Why Post-Training Is the New Pre-Training in AI
_Personal notes on data labeling, RLHF, and where the real gains are happening now._ --- I used to think model quality was mostly a function of scale—more tokens: more parameters, more compute, etc....