The most useful thing in this book is a piece of reframing that arrives early. Most public argument about AI risk splits into two camps that talk past each other, one worried about a superintelligence that decides against us and one insisting that the whole thing is overblown pattern matching. This book proposes a different question, which is not whether the technology will be dangerous but whether it can be contained. Containment means the ability to limit, direct or stop a technology once it exists, and the argument is that the properties of this particular wave make containment far harder than for anything comparable.
It is asymmetric, in that small actors can produce large effects. It is hyper-evolutionary, improving faster than governance can respond. It has general purpose applications, so you cannot restrict it to a domain. And it is increasingly autonomous.
That framework is genuinely useful and I have found myself using it since. The decision to treat AI and synthetic biology together is the book's most distinctive move and the strongest editorial judgement in it. These fields are usually discussed by separate people in separate venues, and the argument that they share the relevant structural properties, falling costs, spreading capability, dual use by default, is convincing. The biology material may be the most alarming part of the book for most readers, and it is not alarmist in tone.
It simply notes what the cost curves for synthesis and design have been doing and asks who will be able to do what, at what price, in a decade. The proliferation argument deserves attention because it is where a lot of policy thinking is weakest. Restricting a capability works when the capability requires resources only a few actors have. The concern here is that model weights get released, techniques get published, and the capability required to do something serious keeps falling, so that a control regime designed around a handful of large labs is aiming at the wrong target within a few years.
Whether or not you accept the conclusion, the mechanism is real and the book explains it clearly. The author's experience is an asset in the descriptive chapters. He has built and shipped these systems, run a large lab, and dealt with governments, and the passages about what actually happens inside these organisations and in those conversations have a specificity that outside commentators cannot manufacture. Now the difficulties, and the first is unavoidable.
This is a book about whether a technology can be contained, written by someone whose career and current position depend on continuing to build that technology at the frontier. That is not disqualifying and it is present on every page. It shows most clearly in what does not get examined. There is a great deal about state capacity, international coordination and the difficulty of regulation, and remarkably little about the concentration of capability and power among a small number of private companies, one of which the author leads.
For a book whose central concern is who gets to control a transformative technology, that omission is not small. The prescriptions are thin. After a long, detailed and often persuasive diagnosis, the proposals amount to a list of things that would help, better auditing, licensing regimes, international coordination, an aligned technical safety effort, a critical mass of public awareness. None of these are wrong.
None are developed to the point where you could act on them, and the difficult question of what to do when the actors who would need to coordinate have strong incentives not to is largely left alone. The book is at its most confident describing the problem and at its vaguest on the only part that would change anything. There is repetition. The central claim is restated at length across the middle chapters without being advanced much, and the book would be sharper at two thirds of its length.
And a smaller point that bothered me. The historical analogies to previous technological waves are used to establish that containment has never worked, and then the argument requires that this time we manage it anyway, without much explanation of what would make that possible. My three point four is for a genuinely useful reframing, for the correct and unusual decision to treat AI and biology as one problem, and for an insider account that outside writers cannot match. Marked down for a solutions section that does not match the quality of the diagnosis, for a conspicuous silence about the power of the companies at the centre of this, and for length that dilutes a good argument.
Read the first two thirds.