Aaron Bergman comments on Half-baked alignment idea: training to generalize