Morgan McGuire says SENPAI autoresearch cut training time by 22%

Morgan McGuire says SENPAI autoresearch reduced their training time by 22% by writing a SliceAttention kernel and BatchedMuon. He says they are using Transolver for now, and that a full training run fell from 26 hours to 22 hours with no significant change in error or quality. The project repository is SENPAI on GitHub.