Tinker Interpretability

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

Loading Larkspur mini

We reply within five business … · “days” marked in the Lens tab

Chapter 3 of 7

The guess forming

Fuses showed which parts matter. Now the team wants to see the guess itself take shape, layer by layer, through Larkspur mini’s 2 blocks.

This chapter: Where inside the model does the answer first appear?

About this chapter

You will learn

  • How the logit lens reads a guess partway up
  • What the model guesses before any block has run
  • Which block makes the guess in Larkspur mini

Words you’ll meet

Block
One layer of the model: its attention heads, then its MLP, each adding to the residual stream.
Logit lens
nostalgebraist's logit lens (2020) reads a model's middle layers through its own final output step, and finds a fair guess often forms partway up and is refined after.

Step 1 of 5

The reads the guess partway up, as if the model stopped there and answered.

Open the Lens tab.

Loading Larkspur mini