Connor Kissane

SAEs are highly dataset de­pen­dent: a case study on the re­fusal direction

Open Source Repli­ca­tion of An­thropic’s Cross­coder pa­per for model-diffing

Base LLMs re­fuse too

SAEs (usu­ally) Trans­fer Between Base and Chat Models

At­ten­tion Out­put SAEs Im­prove Cir­cuit Analysis

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