We already pretrain them on random internet crap then finetune on supervised data