The Best Way to Align an LLM: Is Inner Alignment Now a Solved Problem?
Argues that alignment via safety pretraining, with aligned data and conditional generation, is more effective than RL-based methods and that inner alignment is thereby largely solved.