Tinker Interpretability

by Samy (Kondasamy Jayaraman) · drawn with Colorfigures by April Zhu

Loading Larkspur mini

We reply within five business … · threads from the last token, layer 1 head 1

Chapter 1 of 7

Why did it say that?

Larkspur’s team watched Larkspur mini work in Transformers. Now a reviewer asks why it says what it says, and an attention map won’t settle it.

This chapter: What can an attention map tell you about why a model answered, and what can’t it?

About this chapter

You will learn

  • What interpretability asks: which parts made this guess
  • Why an attention map describes where a head looks, not what it caused

Words you’ll meet

Interpretability
Finding out how a model reaches its output from what happens inside it: which parts, doing what.
Attention map
A picture of one head’s weights: how much each token looks at each earlier one. Transformers drew them as threads and a heatmap.

Step 1 of 5

Larkspur mini says “days” next. asks why: which parts inside made that guess?

Show its top 5.

Loading Larkspur mini