The one-sentence takeaway

LoRA fine-tunes a giant model cheaply by freezing the original and training only a tiny pair of low-rank matrices that capture the small adjustment a new task requires, which slashes memory and storage, lets one base model wear many swappable adapters, and adds no slowdown when the model runs.

Next: Chapter 5, Mixtral and Mixture of Experts, where models get a bigger brain without getting slower.