logoalt Hacker News

p0w3n3d • yesterday at 7:27 PM • 0 replies • view on HN

It's because LLMs' pretraining is basically a compression. Later post-training and fine tuning is something else, based on the compressed data.