20. Refine large models with smaller updates
Use adapters, LoRA, prompt tuning, and other parameter-efficient methods to change large models without updating every weight. This chapter covers the tradeoffs in memory, cost, quality, and maintainability.