gwern comments on Self-Coordinated Deception in Current AI Models