Roughly. It's about learning a decoder for a blackbox decoder. It's much harder in an environment that is not end-to-end trainable.