Hacker News Reader

If you're able to find a feature, is it possible to selectively replace it to optimize it?

Kind of like replacing a portion of unoptimized compiler code with hand written assembly?

It's a super interesting direction! That's one of the long term goals of interp research: deconstruct model behavior into circuits of features, and then turn those circuits into code (that we can maybe even formally verify!).

Reply on news.ycombinator.com