Not All Language Model Features Are Linear

Re:Not All Language Model Features Are Linear by kromem on Friday May 24, 2024 @11:24PM (#40472424)

One of the questions I've been thinking about a lot looking at the past year of interpretability research is just how much of what we are finding is "what we're attuned to find" as opposed to "what's actually there."

Are we only measuring the tip of the iceberg, and have coalesced towards getting better at iceberg tip measuring?

[ Parent | on HN ]

Re:Not All Language Model Features Are Linear by jengels_ on Friday May 24, 2024 @11:40PM (#40472477)

I feel like un-supurvised methods like Anthropic's SAEs can be argued to find things we're not looking for (their most recent work is from a couple days ago: https://transformer-circuits.pub/2024/scaling-monosemanticit...). And we can get some sense of how "much" of the model they're recovering by looking at their downstream reconstruction loss.

[ Parent | on HN ]

Re:Not All Language Model Features Are Linear by kromem on Saturday May 25, 2024 @01:11AM (#40472800)

I have skepticism regarding the 'completeness' of SAE in comprehensive discovery of features:

https://www.lesswrong.com/posts/BduCMgmjJnCtc7jKc/research-r...

[ Parent | on HN ]

Re:Not All Language Model Features Are Linear by jengels_ on Saturday May 25, 2024 @01:55AM (#40472918)

Sure, 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

[ Parent | on HN ]