Instead of building a model from scratch, you start with a general one and continue training on a narrow dataset — your support tickets, your codebase's conventions, a medical corpus. The model keeps its general ability and gains your specifics.

The practical tension: fine-tuning competes with just putting examples in the prompt or using retrieval (RAG). Prompting is cheaper and instantly reversible; fine-tuning wins when the behavior change is deep, the examples are many, or per-request prompt length is too expensive.