The type of data used and the generative AI outcome varies depending on the type of data being analyzed. While the focus of this project is on LLMs, other types of generative AI models can produce other types of output.
- Text—LLMs can be used to generate new text based on a specific prompt or to compile long sections of text into shorter summaries. For example, ChatGPT can generate news articles or product descriptions from a few key details.
- Image—Generative AI models for images can be used to create realistic images of people, objects, or environments that do not exist. For example, StyleGAN2 can generate realistic portraits of nonexistent people.
- Audio—Generative AI models for audio can be used to generate new sounds or music based on existing audio samples or to create realistic voice simulations. For example, Tacotron 2 can generate speech that sounds like a specific person, even if that person never spoke the words.
- Video—Generative AI models for video can be used to create videos based on existing footage or to generate realistic animations of people or objects. For example, DALL-E can generate images of objects that do not exist, and these images can be combined to create animated videos.
In each case, the generative AI model must be trained on large datasets of the appropriate datatype. The training process is tailored to the requirements of the datatype and to the specific datatype because different input and output formats are required for each type of data. Recent advancements are now capable of integrating differing datatypes, for example, using a text entry to generate an image.