Huyen starts from a distinction that gets lost in sales conversations: training a model is one thing, building a product on top of a model that already exists is another. The second is almost everything a mid-sized company does today, and it has problems of its own: picking the model, measuring whether it helps, adapting it, controlling cost per answer, and closing the loop with what users say.
What I did with it
The order of adaptation is what I use most. Cheap and reversible first — change what you ask the model — then bring the company's own documents into the context, and only at the end adjust the model itself. Done backwards you spend a lot and learn little, and fine-tuning before knowing what is failing is the most expensive way to discover that the problem was the data.
The evaluation chapters left me a working rule: do not start building until there is a way to tell whether it came out right. That means a set of cases with their expected answers, written with whoever knows the business, before the first line of code. It costs a day and saves the month in which nobody can agree on whether anything improved.
And a third one, architectural: the model is an interchangeable part, not the centre. If the application is written so that switching provider is a decision measured in days, you negotiate with today's prices rather than the ones you signed for a year ago.
What I left out
The inference optimisation section, which is excellent and is not my problem: if you pay per call, you are not fighting for room on a graphics card. And the specific model and price tables, which aged while the book was being printed. What does not age is the method, and that is half the book.