Huyen writes from the position of someone who had to keep these systems running, not from the position of someone publishing results. That is why the book spends more room on data, deployment, monitoring and organisation than on the model, which takes up a chapter and a half out of eleven.
What I took from it
Data distribution shift, which is the polite way of saying the world moves and the system stands still. A model does not break with a visible error: it drifts quietly until it is answering questions nobody asks any more. The criterion it left me is a question I ask before promising anything that learns: who will notice it has degraded, from which signal, and how long will that take. With no answer, the system is not finished however well it performs today.
The choice between answering in batches and answering on request, framed as what it is: a business decision with cost consequences, not a technical preference.
And the chapter on the human side, which almost no book in this field includes. These systems do not fail alone. They fail inside organisations where nobody is clearly accountable when an output looks odd.
What I left out
It is from 2022 and it shows. The chapters on training data and feature engineering assume the company trains its own model, which was normal then and is the exception now. Anyone reading it to learn how to build on top of a language model will find half the book out of date. The other half, the operational half, is still among the best available.