[jedharris]

"Our results show that simple algorithmic improvements can enable significantly more data-efficient pre-training in a compute-rich future."