Generative AI: construct an output from a learned distribution

Generative AI produces content such as text, images, audio or code using learned structure. The output is constructed through a model rather than merely selected as one fixed class label.

Different generative models organize that construction differently. An autoregressive language model chooses one token after another. A diffusion model repeatedly transforms a noisy state toward an output.

Conditioning directs the generation

A text prompt, an image, an audio segment or a partial output can constrain what is generated. The same base model can therefore support several tasks without being retrained for each individual request.

Generation is not the same operation as retrieval. A system can first retrieve evidence and then generate an answer from it, combining the two mechanisms.

A newly produced sequence is not necessarily a novel idea or a true statement. Models can reproduce familiar patterns and sometimes memorized material. Judge the result by the task: a useful image, a supported explanation and executable code require different kinds of checking.