Post #1973409

@lmorchard @leeloo These specific models - yes, probably. One plausible argument I heard for it is that there's a common failure mode in ML where the model fails to generalize, but if the verification set overlaps the training set, then data leakage will fool the authors into thinking it generalized. Another one is that these models were "rewarded" for saying plausible things, not for interacting with a world in a way that doesn't get them killed. But these arguments are specific.