On “interpretability creationism” – interpretability methods that only look at the final state of the model and ignore its evolution over the course of training