gwern comments on Risks from AI persuasion