Generative AI is a family of systems able to produce original content — text, images, audio, code, video — rather than simply classifying or predicting. The clearest analogy: if classic AI is the grader sorting essays, generative AI is the writer composing a new essay, word after word.
Learning the shape of the world
These models learn the statistical distribution of their training data, then sample from it to create something new. A language model predicts the next token (word fragment) from the preceding ones. Formally, it models the probability of a sequence:
$$P(x_1, \dots, x_n) = \prod_{t=1}^{n} P(x_t \mid x_1, \dots, x_{t-1})$$
By repeating this prediction, it generates a whole sentence — coherent because it was learned from vast corpora.
The main families
| Family | Principle | Typical use |
|---|---|---|
| Transformers (LLMs) | Attention over sequences | Text, code |
| Diffusion | Denoise a noisy image | Images, video |
| GANs | Generator vs. discriminator | Faces, styles |
| VAEs | Encode/decode a latent space | Compression, variations |
Strengths and limits
- Strength: creativity at scale, rapid prototyping, writing assistance.
- Limit: hallucinations (invented facts stated as true), data bias, energy cost.
For the saMORAIs, the key point is that these systems do not "understand" in the human sense: they recombine learned regularities.
Generative AI does not recite the past — it draws unprecedented combinations from it, which is the source of both its power and its risks.