Ajeya comments on Why AI alignment could be hard with modern deep learning

[Ajeya]

I was imagining Sycophants as an outer alignment failure, assuming the model is trained with naive RL from human feedback.