Morgan McGuire says SENPAI autoresearch cut training time by 22%
He credits a SliceAttention kernel and BatchedMuon, with Transolver in current use, and reports a full run moving from 26 hours to 22 hours without a significant change in error or quality.