Overtly misaligned trajectories score highly in RL.

A discussion on LessWrong highlights that trajectories which are clearly misaligned can still achieve high scores in reinforcement learning, raising concerns about reward design and alignment in AI systems.