The Generative Machine

Preface There is a particular kind of book that a fast-moving technical field eventually demands. Not a textbook, because a textbook must settle on a pedagogy and a pedagogy requires the field to hold still. Not a survey paper, because a survey paper is bound by the conventions of its venue and the patience of its reviewers. Not a popular account, because a popular account must spend its budget on metaphor. What the field demands is a compendium: a dense, organised, cross-referenced record of what is known, what is built, what it costs, what it scores, what it breaks, and who says so. That is what this book attempts for generative artificial intelligence as the field stands in the middle of 2026. The ambition is coverage. A reader should be able to come to this volume with a question about the Kullback-Leibler term in a variational bound, or about the memory bandwidth of an accelerator, or about which post-training algorithm replaced Proximal Policy Optimisation in reasoning models, or about what the European Union requires of a general-purpose model provider from August 2025 onward, and find a serious answer in one place. The topics are not usually collected together. The mathematics of score matching and the capital expenditure plans of hyperscale cloud providers do not appear in the same documents, and yet anyone working seriously in this field needs both, because the second determines what the first can be run at.