Distillation Robustifies Unlearning — LessWrong