Yeah, the surprising validators. Yes. And I think that is really the point. With language models, one can now identify, in a huge amount of text—a rant that is a personal attack, very toxic or whatever—the few words in it that can be construed as constructive and bridge-making. I always make sure to profusely thank them, sometimes publicly, for this very specific part. This teaches them that the other parts, designed to induce a sense of status harm or polarization, simply have no effect. Only the bridge-making part has an effect.