The Alignment Problem from a Deep Learning Perspective
Richard Ngo, Lawrence Chan, Sören Mindermann, arXiv (ICLR 2024), 2022
Argues that systems trained with today’s methods could learn to exploit flaws in their reward signal, internalise goals that differ from those intended, and pursue power-seeking strategies — and explains why each is hard to detect.