Artificial Intelligence: A Modern Approach

Russell & Norvig

The standard comprehensive textbook of artificial intelligence, organizing the field around the idea of the rational agent. It spans problem-solving search, logic and knowledge representation, planning, probabilistic reasoning under uncertainty, and machine learning, with the fourth edition adding deep learning, transformers, and the safety of advanced systems.

Key takeaways

Organizes the whole field around the rational agent: an entity that perceives its environment and acts to maximize a performance measure, a frame that unifies otherwise disparate AI techniques.

Covers classical problem solving through search (uninformed, heuristic, and adversarial), the algorithmic backbone (A*, minimax, constraint satisfaction) that predates and underlies modern learning methods.

Develops knowledge representation and reasoning via propositional and first-order logic, planning, and inference, treating explicit symbolic reasoning as a first-class capability.

Presents probabilistic reasoning under uncertainty with Bayesian networks, Markov models, and decision theory for acting rationally when the world is only partially known.

Situates machine learning as one component of a broader intelligent system rather than the whole of AI, placing supervised, reinforcement, and deep learning inside the agent framework.

The fourth edition expands coverage of deep learning, transformers, probabilistic programming, and the ethics and safety of increasingly capable systems.

Reflections

AIMA is the map I keep coming back to when the discourse forgets there was a field before large language models. Russell and Norvig frame everything around the rational agent, a thing that perceives and acts to maximize an objective, and that framing has aged better than most of the specific techniques in the book. It’s a useful corrective to recency bias: search, logic, planning, and probabilistic reasoning didn’t stop mattering because transformers arrived; they’re the substrate a good agent still needs underneath the model. I treat it less as a how-to and more as a sense of proportion. The current moment is one chapter in a long book, and knowing the earlier chapters is what keeps me from mistaking a powerful tool for the whole of intelligence. The fundamentals are cheaper to learn than to rediscover the hard way.