Профиль

🫨

ah, thanks, arXiv version now shows the most updated version so you can refer to that: arxiv.org/abs/2506.23845

Position: Use Sparse Autoencoders to Discover UnknownsWhile sparse autoencoders (SAEs) have generated significant excitement, a series of negative results have added to skepticism about their usefulness. Here, we establish a conceptual distinction that r...arxiv.org

Here's the arXiv link: arxiv.org/abs/2506.23845

Use Sparse Autoencoders to Discover Unknown Concepts, Not to Act on Known ConceptsWhile sparse autoencoders (SAEs) have generated significant excitement, a series of negative results have added to skepticism about their usefulness. Here, we establish a conceptual distinction that r...arxiv.org

openreview.net/pdf?id=x1Px7... (the arXiv draft is slightly outdated but the update should be posted tomorrow)

This work was co-led by @rajmovva.bsky.social , and with Jon Kleinberg, @emmapierson.bsky.social , and @nkgarg.bsky.social Stop by our poster to chat!

Check our @rajmovva.bsky.social 's earlier thread for details! One difference from our previous draft: we leave open the possibility that SAEs can also be a useful tool for acting on knowns (there is new evidence for this) bsky.app/profile/rajm...

I enjoyed making some ruler-and-compass constructions of some of the mentioned triangle centers. CCW from bottom left: the centroid, 1st Napoleon point, Feuerbach point, circumcenter, Exeter point, 2nd Napoleon point, and incenter.

I wrote the essay in 2020, for a creative non-fiction class taught by John McPhee, whose editing I benefited enormously from. I've published it here essentially unchanged (with a short addendum including an update from Clark Kimberling).

Thanks, that’s very nice! And glad you like the idea.

Better than I 🤷‍♂️ I got from SPEAKER to ZERO in 15 steps in today's Traversle puzzle! Play here: traversle.io

Traversle | Find the path between wordsStart from one word, choose a connected neighbor each turn, and try to reach the target before you run out of steps.traversle.io

More broadly, we've been interested in using the internal representations of language models to build information structures that allow human navigation and agency.

This was led and built by Nicholas Goguen-Compagnoni, an MEng student at @cornelltech.bsky.social. This is part of some ongoing work we're doing with Jon Kleinberg and @nkgarg.bsky.social

Thanks! Vannevar Bush’s essay As We May Think is our main source for “trails” web.mit.edu/sts.035/www/...

So you can craft your feed as you like, and then jump onto trails if and when you choose. We see trails as an anti-paternalistic vision. Would be curious to hear what you think!

Thanks -- interestingly, this sentiment is part of our motivation. Social media, I think, has two functions: 1) help you see what you know you want, 2) connect you to things you didn't realize you wanted. Rather than try to solve both within the feed, we want to let users do (2) by choice.

glad you're enjoying!

I also think there could be a way for authors to directly select/unselect trails! Thanks for the questions, they're very generative/interesting to think through

I think very similar to hashtags, but I think hashtags are difficult from a coordination perspective (certain hashtags arise for certain topics on Bluesky, but I think this is a fraction of possible trails that could exist—basically, we could augment the hashtag functionality)

Re visibility, one interface we've played with lets the user to hover on a tag to preview a trail. Definitely would be curious about other ideas here.

Definitely see as complementarity to directed queries on the user's part (one thing we're trying to solve though is when users don't know a topic/trail exists; i.e., users may struggle to articulate destinations they don't yet know of)