How a language model writes

The weights are frozen and nothing is remembered. Pull the sliders and watch the model pick the next piece of text, over and over.

Temperature 1.00
Top-P 0.90

Tap a part of the diagram to open that step.

The whole loop

String → pieces → links to the earlier pieces → frozen layers → probabilities → cut the tail and draw one at random → the piece is appended. Then the same again from the top. Nothing else happens while a model is being used.

What's inside

  1. 1 Context and pieces — The window is finite. Once it fills up, generation stops — the model holds nothing outside it.
  2. 2 Attention over the context — The links are recomputed from scratch on every step: add one piece and the whole picture of links changes.
  3. 3 Frozen weights — Training changes the weights, using them does not. Two different modes of the same network.
  4. 4 Probabilities, temperature and Top-P — Move both sliders and watch " bananas": that is the price of a high temperature and a Top-P close to one.
  5. 5 The generation loop — The model does not write a sentence. It takes one step, then another, and that is all it ever does.
Read more On AI Dependency: Why the Older Brother Turned Out to Be the Younger One How honest dialogues with AI became a gift for corporations, and why the thinking should stay with the person while the routine goes to the model.

Explorable Explanations

Temperature 1.00
Top-P 0.90
1 / 5