AI safety theory of infinite deferral: postponing self-modification, pursuing resources and recursive self-improvement, and risks of betrayal and resistance to human control.