Back to index
OtherA full academic year if worked properly, reference for life·Around $150 to $250 new, cheaper used, international editions widely available

Artificial Intelligence: A Modern Approach

4.4

The book that defines what the field considers AI, written with unusual clarity and breadth. It is a textbook rather than a course, and the deep learning coverage is thinner than its reputation suggests.

What We Liked

  • Unmatched breadth across the whole field, not just the fashionable parts
  • Writing is genuinely clear, which is rare in a text this comprehensive
  • Pseudocode for every algorithm, so you can implement rather than admire
  • Gives you the vocabulary and mental models the field actually uses
  • The ethics and philosophy chapters are thoughtful rather than bolted on

What Could Be Better

  • Enormous, and nobody reads it cover to cover successfully
  • Deep learning coverage is present but not where the book is strongest
  • Exercises range from trivial to research problems with little signposting
  • Large sections cover material you may never touch in practice
  • New copies are expensive for something you will read selectively

Detailed review

Almost every university AI course in the world either uses this book or was designed by someone who did. That kind of dominance usually signals inertia rather than quality, so it is worth saying plainly that the reputation here is deserved. Russell and Norvig wrote something genuinely good and then kept improving it across four editions and thirty years. The organising idea is the intelligent agent, something that perceives an environment and acts in it.

Everything hangs off that frame, which sounds like a textbook device and turns out to be a real intellectual contribution. Search, logic, probability, planning, learning and communication all become different answers to the same question of how an agent should decide what to do. Once you have that frame, a lot of AI stops looking like a pile of unrelated techniques and starts looking like a discipline. The writing is the thing people underestimate.

Comprehensive textbooks are usually turgid. This one is not. Explanations are patient, examples are chosen well, and the authors have a dry sense of humour that surfaces just often enough to keep you awake. Norvig in particular writes prose that reads like a person thinking rather than a committee documenting.

Every algorithm comes with pseudocode. That matters more than it might sound. You can implement A star, or minimax with alpha beta pruning, or variable elimination, directly from the page, and implementing is where understanding actually happens. The Berkeley course materials and the associated code repositories give you working versions in several languages if you get stuck.

The search and logic chapters are the best available treatment I know of at this level. Constraint satisfaction, adversarial search, knowledge representation, first order logic and inference are covered with a thoroughness that pays off for years. The probabilistic reasoning material on Bayesian networks and inference is similarly strong. The planning chapters are dense and reward slow reading.

Now the honest limitation. This is a book about artificial intelligence broadly, and the last decade of the field has been overwhelmingly about deep learning. The fourth edition expanded that coverage and it is competent, but it is not why you should buy this book. If you want neural networks, use Goodfellow or Prince or Zhang, and if you want language models specifically, use Jurafsky and Martin or Raschka.

Come here for everything that surrounds them. The size is a real problem in practice. Well over a thousand pages, and almost nobody finishes it. I think the correct approach is to stop treating it as a book to be read and start treating it as a reference with a good table of contents.

Read the agent chapters, then read whatever your current work or curiosity demands, then come back when you need something else. A term of a course covers maybe a third of the material and that is normal. Exercise quality is uneven. Some are five minute checks that you were paying attention.

Others are genuinely hard research adjacent problems that a strong graduate student would spend a week on. There is not enough signposting to tell them apart before you commit an evening. The ethics, safety and philosophy material is better than the equivalent chapters in most technical books, which usually treat the subject as an obligation. Russell has spent years thinking seriously about the control problem and it shows.

These chapters engage with hard questions rather than reciting principles. There are large sections that many working practitioners will never use. Logical agents, classical planning and formal knowledge representation are foundational to the field and marginal to most industry work. That is a fair criticism of the book as vocational training and not a criticism of it as education.

My four point four is for the best single overview of artificial intelligence as a discipline, written clearly, organised intelligently, marked down because the length defeats most readers, because the deep learning coverage will send you elsewhere anyway, and because a meaningful fraction of the content is history for most practical purposes. Every serious AI practitioner should own it. Almost none of them should try to read it in order.

[ final ]

The verdict.

Buy it, keep it, and read the chapters you need rather than attempting the whole thing. If you want to build language model applications this week it is the wrong book. If you want to understand what artificial intelligence means as a discipline, nothing else comes close.