I feel like there is still low hanging fruit on the auto regressive LLMs; the encoder