The Hundred-Page Machine Learning Book — Burkov
Review of Andriy Burkov’s compact ML reference: surprisingly dense, best used as a refresher rather than a first introduction.
Andriy Burkov’s The Hundred-Page Machine Learning Book (2019) is exactly what the title promises: a compressed tour through the core of machine learning in roughly 100 pages of actual content. It covers supervised learning, unsupervised learning, neural networks, SVMs, kernel methods, regularization, model evaluation, feature engineering, and even touches on deep learning architectures—all at a pace that would make a PowerPoint deck jealous.
The obvious question: can you actually learn machine learning from 100 pages? The answer is no, and Burkov seems to know it. The book reads less like a textbook and more like a well-organized cheat sheet written by someone who has taught the material many times and knows exactly which corners can be cut without total loss of meaning.1 If you already understand gradient descent and have implemented a few models, this book serves as an excellent refresher—the kind of thing you flip through before an interview or when you need to quickly recall the difference between L1 and L2 regularization. If you don’t have that background, you’ll find yourself constantly reaching for supplementary materials.
What Works
The coverage-to-page-count ratio is genuinely impressive. Each chapter is 5–15 pages and manages to convey the essential intuition behind each algorithm. The notation is consistent throughout—something that even 1,000-page textbooks frequently botch.2 The diagrams are clean and purposeful. There’s a chapter on best practices for ML projects that, while brief, contains advice that many practitioners learn only after years of painful experience: always establish a baseline, be skeptical of accuracy as a metric, watch for data leakage.
The mathematical treatment hits a sweet spot for working engineers. Burkov includes just enough formalism to be precise—he gives the actual loss functions, the actual update rules—without drowning the reader in measure theory or convergence proofs. For someone who needs to use ML rather than prove theorems about ML, this is the right level.
What Doesn’t
The brevity that makes the book useful as a reference also makes it inadequate as a learning resource. Ensemble methods get maybe three pages. Bayesian learning is compressed to a point where anyone who doesn’t already understand it will come away more confused than they started.3 The neural network chapters are too thin to be useful in the post-transformer era—though this is partly a consequence of the 2019 publication date.
There’s also a structural issue: by covering everything at the same depth, Burkov gives equal weight to topics of wildly different practical importance. k-nearest neighbors gets roughly the same treatment as gradient boosting, despite the latter being vastly more important in applied work.4
Detailed Chapter Notes
The book proceeds in a logical order: mathematical notation → supervised learning fundamentals → basic algorithms (linear regression, logistic regression, decision trees, SVM, kNN) → neural networks → problems and solutions (overfitting, regularization, model selection) → advanced topics (ensemble methods, representation learning, dimensionality reduction) → unsupervised learning → other forms of learning (semi-supervised, reinforcement learning). The final chapters on best practices and conclusion are arguably the most valuable for practitioners. Each chapter ends with a brief summary that could itself serve as a cheat sheet of the cheat sheet. The appendix includes a notation reference and recommended reading list that points to the deeper treatments one should consult for any topic that the book merely sketches.
Who Should Read It
If you’re a software engineer preparing for ML interviews, this is nearly ideal preparation material. If you’re a student who has taken one ML course and wants a compact reference, it works well. If you’re trying to learn machine learning from scratch, start with Andrew Ng’s Coursera course or Bishop’s Pattern Recognition and Machine Learning instead, then come back to Burkov as a reference.5
The book has aged tolerably well despite the rapid pace of ML progress. The fundamentals it covers—bias-variance tradeoff, regularization, cross-validation—remain as important as ever. What’s missing is any serious treatment of transformers, attention mechanisms, or large language models, but adding those would have blown the page budget and the book never promised to be comprehensive.
Verdict: A 6/10 in importance. Useful, competently executed, but ultimately a reference card rather than a book that changes how you think. Keep it on your desk; don’t make it your primary learning resource.