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