Human Compatible
Artificial Intelligence and the Problem of Control
Berkeley AI pioneer Stuart Russell argues that the standard model of building ever-smarter machines to pursue fixed goals is a mistake, and proposes a new foundation for keeping AI provably beneficial to humans.
Stuart Russell literally wrote the textbook on artificial intelligence, and here he argues that the field has been building it wrong. Machines told to pursue a fixed objective as capably as possible become dangerous precisely when they succeed, since a goal that leaves out something we care about can be optimized to our ruin. His fix is to design machines that are inherently uncertain about what humans want and defer to us, a redesign he argues is essential before AI grows more capable. Elon Musk flagged it as worth reading on exactly those risks.
Recommended by 2 People
Scott Alexander
I recommend this book both for the general public and for SSC readers. The general public will learn what AI safety is. SSC readers will learn what AI safety sounds like when it’s someone other than me talking about it. Both lessons are valuable.
Book Review · January 2020
As an Amazon Associate, bookstoread.org earns from qualifying purchases. Links to Amazon on this page may earn us a commission at no extra cost to you. Spotted an error on this page? Tell us and we will fix it.