The Pretrained Era: Learning What to Reuse

Part IV leaves us with a powerful pretrained backbone and a new question: what should be reused, and what still needs to change? This Part separates choices that are easy to conflate: where task-specific information lives, what objective assigns the update, and how a model produces or compares outputs. Prompting, parameter-efficient adaptation, and quantization move different resource bills; instruction and preference learning change which behavior is rewarded. Generative models add the missing sampling contract, while paired image–text data makes a cross-modal comparison learnable. Each gain carries a boundary: a judge is not a generator, a shared embedding is not a world model, and a smaller checkpoint is not automatically faster. The Part ends where the book must: objectives, data, evaluation, and values remain choices no architecture makes for us.