Chapter Nine · Prediction, Run Forward
Generative AI
Every model so far has judged — a category, a number, a group. The systems everyone now talks to produce — sentences, images, working drafts. This chapter closes that gap with one deflationary secret: generation is prediction, pointed forward and run in a loop. Then it follows the idea all the way out — how these models are made, how to work with them honestly, what they generate beyond text, and where they are genuinely strong versus reliably shaky.
You already own every piece this chapter needs. Classification that outputs probabilities (Chapter 4). Labels the world writes for free (Chapter 2). Training as nudging numbers until predictions match examples (Chapter 3), and reward signals shaping behavior (Chapter 5). The three ingredients that made deep learning take off, and words placed in a space of meaning (Chapter 8). Generative AI — the chatbot on your phone, the tools that paint from a sentence — is those pieces assembled and scaled. No new magic arrives in this chapter, because none is needed.
Five topics carry it. First, the secret itself: a paragraph is grown one predicted word at a time, and the strip below is the whole trick. Then the three-stage recipe that turns raw autocomplete into an assistant. Then the two behaviors every user of these systems must metabolize — why what you type in changes so much, and why fluent falsehood is the native failure mode. Then the same idea painted onto images and sound. And finally the book's most practical page: an honest map of what these systems can and cannot do, derived from the mechanism rather than from vibes.