Tinker Interpretability

by Kondasamy Jayaraman · drawn with Colorfigures by April Zhu

Loading Larkspur mini

All six fuses in.

Chapter 2 of 7

Pull a fuse

Looking didn’t prove anything. The team tries the electrician’s test: switch one part off and see what stops working.

This chapter: What changes when one head is switched off, on one prompt and over many?

About this chapter

You will learn

  • What ablation is: silencing one part and watching the guess
  • Why one prompt can mislead, and how text-wide counts help
  • That in Larkspur mini no single head carries the work

Words you’ll meet

Ablation
Switching one part of a model off (here, zeroing a head’s output) and watching what changes.
Fuse
This module’s switch for one part of Larkspur mini: one per head and one per MLP. Pulled, the part is off.
Held-out text
Larkspur documents kept out of training, so a score on them isn’t just the model remembering.

Step 1 of 5

tests a part: switch it off and watch the guess. Layer 1, head 1 finds where words start.

Pull Layer 1, head 1’s .

Loading Larkspur mini