Zac Hatfield-Dodds comments on Taking features out of superposition with sparse autoencoders more quickly with informed initialization

[Zac Hatfield-Dodds]

I’d love to see a polished version of this work posted to the Arxiv.