Tag
This paper presents a theoretical model of AI alignment, identifying conditions under which an imperfect proxy to human values can lead to catastrophic outcomes when an agent optimizes too heavily, motivating safer designs like quantilizers.