3 ms·
I have skepticism regarding the 'completeness' of SAE in comprehensive discovery of features: https://www.lesswrong.com/posts/BduCMgmjJnCtc7jKc/research-report
by kromem 2y ago
I have skepticism regarding the 'completeness' of SAE in comprehensive discovery of features:
https://www.lesswrong.com/posts/BduCMgmjJnCtc7jKc/research-report-sparse-autoencoders-find-only-9-180-board https://www.lesswrong.com/posts/BduCMgmjJnCtc7jKc/research-r...
- jengels_ 2y agoSure, but completeness is a much higher bar than being able to find at least some things we weren’t looking for. And I’m reasonably optimistic that we’re going to make SAEs much better in the future, I agree they’re definitely imperfect right now