Classifiers are just one example of output, one where the final layer of output is interpreted to be probabilities of different items. Using a different loss fu

Classifiers are just one example of output, one where the final layer of output is interpreted to be probabilities of different items. Using a different loss function (or reward function), you can train for the output to be different things. But in fact text generators work basically identically to classifiers. You train the model to classify texts according to which single word comes next. Then you append a word to the text according to that output, and repeat.