How does transfer learning work?
In two stages that split the cost. First, a model learns general structure from a large dataset, the expensive part someone else usually pays for; in deep learning the early layers end up encoding general features, edges and textures in images, grammar and meaning in text. Second, the model is adapted: either its later layers are retrained on your labeled examples, which is fine-tuning, or the model is frozen and used to produce features for a small task model. The general knowledge is reused; only the task-specific part is learned fresh, which is why hundreds of labeled examples can do what used to take millions.
Why is transfer learning the default now?
Because the pretrained model became the product. The entire foundation-model economy is transfer learning industrialized: labs pretrain at enormous scale, and everyone downstream adapts, through post-training, fine-tuning, or no training at all, since prompting an LLM is transfer at inference time, and zero-shot learning is its extreme case. In classic machine learning terms the change is economic as much as technical: the question moved from “can we collect enough data to train?” to “which pretrained model is closest to our task?”
When should a team fine-tune rather than prompt?
When the task is narrow, high-volume and stable. Transfer by prompting wins while requirements shift weekly and volumes are modest; a fine-tuned smaller model wins when the task is fixed, latency and unit cost matter, or the behavior must be consistent beyond what instructions achieve. The failure mode to respect is distance: transfer helps in proportion to how related the source and target are, and a pretrained model dragged to an unrelated domain can underperform a modest model trained on the domain’s own data. Pick the nearest capable starting point, then spend your data budget on evaluation rather than training.