Gary Marcus has spent years arguing that deep learning is overrated and that the field is overselling its progress, a position that has made him a prominent critic and a frequent target. This book takes that scepticism into policy territory, arguing that AI development is currently shaped by commercial incentives that do not align with public interest, and proposing specific regulatory responses. It is short, it is direct, and it is angry in a way that will either sharpen your attention or lose you entirely. The strongest material is on incentives rather than technology.
The argument that companies racing for market position will systematically underinvest in safety, overstate capability, resist oversight and externalise harms is not controversial in other industries and is well made here. The examples of misleading demonstrations, quietly retracted claims and the gap between marketed and actual capability are specific and documented, and that section is genuinely useful regardless of your view on the broader thesis. The policy proposals are more concrete than most books of this type manage. Rather than calling vaguely for oversight, there are specific suggestions about disclosure requirements, liability, data rights and independent auditing.
Whether you agree with them or not, having something specific to argue with is a considerable improvement on the usual hand waving, and it makes the book more useful for anyone actually thinking about governance. The technical criticism is where the author's history intrudes. Marcus has argued for a long time that scaling would not deliver reasoning, and that prior commitment shapes how capability is assessed here. Some of the specific limitations described have been substantially reduced by subsequent models, which is the occupational hazard of writing about a fast moving field and also a pattern in this particular author's work.
The structural critique of incentives ages far better than the capability claims, and a reader should weight those sections differently. Tone is a real barrier. The book is written with visible frustration, and while that frustration is often earned, it makes the argument easier to dismiss for anyone not already inclined to agree. A cooler treatment of the same material would persuade more people, and the author appears uninterested in persuading the unpersuaded.
Whether that is admirable or self defeating depends on what you think the book is for. Implementation difficulty gets less attention than it should. Regulating a fast moving technology across jurisdictions, defining terms precisely enough to be enforceable, and avoiding rules that entrench incumbents are genuinely hard problems, and the proposals here are stronger on what should happen than on how. That is a common weakness in policy books and it does limit their practical value.
The length is a real virtue and worth saying so. You can read this in an evening, which means it can function as a counterweight in a reading diet rather than a commitment. I would recommend exactly that use, placed against a more optimistic book, rather than treating it as the final word. Three point seven.
A sharp, efficient critique of the commercial forces shaping AI, with concrete policy thinking and a tone that limits its reach. Best read as one voice in an argument rather than a settled account, and the incentives material is worth the time even if the technical scepticism does not land for you.