ParentFull threadbitL·There was no hugely scalable approach before transformers, RNNs, the previous SOTA, were notoriously bad at scaling.View on HN