They show that extended pre-training can actually make language ... on a surprising trend observed in modern LLM development: while models are pre-trained on ever-expanding pools of data ...