It's not that surprising that persuasion is more effective on more advanced models. As Hinton said, predicting the next token takes knowledge. And as we've see

It's not that surprising that persuasion is more effective on more advanced models. As Hinton said, predicting the next token takes knowledge. And as we've seen, complex models develop linear representations of world models. We should expect that more advanced models trained on anthropomorphic data should be vulnerable to anthropomorphized persuasion techniques. The more surprising thing is the bull headed resistance to this idea we continue to see so commonly.