Generative AI
In brief
Generative AI refers to models that can produce new content (text, images, sound, video, code) from a request, based on what they have learned from very large numbers of examples.
Definition
A generative AI creates content that did not exist before: a text, an image, a piece of music, a video, some code. It does not copy a stored document: it produces a new result that is statistically consistent with the examples seen during training and with the request (the prompt) it is given.
The main families of models
- Language models (LLMs): they generate text token by token (a token is a word or part of a word). They are based on the Transformer architecture (2017). ChatGPT, Claude, Gemini and Le Chat are examples.
- Diffusion models: they start from random noise and "denoise" it step by step until an image or a video emerges. This is how many current image generators work.
- Generative adversarial networks (GANs): two networks train against each other, one to produce, the other to spot fakes. Widely used before diffusion took over.
Benefits
- Write, summarise, translate or rephrase in seconds.
- Produce visuals, mockups or prototypes without design skills.
- Help with programming, data analysis or exploring ideas.
Limitations
- Convincing errors: generated content can be wrong while looking credible (see hallucination).
- Rights and data origin: copyright questions about training data and generated content are still debated.
- Bias: models reproduce the biases present in their data.
- Disinformation: how easy it is to produce fake texts, images or voices makes it essential to check where content comes from.
Spotted a mistake or have an idea about this entry? Let us know