The challenges, the challenges are very probabilistic and a different interface to anything we know. AI is kind of wild beast to tame. You ask first the interface is language not code. And then what comes back is like semi-probabilistic in nature. The model is supposed to be static. But if you use GPT 5 3.5 turbo, for instance, that's just one trained model. And I believe that that is immutable. It feels like it should, but it doesn't. That starts to get into reproducibility, accuracy, and just those limitations, which is kind of an engineering type of thing. Yeah.