labs be like "misalignment is fake and just caused by bad things in the training data", and then not filter out the bad things from the training data
10,27K