Can we predict the kinds of domains and tasks for which large generative models are more likely to be robust (or not robust), and how does this relate to the training data? Are there techniques we can use to predict and mitigate worst-case behavior?