Can we gain insights about induction heads without sampling from the underlying LLM? In this writeup, we explore this question with help of Gemma-2 2B and Gemma Scope SAEs trained on the residual stream.