I think a better simplification would be to remove the position-encoding altogether, and just discuss transformers as set models.