https://youtu.be/Up9CvRLIIIw?t=478 This might let us decorate `run` with `tf.function`, resulting in better performance.