Chip Huyen's AI Engineering arrived at just the right moment, because the last couple of years produced a huge wave of people whose job is no longer to train models from scratch but to build reliable applications on top of foundation models they call through an API, and until recently there was no single resource that laid out what that job actually involves. This book is the best attempt I have seen at doing exactly that, and it follows on naturally from her earlier Designing Machine Learning Systems, which was itself widely respected. The framing is the key thing to understand before you buy it. This is not a tutorial where you type along and end up with a finished app, it is a book about how to think, walking through the real decisions an AI engineer faces and the trade offs behind each one.
It covers how to evaluate models and, crucially, how to evaluate your own system so you can tell whether it is genuinely working, which is a subject most courses and blog posts skate over despite it being where most real projects quietly fail. It goes through prompt engineering with a clear head rather than as folklore, retrieval augmented generation, when and how finetuning is actually worth it, inference optimization and the latency and cost realities of running these systems, and the architecture of AI applications including agents. The strength that runs through all of it is judgement. Huyen writes as someone who has shipped this stuff, so instead of cheerleading for whatever framework is trending, she explains why you might choose one approach over another and what it will cost you, and that kind of reasoning ages far better than any specific library walkthrough.
The limitations are the natural flip side of what makes it good. Because it prioritises concepts and decisions over code, you have to do the building yourself and bring your own hands on practice, so it works best read alongside a real project rather than in a vacuum. It also assumes a baseline of software and machine learning literacy, so someone who has never written code or trained a model will find stretches of it heavy. And in a field moving this fast, a handful of specific tool references will inevitably date, though the underlying principles are chosen well enough that the core will stay relevant far longer than the average tutorial.
My recommendation is strong and simple. If you are building or about to build applications on foundation models, from either direction, coming in as a software engineer or coming from traditional machine learning, this book is one of the highest value things you can read at the moment. Go in expecting a mental model of the entire field and the judgement to navigate it rather than a copy paste build along, read it at a deliberate pace with a real project open beside you, and it will sharpen how you work well beyond its cover price.