Large language models don't plan a whole sentence and then write it down. They generate text the way you'd guess the next word in a game — one piece at a time, by ranking every possible next "token" by probability and picking one, then doing the exact same thing again with the new, slightly longer text. Click through the widgets below to watch that process happen, one token at a time, in front of you.
Below is a partial sentence and the model's ranked guesses for what comes next, each with a probability. Click the button to let the model sample one token, append it to the sentence, and recompute a brand new probability distribution for whatever comes after that — conditioned on everything written so far.
Each bar is the model's estimated probability that token comes next, given everything before it. Watch how the ranking shifts completely once a new token is locked in.
"Temperature" doesn't add new candidates or remove any — it reshapes how sharply the model favors the top choice. Drag the slider and watch the same five candidates redistribute their probability in real time.
Five short questions. Pick an answer to see if you're right, then move to the next one.
Once you've seen next-token prediction in action, a lot of "AI behavior" stops looking mysterious and starts looking mechanical.
Once you understand that a model is sampling one token at a time from a probability distribution, a lot of other AI behavior — confident-sounding nonsense, creative variation, why the same prompt gives different answers twice — starts to make sense. A few related tools dig into those ideas further.
Want your team to actually understand what you're buying before you bet the roadmap on it?
Talk to us about your AI roadmap