Other AI models
Other AI models
This module will grow. Each part stands alone, and each names products and figures that date quickly. It covers kinds of model that the other modules do not: small models such as the Smol, Phi and Needle families, world models such as NVIDIA’s Cosmos, DeepMind’s Genie and World Labs’ work, language models built on diffusion, and models such as Jev that return a decision and no text. Part 4 describes a product announced days before it was written, from its maker’s own figures, and says so. The whole module is about two hours of reading.
Part 1: Small models
After it you can: Explain how quantising, distilling and pruning make a model small, what each costs, and what a very small model is good for
30 minutes
Part 2: World models
After it you can: Say what a world model predicts, how it differs from a video model, and what it is for
25 minutes
Part 3: Diffusion language models
After it you can: Explain how a language model can write a whole passage at once and refine it, and what that trades against writing word by word
25 minutes
Part 4: System one models
After it you can: Tell when a decision needs a typed answer with a probability, and not generated text
30 minutes