Comment by philkuz

Caveat buried in the abstract is that this beats BERT and non-pretrained Transformers. Looks like GPT style should still be better, but naturally requires a higher computation cost