It is a wide book: model architecture, fine-tuning, retrieval, agents, reinforcement learning, infrastructure, governance and compliance, time series, and a closing chapter of predictions. That breadth is its flaw — each topic gets whatever room is left.
What I took from it
First, the idea that reasoning may be probabilistic but execution has to be deterministic. A system can let the model decide what ought to happen, and still the action that touches the real world — issuing, paying, posting — goes through a closed list of permitted operations rather than through free text. That is a boundary, and drawing it is an architectural decision, not a coding one.
Second, the policy gate: before a tool runs, a rule looks at the proposed operation and the current state and decides allow, deny, or escalate to a person. It is simpler than most of what gets published about agents, and it is what makes the whole arrangement defensible in front of an auditor.
Third, and most useful to me: measure faithfulness per claim, not per answer. A long answer may contain five claims of which four are supported by the documents and one is not; scoring the whole answer right or wrong hides exactly the one that matters. Since then, when I define how a question-answering system over a company's own documents will be evaluated, the unit of measurement is the claim, with its supporting evidence beside it.
What I left out
The financial context, which is most of the book: earnings calls, models trained on market data, price impact. None of it transfers to a mid-sized company. I also left the reinforcement learning and time series chapters, which in a book this broad work as summaries of other books.