noun · AI process
Next-token prediction
also called autoregression
Autocomplete on a huge scale. The AI picks one small chunk to write next, adds it, then looks at everything again and picks the next.
The core loop of a language model: given the tokens so far, produce a probability for each possible next token, pick one, append it and repeat.
At each step the model outputs a score for every token in its vocabulary. A sampling rule picks one, it is appended to the input, and the loop runs again until a stop token or a length limit.
Real-life example
You start a sentence in your phone's autocomplete and keep tapping the middle suggestion. After ten taps you have a sentence nobody planned. A language model is the same idea at a far larger scale, and with a much better sense of what fits. It writes the answer one piece at a time and re-reads everything before each piece.
Where the analogy breaks
-
There is no finished answer waiting inside. The reply is built one token at a time, so earlier choices shape later ones, including mistakes.
-
Most likely is not the same as true. The model scores what is plausible after the text so far, which is why a fluent sentence can still be wrong.
-
Every extra token needs another pass. That is why long replies take longer and cost more than short ones.