Generative AI
Generative AI
Models that produce new artefacts rather than predicting a label. The densest region of the Atlas.
Grasp
A discriminative model answers "which category is this?". A generative model answers "what does a plausible example look like?". That shift, from drawing boundaries between things to modelling the distribution of things, is what made the last few years feel discontinuous.
Two families dominate practice. Autoregressive transformers generate sequences one token at a time, each conditioned on everything before it, and they power every large language model you have used. Diffusion models start from noise and iteratively denoise toward a sample, and they power most image and video generation.
The engineering consequences are unlike anything in classical ML. Output is non-deterministic, correctness is often a matter of judgement rather than a label, and evaluation is genuinely unsolved. Most of the difficulty in shipping generative systems lives in those three facts, not in the models themselves.