A word enters the model. Light moves through every layer. Eight possible next words come out.

  2. 02

    See the learning step.

    The model measures how wrong the guess was, then adjusts every number a little toward a better guess.

  3. 03

    Train it yourself.

    Take over a fresh model, teach all six context words, and watch the loss fall as the guesses improve.

Your turn

Training is the same step, repeated.

Make a guess. Measure how wrong it was. Adjust every number a little. Repeat.

Teach the model