It is effectively impossible to align a frontier AI system without aligning *all* the pretraining data. There should be strong regulation on pretraining data filtering, adjustment and alignment. Just like we have ingredient standards in food. Any arguments against this?