Summary: The four-paradigm landscape and where modern systems sit
The track opened with a map. Thirteen lessons later, you have built every paradigm from its foundations. This lesson is what closes the map.
What this lesson did
Section titled “What this lesson did”- Recapped the four paradigms (autoregressive, latent-variable, adversarial, score-based / diffusion) with each one’s training objective, sampling procedure, and trade-off profile filled in from the intervening derivations.
- Placed modern systems on the map explicitly: a modern autoregressive language model on paradigm 1, a Stable-Diffusion-style latent diffusion image generator as a paradigm-2-plus-paradigm-4 hybrid, a GAN-based face generator on paradigm 3, a modern text-to-video diffusion system on paradigm 4 extended to video, a multimodal language-plus-image system as a paradigm-1-plus-paradigm-4 hybrid.
- Named what the four paradigms share: a trained network, an information-theoretic objective tied to the data distribution, a paradigm-specific sampling procedure, and a trade-off set that cannot be jointly optimized.
- Built the paradigm fluency procedure: identify the training objective, identify the sampling procedure, predict the trade-offs, place the system on the four-paradigm map. The procedure is the deliverable of the track.
What to remember in three lines
Section titled “What to remember in three lines”- Four paradigms of generative modeling: autoregressive, latent-variable, adversarial, and score-based / diffusion. Each picks a training objective, a sampling procedure, and a trade-off profile. Modern systems often combine paradigms; identifying components is the first step in reading any release.
- Paradigm fluency is the deliverable. Identify the training objective, identify the sampling procedure, predict the trade-offs, place on the map. This is the procedure for reading any new generative-AI system release critically; the math you built through the track is what makes it precise.
- Hybrids are common. Latent diffusion uses a VAE for compression and a diffusion model in the latent space. Multimodal systems pipe autoregressive language through diffusion image generation. Recognizing the components and how they fit on the map is what makes hybrid systems readable.
Where this is going
Section titled “Where this is going”The track ends here. After this lesson, every model you can read about across the field has a precise home in one of four paradigms (or a named combination of paradigms). The map is the spine; the math is what makes the spine precise. Generative models are not magic; they are math, and the math is the same math you already have.