But fine-tuning is very different from (pre)training. Pretreating proceeds via unsupervised learning on massive amounts of data and compute, while fine-tuning uses much smaller amounts, with supervised learning (instruction tuning) and reinforcement learning (RLHF, constitutional AI).