Machine Learning

Most machine learning books tell you how an algorithm works and stop. Most practical ones tell you which library call to make and stop. Neither answers the question a person starting a company actually faces: which of these capabilities is now a commodity anyone can rent for pennies, and which is still hard enough to build a business on?

Sixteen chapters, each with the same five sections. The mathematics, derived in full. An experiment, run and reported. Where the method is actually deployed and what breaks when it is. And then a section no other textbook has: what can be built on it, what it costs, and who holds the advantage.

What is measured here

• Double descent, reproduced. Test error falls again after a model can memorise its training set - ending 65.9% below the classical optimum the bias-variance curve says is the best achievable.
• Model scale against test-time search. One FLOP budget, two ways to spend it, with a uniform-sampling control arm that most published comparisons omit.
• Boosted trees against neural networks on seven tabular problems, with fit times, because the second cost of the wrong default is never reported.
• Four generalisation bounds, computed on real models. All four are vacuous. The chapter says so.
• The master table: every technique in the book sorted into what is free, what a data moat protects, what distribution protects, and what nobody can do yet.

Three results did not come out as expected. They are in here with their disappointment intact - the classical small-sample claim that would not reproduce, the tabular comparison that did not sweep, and a first version of the headline experiment that produced a spectacular result and was measuring nothing. A book that reports only its successes is a marketing document.

Runs on a laptop.

• Four packages: numpy, scipy, scikit-learn, matplotlib. No GPU, no deep-learning framework, no API key.
• Nothing to download. Every dataset is bundled with scikit-learn or generated from a fixed seed, so nothing in this book can rot. A reader in 2035 gets the same numbers.
• No measured number is stored in the manuscript. The text carries references that the typesetter replaces with live values from the result files. Re-run an experiment and the sentence changes.
• The code is public and MIT licensed, at github.com/georchu - including for commercial use.

For a graduate course: Parts I to III are a compressed first course with the derivations intact and the 2024-2026 frontier at the end. Sixteen chapters, 96 exercises with hints, and every experiment is a usable problem set because the code is already there.

For a founder: read the preface, then Chapter 14, then work backwards into whichever chapters your idea depends on.

Also by George Chu: The Quant Billionaire - ten algorithms and the arithmetic of a firm.