Comment

Comment by krackers

You don't even need SAE for residual stream based steering, you can do so much more naively by just taking the average difference in activations of prompted vs. unprompted versions, then PCA to get a "control vector" that you can use to elicit the prompted behavior

https://vgel.me/posts/representation-engineering/

parent krackers 3 days ago

No comments yet

Be the first to share your thoughts.

Live from the official Hacker News API

A modern reader for Hacker News · SvelteKit + Skeleton