logoalt Hacker News

supermdguyyesterday at 9:33 PM0 repliesview on HN

It's interesting because their last model series (Phi) was based around the thesis that high-quality synthetic data is better than a large pre-training corpus.