Hacker News new | past | comments | ask | show | jobs | submit
I think the idea is you sink the pretraining costs once and then you can distill multiple specialized models from that