Human Compatible: Artificial Intelligence and the Problem of Control Paperback – 17 November 2020

Excellent book.

For the price of Rs 1231 + the book looked used and a second hand one ..

The book primarily contends that the formulation of artificial intelligence as the ability of a machine to achieve it's stated goal leads to problems in value alignment as well as the machine coming up with non cooperative subgoals. The author lays out the foundation of a provably beneficial AI which instead learns it's objectives from the revealed preferences of human behavior. Thus the machine achieves more and more certainty of human objectives with time and allows the human greater control over how the machine behaves. The author delves deeps into economics and has discussions on state of art in rationality, utilitarianism, preference comparison and updating as well as recent findings from cognitive science. The author lays open his foundation of provably beneficial AI (or inverse reinforcement learning) to be extended to cases with multiple humans, sadism or altruism or population dynamics which are challenging. Lots of good and simple examples are provided to illustrate the authors points. In addition, the author provides optimistic narratives for an AI powered abundance as well as pessimistic narratives where humans are sidelined by a superintelligent AI. The language of the author is free of jargon and good appendices and references are provided. The author also takes on luminaries in AI and addresses their thoughts, which is useful if the reader has been following them. A very useful book for an important problem.