0/8 I’m super excited about work done by my LASR scholars David Chanin, Tomáš, Hardik Bhatnagar and James Wilken-Smith. This work demonstrates a critical but likely solvable issue with SAEs! Arxiv link: arxiv.org/abs/2409.14507. Blog post: tinyurl.com/u45waafj

I'm really excited to see cool research begin to be made possible with my team's Gemma Scope Sparse Autoencoders! This paper finds an interesting problem with SAEs, feature absorption: SAEs sometimes learn weird gerrymandered features, like "starts with S, excluding 'seeking'"