“Misha Laskin asserts that reward models for vision or other sensory signals in robotics are significantly more prone to being hacked by RL agents than reward models for language.”