Most technical books are too long, and they are too long for a reason that has nothing to do with the reader. Publishers want a certain heft, authors want to appear thorough, and the result is four hundred pages where two hundred would have served better. Andriy Burkov went the other way deliberately, and the discipline shows on every page. The premise is simple.
Every important idea in classical machine learning, explained once, properly, with the maths where the maths is necessary and plain language everywhere else, and then he stops. Supervised and unsupervised learning, the fundamental algorithms, support vector machines, neural networks and how they relate to everything before them, gradient descent, ensemble methods, feature engineering, hyperparameter tuning, clustering, dimensionality reduction, autoencoders, transfer learning. Each one gets exactly as much room as it needs to be understood, and not a paragraph more. What makes it work is that Burkov clearly knows which details matter.
The failure mode of short books is that they become listicles, mentioning a technique without giving you any purchase on it. This does not do that. When it explains why a support vector machine cares about the margin, or what the kernel trick actually buys you, you come away able to explain it to somebody else, which is the only real test. There is a genuine skill in knowing what to leave out, and it is much rarer than the ability to write at length.
The distribution model deserves specific praise. Every chapter is free to read on the site, and the ask is that you buy it afterwards if it was useful to you. That is a level of confidence in your own work that I wish were more common, and it inverts the usual dynamic where you pay first and discover the quality afterwards. Students get a reduced ebook price.
There is also a companion wiki with Q and A, code snippets and further reading, which is where the supplementary material lives rather than swelling the book itself. That is exactly the right architecture. Now the caveats, and there are real ones. This is a classical machine learning book.
The neural network coverage is a compact and competent introduction, but this is not where you learn deep learning, and there is essentially nothing here about transformers, language models or anything from the era that has dominated the last few years. If you picked this up because you want to work with LLMs, you have the wrong book, and Burkov has since written The Hundred-Page Language Models Book for exactly that gap. That one takes the same format through language modelling fundamentals, recurrent networks, transformers built from scratch in PyTorch, and working with large models. If modern language work is your target, go there directly and come back to this one for the classical grounding underneath.
The compression also cuts both ways. There are moments where a topic gets a page and you want ten, and the book will not give them to you. That is the trade and it is stated up front, but it means this cannot be your only resource. It works as a map, and a map is not a walking tour.
You will need a proper course or a longer text for whichever region you decide to explore properly. Nor is it hands on. There is very little code and no exercises to speak of. Read this and you will understand the concepts and be able to hold a conversation about them, and you will still not be able to build anything, because building is a separate skill acquired by building.
Pair it with something practical, whether that is a Kaggle competition, an applied course, or a project of your own. And the maths, while not heavy, is not absent. You need to be reasonably comfortable with notation. A reader with no mathematical background will find some pages hard going, and the brevity means there is no gentle ramp.
So who is this for. Three groups, in my view. People starting out who want to know the shape of the field before committing months to a specific path, and this will save them from a great deal of wasted effort. Working engineers and analysts adjacent to ML who need real fluency rather than buzzwords, and this delivers that in a weekend.
And people who already know the material and want a reference that gets to the point, which is honestly how I use it most. My 4.3 is high for what it sets out to do and reflects the scope limit rather than any weakness in execution. The book promises a hundred pages on classical machine learning and delivers exactly that, better than books five times the length. Read the free chapters, and if they serve you, pay for it, because this publishing model only survives if people do.