Training large models from scratch is expensive and rarely the right first move. It requires data at scale, specialized training infrastructure, and enough iteration budget to recover from failed runs.

In practice, the goal is almost always a model that performs well on one specific task under production constraints, not a rediscovery of general visual, linguistic, or multimodal representations from raw data.

Transfer learning starts from a model that already learned useful representations, then adapts it to a target domain or task. Fine-tuning is one adaptation method; prompting, retrieval augmentation, adapters, LoRA, and distillation are often part of the same design space.