Peter Norvig: Artificial Intelligence: A Modern Approach

Lex introduces Peter Norvig, Google's director of research and co-author with Stuart Russell of Artificial Intelligence: A Modern Approach, the field-defining textbook that educated and inspired a whole generation—including Lex—to enter AI.

Asked how the book changed across editions, Norvig points first to the rise of computing power: first-order logic once chosen for conciseness gave way to realizing fast SAT solvers and ample memory made predicate logic practical, and now GPUs, TPUs, and custom ASICs drive another hardware advance.

Another shift, Norvig says, is that the first three editions framed AI as maximizing expected utility with 27 chapters of optimization techniques, whereas this edition suggests optimization may be the easy part and the hard part is deciding what your utility function is—and, for a society of agents, what we want as a whole. The philosophy chapter now leans less on the Chinese room and more on ethics, fairness, bias, and aggregating utilities.

Asked how to encode human values into a utility function, Norvig says there's no single answer but techniques like inverse reinforcement learning: rather than taking actions for rewards, you observe someone's actions and infer what they were trying to do—though people act self-destructively or suboptimally, so you want to learn their ideal actions, a challenge for the field.

Norvig describes the theoretical side through recidivism-prediction programs for parole and bail: you want a score to mean the same across protected classes like race, and you want mistakes to affect both sides equally. A commercial program achieved the first but was twice as likely to harm a black defendant—and analysis shows both goals are theoretically impossible to satisfy at once, so you must trade them off. The analysis tells you what's achievable, but not where to set the trade-off.

Lex calls removing undesirable bias across protected classes a powerful, important, and doable first step, now informed by what's actually achievable, with humans deliberating the trade-off point.