It doesn't pick one answer. It rolls loaded dice.
After all that processing, the model produces a probability for every possible next token — a giant ranked list of "how likely is each word to come next." Then it picks one. Here's that list, live:
"The cat sat on the ___"
These are the model's guesses for the next word, with their probabilities. Drag the slider to change the temperature — the creativity dial — and watch the odds reshape.
Safe and predictable
Almost always picks the top word. Repetitive but reliable — good for facts, code, and structured answers.
Creative and surprising
Gives unlikely words a real chance. Varied, sometimes nonsense — good for brainstorming and stories.
One word at a time, in real time
This is the whole engine running. It picks a word, adds it, then predicts the next from scratch. The bars show its live guesses for the very next word.
It never drafts an outline first. Each word is chosen as the likely next one — yet whole coherent paragraphs emerge from that single move, repeated.
That's the engine. Next we'll watch all the pieces run together — and then spend the rest of the course on the big question: how did it ever learn which words are likely?