The mechanism is elegant: two networks trained against each other, one generating and one judging whether what it sees is real. The book walks the whole road, from autoencoders through conditional and domain-translation variants, and closes on adversarial examples and a chapter on ethics.
What survived
The generator-and-critic pair, which turned out to be a more general idea than the technique that made it famous. A generator with no critic converges on whatever is easiest to produce, not on what is best. That is exactly what happens to an AI system writing without anyone scoring the output: it produces plausible text and settles there. Which is why anything I deliver that generates content ships with its evaluator, even if that is a set of automated checks.
Mode collapse, the technical name for something I later saw in production: the generator finds one answer that fools the judge and repeats it with minor variations. Translated, an assistant that always says the same thing in different words because it found the formula nobody corrects.
And the chapter on adversarial examples, which aged in the opposite direction to the rest: it matters more now than when it was written.
What I left out
The technique, which is ninety per cent of the book. Training these networks was notoriously unstable, and that problem was not solved: it was sidestepped by changing approach, which is different. Anyone generating images today neither starts here nor passes through here. I say it flatly because a reading list usually hides this sort of thing: it is a well-written book about something that stopped being the way forward.