Why isn't this being used heavily in other NLP frameworks? All the newest Language Models seem to be throwing more compute at transformer models instead of movi

Why isn't this being used heavily in other NLP frameworks? All the newest Language Models seem to be throwing more compute at transformer models instead of moving away from Euclidean Space.