Human Compatible by Stuart Russell argues AI should stay uncertain about human goals, not certain about the wrong ones. Summary, 3 principles, and a 7-day plan.