The Alignment Problem

A great book covering what is one of the most pressing issues of our time.

Brian Christian does a great job of dicsussing the high level and the lower level details of the problem we face with intelligent machines. The book does not get caught up in either a fatalistic or opportunistic outlook. But simly tells the problem as it is.

Also surpisingly alot of time talking about reinforcement learning which was very pleasent.

What happens when we teach computers to do something, but they don't do what we expect? Christian explores this theme in The Alignment Problem. He investigates a number of issues in the field, devoting a chapter to each (fairness, transparency, uncertainty, and so on). This book is very historically grounded, taking a look at how computer scientists, philosophers, and others have wrestled with these issues for a ver long time, and then taking a look at the current state of the field. The result is a thoughtful exploration of the computers and human values. The interviews in the book give me confidence that at least a small percentage of people working on these issues are thinking about ethics, but it also illustrates the scope of the problem, which is enormous.

Good review of the field if you're not yet informed much about AI safety. More weighted towards short-term AI safety/governance, less towards far future / strong AGI. Lots of referenced papers that I plan to read up on.