Well, since we haven't solved alignment, there's not even a guarantee that unpoisoned training data is good enough.