Alignment 'train/test splits' should be sacred
Tl;dr: Alignment techniques can be used for training alignment or for testing it, but they cannot be used for both! As we build out the set of alignment techniques available to us, we should enforce that some are judiciously reserved only for evaluating alignment.See also: Validating against a misalignment detector is very different to training against one.Here is a guiding principle regarding alignment-minded AI training: whatever you do, you must maximally avoid teaching the model to take misa...
Read full article →