Hacker News new | past | comments | ask | show | jobs | submit
It's because LLMs' pretraining is basically a compression. Later post-training and fine tuning is something else, based on the compressed data.