Foster goes family by family: variational autoencoders, adversarial networks, autoregressive models, normalising flows, energy-based models and diffusion, then the jump to transformers and applications. Every chapter carries a small model that trains on a normal machine, and that decision about scale is what makes the book teach something.
What I took from it
Understanding that generating means sampling from a distribution changed how I answer the most repeated question in a management meeting: why did it say one thing yesterday and another today. It is not an error and not a configuration problem, it is the design. From which follows a practical decision I use constantly: where the business needs the same answer every time, you do not put a generator there, you put a rule; and where variety is the point, a generator is the right tool.
The comparison across families, which explains why each kind of content ended up with a different method, and helps avoid importing one domain's fashion into another where it does not pay.
And the value of building small. Watching the mechanism work in a model that fits on a laptop is the fastest way to lose your reverence for the technology, which is a precondition for being able to decide about it.
What I left out
The chapters on music and world models: entertaining, and a long way from anything I get asked for. And a limitation of the whole genre: the code in a book like this ages at the pace of the libraries it uses, so two years on, half the examples no longer run. The explanations, which are what you came for, do not age.