Is Differential Privacy a Good Idea?

The models are so powerful, but they definitely have a huge amount of unintuitive properties that need to be studied quite carefully. On the on the one silver lining on this is that now that we can create these models that are getting in at least some small subsets closer to what humans might be able to do, they will be easier to study than at least the brain. Trey: I wanted to ask about to Richard's point about explainability based on different training examples. How you think about knowing when a model is safe enough for critical slices of data?