[kadushka]

There’s not enough improvement over regular LLMs to motivate optimization effort. Recall that the original transformer was well received because it was fast and scalable compared to RNNs.