Output

The output is the result a function computes from an input.

For a language model, the immediate output isn’t a finished word, let alone a finished sentence — it’s a list of scores, a numeric value for every single piece of text in the entire vocabulary, including ones that fit badly. A following selection step uses those scores to choose one concrete piece, either always taking the highest score or allowing some controlled randomness.

An example: A spam filter adds up the weights of the words in an email. Its output is first a number, say 5. Only the comparison with a fixed threshold turns that into the verdict “spam.” Image recognition, by contrast, outputs a whole row of numbers, one per class such as cat, dog or car. That whole row together is the one output.

Not to be confused with the finished answer: What you read in the chat window is the result of many rounds. In each round, the model outputs a score list, one piece is chosen and appended, and the longer text goes back in. The list itself always has the same length: for the older language model GPT-2, it has 50,257 entries, no matter how short or long the input is.

Where you’ll come across it: When a chatbot builds its answer almost word by word, you are watching this loop. The word also shows up in the settings of AI services, for example as a length limit on the output.

Explained in more depth in Input and Output: What a Function Does.