TinkerLoading Larkspur mini
Every enterprise account has a named customer success … · the clean run
Chapter 4 of 7
Patching
Swap one word and Larkspur mini’s answer changes. The team wants to know where, inside the run, that word’s effect travels.
This chapter: How do you find where a piece of information sits and where it moves?
About this chapter
You will learn
- What activation patching is: copying one activation from another run
- How a grid of patches shows where information travels
Words you’ll meet
- Activation
- The numbers a part of the model computes for one token on one run, such as the residual stream at one layer.
- Activation patching
- Copying one activation from a run on another prompt into this run, to see whether the answer follows it.
Step 1 of 5
Swap “customer” for “Kestrel” and the guess flips from “manager” to “team”. Where does that change travel?
Run both prompts.
Loading Larkspur mini