15.22 Where we are
A reader who started this book at chapter 1 has travelled from the perceptron to AlphaFold 3 in twenty chapters. The final stretch, the past five years, is where almost all the visible capabilities live. It is also the part of the field least settled and most rapidly changing. We have tried to write this chapter in equations and design principles rather than model names so that it ages gracefully, but no chapter on modern AI in April 2026 will look the same in April 2027.
Three observations to close.
First, the recipe is simple. Pre-train on a lot of diverse data. Fine-tune on instructions. Apply preference learning. Add tools, retrieval, and inference-time search as needed. The conceptual machinery in this chapter is small. What is large is the engineering, the data, and the compute.
Second, the frontier is a moving target. Every section in this chapter has a 2024 version that is already wrong in 2026, and a 2026 version that will be wrong in 2028. The mathematical foundations, the scaling laws, the Bradley–Terry model, the closed-form RLHF optimum, the GRPO objective, are durable. The system names are not.
Third, the relationship between this chapter and the rest of the book is the relationship between an industry and the science it grew from. Modern AI is not a separate discipline; it is the dense application of every chapter that came before it, plus engineering at unprecedented scale. The reader who internalises chapters 1–14 has the tools to read any frontier paper that comes out next year. That is the goal of this book.