Nice to see successes in deep learning moving beyond transformers. This one is more accurate and scales much better without ridiculous memory requirements.