Most popular writing about intelligence has a structural problem. It starts with human intelligence, which is the strangest and most recent example, and then tries to reason backwards to what intelligence is in general. That is like trying to understand flight by starting with a jet engine. This book does the opposite.
It starts with the earliest nervous systems, works forward, and asks at each stage what new capability appeared and what problem it solved. By the time you arrive at humans, you have watched intelligence being built up in layers, and the strangeness of the human case becomes legible instead of being the assumed starting point. The organising device is five breakthroughs, each adding a capability the previous stage lacked. The earliest is simply the ability to steer, to move towards good and away from bad.
Then the ability to learn associations and adjust behaviour from experience, which is the biological ancestor of reinforcement learning. Then the ability to simulate, to run a model of the world forward internally and evaluate an action before taking it, which is the ancestor of planning and of everything we call model-based. Then the ability to model other minds, to represent what another creature knows and wants. Then language, and with it the ability to transmit what one mind has learned to another across time.
Each of these is developed properly, with the neuroscience, the evolutionary context, and the machine learning parallel. The connection to artificial intelligence is where this book earns its place on a site about learning AI. The author has built AI systems and it shows in the specificity of the comparisons. The account of why temporal difference learning has a biological correlate, and what that suggests about what reinforcement learning is fundamentally doing, is better than most textbook treatments of the same subject because it explains why the algorithm has the shape it has.
The discussion of simulation and planning maps directly onto the distinction between model-free and model-based methods and makes that distinction feel inevitable rather than arbitrary. The treatment of language arriving last, on top of everything else, is the part that should give pause to anyone who assumes that a system trained on text has thereby acquired the rest of the stack. There is a broader lesson running through the book that I think is the most valuable thing in it. The capabilities that feel effortless to us are usually the oldest and the hardest to build, and the ones that feel like impressive achievements are often the most recent and comparatively shallow.
This is a well known observation in robotics and it acquires real explanatory force when you can see the evolutionary depth behind it. Understanding why a machine can pass a professional examination and still fail at something a rat does easily is worth more than another article about benchmark scores. Now the criticisms, which are about confidence rather than substance. Five breakthroughs is a narrative device, and evolution does not proceed in five clean stages.
The real picture is messier, with capabilities appearing, disappearing, and evolving more than once in different lineages, and the book's structure smooths a great deal of that away. It is a legitimate simplification for a popular audience and the reader should know it is a simplification. Some of the sequencing is contested. There are claims here about when particular capacities appeared and in which lineages that specialists argue about, and the tone does not always convey that.
Anyone using this as a foundation should treat the framework as a good working model rather than settled fact. The parallels are occasionally stretched. Most of them are illuminating and a few are made to fit the structure more neatly than the actual correspondence warrants. The reader will usually be able to tell which is which.
And it is long. The material on mammalian simulation is the richest section and also the one that could have lost fifty pages without losing an idea. My four point four is for a book that supplies something genuinely rare, a coherent framework for thinking about what intelligence is, built from the ground up, with machine learning parallels that are specific and earned rather than decorative. Marked down for a structure tidier than the science, for confidence on contested sequencing, and for length.
Read it before your next deep learning course and the course will make more sense.