Tinker Interpretability

by Kondasamy Jayaraman · drawn with Colorfigures by April Zhu

Loading Larkspur mini

We reply within five business … · 6 of 6 fuses in

Chapter 7 of 7

The smallest set

The bench is yours: six fuses, every behaviour live. The reviewer’s last question: how little of the model does one answer need?

This chapter: What is the smallest set of parts that still gives the answer?

About this chapter

You will learn

  • How to search for the parts a behaviour needs
  • Why some behaviours need almost all of them

Words you’ll meet

Circuit
The smallest set of a model’s parts that still produces one behaviour when everything else is switched off.

Step 1 of 3

A is the smallest set of parts that still gives an answer: here “days” after “We reply within five business”.

Pull all you can.

Loading Larkspur mini