Comment by yuwenlu on Case Study: Interpreting, Manipulating, and Controlling CLIP With Sparse Autoencoders · 2025-02-15T03:33:59.282Z · LW · GW

Hey! Late to the party but this is *really* cool. 

A quick question: any reason to use CLIP embeddings as the SAE input, instead of directly using the images themselves? I understand that the goal is to understand CLIP inner workings, but curious if you have intuitions on whether directly feeding in images would work as well.