Zach Stein-Perlman comments on Why imperfect adversarial robustness doesn’t doom AI control