Connor Kissane

SAEs (usu­ally) Trans­fer Between Base and Chat Models

At­ten­tion Out­put SAEs Im­prove Cir­cuit Analysis

• Connor Kissane 17 May 2024 8:53 UTC

  1 point

  1

  in reply to: Ali Shehper’s comment on: Sparse Au­toen­coders Work on At­ten­tion Layer Outputs

  Thanks for the comment! We always use the pre-ReLU feature activation, which is equal to the post-ReLU activation (given that the feature is activate), and is purely linear function of z. Edited the post for clarity.

• Connor Kissane 31 Mar 2024 16:35 UTC

  7 points

  0

  on: SAE-VIS: An­nounce­ment Post

  Amazing! We found your original library super useful for our Attention SAEs research, so thanks for making this!

We In­spected Every Head In GPT-2 Small us­ing SAEs So You Don’t Have To

At­ten­tion SAEs Scale to GPT-2 Small

Sparse Au­toen­coders Work on At­ten­tion Layer Outputs

• Connor Kissane 14 Aug 2023 14:20 UTC

  1 point

  0

  on: Mech In­terp Puz­zle 1: Sus­pi­ciously Similar Embed­dings in GPT-Neo

  These puzzles are great, thanks for making them!

• Connor Kissane 19 Jul 2023 19:57 UTC

  1 point

  0

  on: Causal scrub­bing: re­sults on in­duc­tion heads

  Code for this token filtering can be found in the appendix and the exact token list is linked.

  Maybe I just missed it, but I’m not seeing this. Is the code still available?