(read a physical book)
An interesting combination of introduction to AI concepts, exposition of a fundamental problem and potential solution, and philosophy. I was drawn to this book because I took an introductory AI course as part of uPenn's MCIT program, and Stuart Russel was one of the authors of the textbook "Artificial Intelligence - A Modern Approach". The textbook dedicates some space to addressing the central thesis of Human Compatible, which is that the "standard" of AI, which relies on maximizing some fixed objective, is flawed and needs to be changed to a model where the AI is uncertain about it's objective. This is because it is very hard to figure out ahead of time exactly what an AI is doing, and as AI becomes more powerful, there is more and more risk that AI will cause harm by focusing on a narrow goal. I think that it makes a lot of sense to develop AI that will keep on seeking input from humans to check that it is indeed being of benefit.

The first third of the book or so is mostly an introduction to various fundamental concepts of AI. Nothing too technical, and could serve as a good intro for people without much exposure. Then talks about the "alignment" problem and solution. Last part of book spends a fair amount of time touch on more philosophical questions, like how to design an AI that balances rights and needs of different people. Interesting, but kind of felt like it was dragging on for a while. The book does a good job of addressing an important topic, but felt a bit padded, even though it wasn't that long.