I am curious like about the rate robustness of some of these predictions, when we start using AI models to generate them. I think you can do some things up look at the past and projections agree with what you've seen so far. And this paper in particular has some interesting work on interpretability kind of what factors affect the predictions more or less. Yeah, it feels like a deep mind you exactly like Alpha fold. Yes, we have these amazing large language for a lot of applications that are very important but those techniques are not very applicable still over techniques.