Popular books about artificial intelligence almost all make the same choice. They decide the mathematics would frighten the reader, remove it, and replace it with analogy. The result is a genre of books that leave you able to talk about machine learning at a dinner party and no closer to understanding it, because the thing that makes it work is precisely the part that was removed. Ananthaswamy made the opposite choice and it is why this book matters.
He shows you the equations. Not all of them and not in full rigour, but the ones that carry the ideas are on the page, explained line by line, with the reasoning for why each piece is there. The bet is that a curious reader will follow real mathematics if it is motivated properly, and in my experience the bet pays off. The structure is historical, and that turns out to be the right pedagogical decision rather than just an appealing narrative one.
Ideas arrive in the order they were needed, in response to problems that were actually blocking people, which means each one feels like a solution rather than a definition. The perceptron appears with its optimism and then its famous limitation, and the limitation is shown mathematically rather than asserted. Nearest neighbours, the kernel trick, support vector machines, backpropagation and the eventual arrival of deep networks each land as a response to something specific. Compare this with the standard course structure, where you are handed a loss function in week one and told to minimise it.
Knowing that gradient descent was not obvious, that people argued about whether neural networks could work at all, that the field went through two winters, makes the current state feel contingent rather than inevitable. That is a more accurate picture of a scientific field and it is a better foundation for judgement. The geometric intuition is the strongest technical thread. The book is very good at getting you to see what is happening in the space rather than manipulate the symbols.
Why a linear classifier is a hyperplane and what that means when you cannot picture the dimensions. Why the kernel trick works, which is one of those results that seems like cheating until somebody shows you the geometry, at which point it becomes obvious. Why distance behaves so counterintuitively in high dimensions and why that breaks methods that seemed reliable in two dimensions. I have read several textbook treatments of the curse of dimensionality and this is the one that made it feel real.
The historical reporting is worth having on its own. Ananthaswamy is a science journalist and he interviewed people, so alongside the mathematics you get the arguments, the funding decisions, the personalities and the periods where the current consensus was a fringe position held by a handful of stubborn researchers. Field history is not decoration. It calibrates you about how confident to be in whatever everyone currently believes.
Now where it falls short. The middle ground is the book's defining feature and also its main risk, and the reviews split predictably. Readers who wanted a popular account find the equations intimidating and stall. Readers with a technical background find the pace slow and the coverage shallow relative to a textbook.
The book is genuinely aimed at a person who is comfortable with algebra, willing to work at calculus and linear algebra, and does not want to be patronised, and that person is a real audience but a narrower one than the marketing implies. Know which reader you are before buying. The bigger practical limitation is that it produces understanding and not capability. There is no code, there are no exercises, nothing to implement.
You will finish it knowing why backpropagation works and unable to write it, which is a legitimate goal but not the one many buyers have. It pairs extremely well with something hands on and functions poorly as a standalone education. My recommendation is to read it before or alongside a serious course, where it does its best work by making the formalism feel motivated instead of arbitrary. Coverage also thins at the modern end.
The foundations get patient, detailed treatment and then the last stretch, covering the deep learning era and transformers, moves much faster and shallower. That is understandable, because the recent material is harder to explain at this level and the book's method depends on ideas being simple enough to show. It does mean a reader hoping to understand why current systems work will find the earlier chapters far more satisfying than the ones about the technology they came for. The mathematical appendices are meant to catch readers who need a refresher and they are uneven.
Some are clear, others compress a semester into a few pages in a way that will not help someone genuinely missing the background. If linear algebra is a gap for you, fix it elsewhere first rather than relying on these. And it is long. Four hundred and eighty pages of narrative to cover ground a textbook handles in far fewer, which is the cost of the historical approach and the reason it works.
Whether that trade is worth it depends on whether you want to enjoy the process. My 4.4 is for a book doing something almost nobody attempts, and doing it with real skill. It respects the reader's intelligence, it shows the mathematics, and it makes a field that usually arrives as a set of arbitrary techniques feel like a story with reasons in it. It will not make you an engineer.
It will make you much harder to fool.