It is three books in one: a refresher on linear algebra, probability and numerical optimisation; a manual of deep networks; and a research section that was the frontier in 2016. It is written for people who train models from scratch, which is neither my work nor what suits most mid-sized companies.
What I took from it
The chapter on practical methodology, one of the shortest and least quoted. Its recipe: fix the goal and the error metric before choosing a model, build the simplest thing that works end to end, and only then instrument the system to find out whether the bottleneck sits in the data, the model or the tuning. Translated into my work: deciding how the word "works" will be measured, before promising anything, is half the engagement.
The relationship between capacity and generalisation, which is the technical explanation of something I see constantly. A model with capacity to spare memorises the examples it saw and falls apart just outside them. That is precisely what happens to a demo built on the three documents the client picked: it impresses, and it says nothing about the other ten thousand.
And one line I use to take the heat out of a meeting: underneath there is no magic, there is linear algebra, probability and a numerical method walking downhill. Being able to say that calmly changes the tone with a management team that arrives either frightened or dazzled.
What I left out
Almost everything else. The third part is decade-old research, and the book is from 2016: the architecture everything now called AI rests on was published the following year. Anyone buying it to understand today's language models is in for a thousand-page disappointment. You read it for the foundations, which do not expire, not for the state of the art, which did.