Part 0: Orientation

"We can only see a short distance ahead, but we can see plenty there that needs to be done."

Alan Turing, "Computing Machinery and Intelligence"

Most readers do not enter AI infrastructure from the bottom of the stack. They enter somewhere in the middle: a model release, a benchmark jump, a GPU bill, a latency incident, or an agent demo that worked yesterday and fails today. Each problem looks local at first. The explanation rarely is. A serving technique can reach back into architecture. What post-training learns to reward is set by a benchmark. Whether an agent is allowed to attempt something comes down to a safety rule, and a modeling idea exists at all only because a hardware limit, several layers down, left room for it.

Part 0 makes the path visible before later chapters put many mechanisms in play at once. Chapter 1 takes a plain task, an assistant asked to find a bug in a small repository, propose a fix, and explain it, and watches that request cross data, tokenization (text cut into integer pieces), base-model formation (training the raw model), adaptation (tuning the trained model's behavior), serving, retrieval, and an agent runtime. The point is not to learn those layers in depth yet. It is to see that later chapters are not separate topics. They are different costs and choices on the same path.

Once that path is visible, Chapter 2 pulls back to the whole book. It gives the stack a stable shape and marks the difference between ground solid enough to learn as engineering practice and ground still under argument: reasoning gains, interpretability, agent architecture, evaluation. The constraint arrow, one of the book's recurring habits, makes its first appearance here: when a design choice on the page seems arbitrary, look one layer down and ask who is paying the cost.

The last piece of orientation is about language. Chapter 3 separates borrowed mechanisms from borrowed pictures. Compression as prediction is not a metaphor; it is the training objective read through information theory. Diffusion's physical origin was real mathematical scaffolding, even if the modern practice no longer talks much like thermodynamics. Neuron, attention, memory, and System 2 are weaker words. They can guide intuition, but they should not be mistaken for the thing itself.

Orientation closes by drawing the subject's boundary. Chapter 4 puts the generative stack next to the AI infrastructure that already ran the world's feeds, search rankings, and ad auctions, and asks what the new stack inherited from that incumbent, what it could not reuse, and where the two are now converging. Knowing where the book's subject ends keeps the later parts from quietly claiming more than they cover.

After Part 0, the book can move faster. The route through the stack is visible, the live claims are separated from settled ground, borrowed vocabulary has been put in its place, and the boundary with the previous AI era is drawn. That is enough orientation to start taking the machine apart.

How to read this book

The book is built to be entered in the middle, at whatever problem brought you here, not read front to back. Three habits make that work. A professional term is set in its full form with the English original on first use in a chapter, and every such term links to the glossary, where a one-line definition and a pointer to where it is first explained are waiting; when a word is unfamiliar, follow the link before reading on. A cross-reference like Chapter 31 points to where a mechanism is developed in full, so a passing mention need not carry the whole weight. And the constraint arrow, introduced in Chapter 2, marks the moments where a choice on the page is dictated by a cost one layer down.

One prerequisite the book does assume rather than teach from zero: how a transformer turns tokens into predictions. A reader who has not met the architecture should read Chapter 7 and Chapter 8 before the serving, reasoning, and orchestration parts lean on them, since those parts treat attention, the key-value cache, and the prefill and decode phases as known ground. The transformer chapter builds each of those from the residual stream up.