This is a special post for quick takes by Raemon. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
This is an experiment in short-form content on LW2.0. I'll be using the comment section of this post as a repository of short, sometimes-half-baked posts that either:
don't feel ready to be written up as a full post
I think the process of writing them up might make them worse (i.e. longer than they need to be)
I ask people not to create top-level comments here, but feel free to reply to comments like you would a FB post.
I've lately been frustrated with Suno (AI music) and Midjourney, where I get something that has some nice vibes I want, but, then, it's wrong in some way.
Generally, the way these have improved has been via getting better prompting, presumably via straightforwardish training.
Recently, I was finding myself wishing I could get Suno to copy a vibe from one song (which had wrong melodies but correct atmosphere) into a cover of another song with the correct melodies. I found myself wishing for some combination of interpretability/activation steering/something.
Like, go to a particular range of a song, and then have some auto-intepretability-tools pop out major features like "violin" "the key" "vocals style", etc, and then let me somehow insert that into another song.
I'm not sure of the tractability of getting this to be useful enough to be worthwhile. But, if you got over some minimum hump of "it's at least a useful tool to have in combination with normal prompting", you might be able to get into a flywheel of "now, it's easier to funnel commercial dollars into interpretability research and there's a feedbackloop of 'was it actually useful?".
And, doing it for visual or musical art would add a few steps between "improved interpretability" and "directly accelerating capabilities."
This is presumably difficult because:
Interpretability just isn't quite there yet, and, you need a real good team to get it there and also continue making progress
Converting it into a product is a lot of work and skill – you need a particularly good UI and product design team.
This came up because I noticed the recent Suno Studio app was actually quite good, UI wise. (They had some nice little innovations over the usual sound-mixing-app suite), and I bet the Suno could figure out how to get from Chris Olah-like interfaces to something actually usable.
I don't think it even has to be that useful in order to become commercially interesting – it could start as a fun toy that's mostly delightful in how it produces weird shit that doesn't quite make sense.
I'm a researcher at Suno, interpretability and control are things we are very interested in!
In general, I think music is a very challenging, low stakes test bed for alignment approaches. Everyone has wildly varied and specific tastes in music which often can't be described in words. Feedback is relatively more expensive compared to language and images since you need to spend time to listen to the audio.
Any advances in controllability do get released quickly to a eager audience, like Studio. The commercial incentives align well. We're looking for people and ideas to push further in this direction.
This is an experiment in short-form content on LW2.0. I'll be using the comment section of this post as a repository of short, sometimes-half-baked posts that either:
I ask people not to create top-level comments here, but feel free to reply to comments like you would a FB post.