What the book underweights
A book with one argument to make has to leave things out, and the useful question is which omissions change what a reader should do. These are the five that do.
Partial measures as a category
The book treats every measure short of a halt as failing to solve the problem and therefore as worthless. The first half is true. The second does not follow, and the argument page explains why: the value of a measure that slows the race or improves visibility depends on whether alignment failures scale smoothly, and the book settles that by assumption. If the halt does not come, and every reviewer thinks it will not come soon, then the partial measures are the entire policy landscape, and dismissing them means having nothing to say about the world we are actually in. The work the book leaves undone is to say which partial measures buy time and which merely buy cover, and that distinction is where an evaluator could contribute most.
Interpretability
The book's mechanism is that grown systems cannot be understood, so their goals cannot be known. Interpretability is the research programme that attacks exactly that, and it has produced real results: reproducible structure inside models, features corresponding to concepts, circuits implementing recognisable computations. None of it is a control theory and the authors are right that it is nowhere near what would be needed before scaling to something dangerous. But it is the only line of work that could make claim 2 false rather than merely delayed, and it gets two paragraphs. A reader who believes the halt will not happen should regard this as the most important thing in the field, and the book gives them no way to know that.
The labs as sources, not only obstacles
Workable Plans Will Involve Telling AI Companies "No" is right that a plan requiring a lab's consent is not a plan. It is also true that almost every piece of published evidence the book relies on for its central mechanism came out of the labs: the alignment-faking results, the strategic behaviour under test, the capability jumps nobody predicted. The labs are where the talent, the compute, the interpretability work and the evidence are. A strategy that treats them purely as adversaries forgoes what they know, and the book's advice to policymakers, to give lab heads no sway, conflates the people who should not decide with the institutions that have the information. Those can be separated. The book does not try.
The non-extinction outcomes
The book's structure allows two worlds: the halt, and everyone dies. The Survive section exists because there are several others and they are where a person's choices bear on their own outcome. Gradual disempowerment, misuse, concentration and dislocation are not consolation prizes; they are the worlds a reader most likely lives in for a while, whatever the eventual outcome, and the book gives them almost nothing. This is the omission that most changes what a reader should do, because it is the difference between a book that tells you to prevent something and one that tells you how to live.
The politics of the treaty
The treaty draft is a serious piece of work, more serious than the reviews acknowledge. What it lacks is a path: which governments first, in what sequence, driven by what, against whose opposition, with what to offer the defectors. Every sympathetic reviewer notices this and the book's answer is a march. The demand side has in fact moved, which the ranking page discusses, and that is evidence the authors' theory of change is not empty. But "build public demand and the treaty follows" is a hope about the second step, not an account of it. The work left undone is a theory of how demand becomes a signed instrument, and it is the work that, on the book's own logic, everything depends on.
What this adds up to
The book is strongest where it argues mechanism and weakest where it argues strategy, and the five gaps above are all strategic. Filling them does not require disagreeing with the book about the risk. It requires taking the risk seriously enough to plan for the case where the book's one remedy does not arrive in time, which is the case every reviewer expects. That is the direction this site takes: not a rebuttal, and not an endorsement, but the part of the work the authors chose not to do.