Books that make confident predictions about technology are usually either forgotten or vindicated, and the interesting ones are the ones that turn out to be partly both. Rebooting AI argued in 2019 that deep learning was fundamentally limited, that scaling would not produce systems with genuine understanding, and that progress required reintroducing symbolic reasoning. Reading it now, with several years of evidence the authors did not have, is a genuinely useful exercise and a more interesting one than reading a book that simply turned out to be right. Start with what holds up, because a fair amount does.
The specific criticisms of brittleness were accurate. Systems that perform impressively on their training distribution and fail bizarrely just outside it. Models that appear to reason and are in fact pattern matching in ways that break under mild perturbation. Failures of common sense physical and social reasoning that no human would make.
These were real problems then and versions of them persist now. Anyone who has watched a capable model fail at something trivially easy has seen exactly what this book was pointing at. The argument that benchmark performance was being over-interpreted was also correct and remains worth hearing. A model scoring well on a test tells you it does well on that test, and the leap from there to claims about understanding was being made carelessly then and is still being made carelessly now.
The book's insistence on that distinction was a service. The writing is clear and the book is accessible. Davis and Marcus explain technical ideas to a general audience competently, and the examples of failure are well chosen and memorable. Someone with no background can follow the argument fully.
Now the central problem, which is substantial. The book's core bet was that scaling would hit a wall and that qualitative advances would require architectural change of the kind the authors advocated. Scaling went considerably further than the book anticipated, and a great many capabilities that are presented here as evidence of deep limitations turned out to be things that emerged with more compute and more data. That is not a small miss.
It is the load bearing claim of the book, and getting it wrong changes how much weight the rest can carry. The tone does the argument real damage. The critique is often combative in a way the substance did not require, and it reads at times as though winning an argument matters as much as being right. That is a shame, because the specific technical criticisms deserved to be heard on their merits, and a more measured presentation would have aged better.
Critics of a dominant paradigm need to be unusually careful with tone precisely because they will be dismissed as contrarian, and this book made that dismissal easier than it needed to be. The proposed alternative is much weaker than the critique. The book argues for hybrid systems combining neural and symbolic approaches, and the case for why such systems should work is more convincing than any account of how to actually build them. This is a long standing problem in this line of criticism.
The gap in current approaches is identified sharply and the replacement remains a sketch. Reading it now, the constructive half feels considerably thinner than the destructive half. Some examples have simply expired. Failure cases presented as illustrating fundamental limits are things current systems handle without difficulty, and each time you hit one it undermines the argument around it, sometimes unfairly, because the underlying point occasionally survives the specific example.
Distinguishing the two requires knowing the current state of the field, which means this book is more useful to someone already informed than to a newcomer who cannot tell which criticisms still bite. What makes it worth three point four rather than lower is the value of reading confident technical prediction against known outcomes. Very few people in 2019 correctly called what would happen next, and watching intelligent, well informed people reason carefully to a conclusion that turned out substantially wrong is genuinely educational about how hard this kind of forecasting is. It should make you more sceptical of confident predictions in either direction, including the ones being made now.
Read it with the dates in mind, keep the criticisms that still land, and treat the whole thing as an object lesson in the difficulty of seeing where a technology is going while you are standing inside it.