fine tuning , sae scalpels stethescope

1. how is full LoRA worse than subset tuned layers selected
2. that too, they fine tune full weights, and only update the ones that are selected by SAE. and rest of them are as it is. how is it even working ? like should training only some of them work?