generative text model

A generative text model writes new text rather than just sorting existing text into bins. A spam filter is the opposite kind of model — it reads an email and outputs one label, spam or not spam. A generative model reads your prompt and produces fresh sentences that did not exist before, composed token by token to fit the request.

It generates by sampling: having predicted probabilities for the next token, it picks one, appends it, and feeds the longer text back in to predict the next. Because there is a choice at every step, the same prompt can yield different answers. This is what makes the output feel creative, and also why it is not perfectly repeatable.