Interesting research from Anthropic. I'm looking forward to reading follow-on work, and I really hope that this will be tested on open source models (like Mistral) to confirm the method.
Anthropic Researchers Map Features in Claude 3 Sonnet
https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html
0 Comments
sjh@sh.itjust.works · 1 pts · 1y