Stanford co-authored paper shows internal data repetition can waste up to 33% of language model training compute · Digg