Amazing. All right. Hello everyone. My name is Sophie. And today I'm going to be presenting my research with my advisor, Nikil Garg, studying social media through the Atmosphere. And I'm so excited to be here today and meet everyone whose presences I'm recognized from online. Cool to see you all in person. Yeah, so just a brief kind of background. Nikil and I are researchers at Cornell Tech, and our research area is studying algorithms and society. So kind of the social aspects of recommendation systems and feeds, understanding dynamics and how to improve systems along social axes, understanding social media as attention markets, and studying human AI collaboration.
Today in our talk, I'm going to be talking first about our work aiding in studying scientific communication on academic Bluesky. Next, I'll talk about what we believe is the potential for studying novel feed experiments through the custom feed ecosystem, as well as novel interfaces through the app protocol. So I'll start with Academic Bluesky. So the backstory for this project is really that Nikil and I had been doing a project researching fairness and recommendation. It was a theory project, which is my kind of original research area. And we had worked with a master student who built a recommender for archive preference, and we used this to simulate kind of market outcomes in this like recommendation system.
And we were curious what happens in real data. How can we study the fairness outcomes in attention markets on a real deployed system? How do we experiment? And what this really looks like in the literature is either you run a lab experiment, you build some sort of simulated simulated recommender, you get people to interact with it and see their behavior. Another option is to deploy browser extensions. So on Twitter, you get someone to download your extension and engage with Twitter, and you will like manually, or you will academically re-rank the posts that Twitter is returning. And this is also frictionful because you have to recruit participants and get them to download the extension and engage with engage with it.
And then the final option is to collaborate with a platform. But these are kind of few, these opportunities are few and far between. So suddenly, November 2024, many academics were using Bluesky, and in particular looking for a personalized academic experience, which is something that we were really excited about building. And so Nikhila had this idea let's build a custom feed for academics on Bluesky. And so we had a little hackathon and built Paper Skygest, which is a custom feed for academics showing posts about papers from accounts you're following. So it's very simple algorithm. And our objective here is to both support academic research discovery, support academics on Bluesky, but also in terms of our research, enable novel experiments to study open research questions.
So yeah, we launched last March, a year ago, and we've seen like kind of sustained usage over that time. And also kind of very positive testimonials from people saying that this has really improved their experience using Bluesky as academics and being able to see filtered academic content in one place that's personalized to them. And with the data that we've collected from Paper SkyGest, we can look at interesting questions here. So one plot from our paper is looking at a histogram of archive categories across archive posts on Bluesky that aren't associated with a bot account, which we've we've labeled.
And so this is sort of the distribution across all posts, but then you can ask what is the distribution of posts that we sent to users on paper skygest when they requested the feed. And you can see that it's a bit like higher on the AI and LG is like machine learning as well as like computer vision and linguistics. And then similarly the posts that people interact with, you can also look at like what are people interacting with in this content on Skydos that we've returned. So that's just an example of like one of the figures we have in our paper and these kind of questions that we can now look at with this data of people using our feed.
And there's like a variety of kind of spin-off projects that we're working on here. One being kind of like looking at uh doing um social science analyses of what people are talking about and working on on Bluesky and leveraging emerging NLP tools like sparse autoencoders to do sort of topic analyses. We're trying to extend paper SkyGest to include trending content and more sophisticated recommendation to a discovery and discourse. And also running experiments on paper SkyGest to A B test different algorithms and also answer interesting science of science questions. And this leads to kind of the novel feed experience detection of the talk.
So we've been piloting experiments on blue on paper SkyGist, particularly, you know, including reposts versus not including reposts. But one thing that I'm really excited to talk about today is some research that we've been doing. Oh, this is out of order, but yes, we did have the pilot experiment which kind of shows some weak results. But excited to talk about this project we're doing in collaboration with Andra who runs the trending newsfeed and Grace, who hosts the trending newsfeed. So this is a really cool feed that includes posts sourced from around 300 verified news organizations and has around 10,000 daily users.
And we're excited to answer some interesting research questions. So there's this exciting paper from 2006, commonly known as like the Music Lab experiment, which studies things about rich get richer effects, which I think got mentioned on the panel earlier today, which basically shows that kind of like little random disturbances in what content gets exposure early on in the process can kind of result in these drastically different long-term outcomes and change like whether posts get exposed according really in terms of like are the highest quality posts being shown. So we want to understand, yeah, to what extent does initial randomness and account size produce these rich get richer effects in a real social setting because this is like a lab experiment.
And also, how can we solve this? Like how can we mitigate these effects and try to get posts being shown that are truly the quality content that people want to see and kind of smooth out this randomness. And kind of the bigger picture here that we're excited about, like custom feed experiments on Bluesky, is that yeah, there's kind of a bunch of new directions for feed experiments. There's new stakeholders, feed creators who are community members who are passionate about showing their community the content that they want to see and potentially those algorithms that are good for different communities are you know yeah, heterogeneous and different communities want to see different things.
As kind of like with this music lab uh extension that we're thinking about. It's an opportunity to ask like these age-old social media questions in new settings and um in deployment. Um yeah, um, as I was saying, new questions about how this is heterogeneous across communities. And finally, I want to talk about um novel interfaces through the app protocol. Um so there's a couple things that we've been working on here. In particular, we've been thinking about how can we kind of allow people customization experiment uh experience on their paper skygest as well as this project led by Kenny Pung from my lab on kind of extending outside of the linear feed structure to trails and these post-linear feeds.
So the customization interface here. Um custom feeds gives user algorithmic choice, and there's kind of been a lot of interest, I think, in academia and industry in sort of giving users more agency over their algorithms and their online experiences. Um but this leads to a lot of research questions, such as like what how much granularity do users want? You know, are they is the feed, you know, choosing between different feeds that are created by other people, is that kind of the sweet spot, or do people want like more fine grain control? Um, what types of controls do they like?
Do they like to control like who they're looking at, or do they want to see kind of different mixtures? Um, and is there some sort of consensus about what's good for people, or are there different kind of types of users? Um to that end, we built this uh customization interface, which for the sake of smoothness, I will probably just show the screenshots instead of the actual demo. Um, but find me after I'll show show the demo. Um but basically this like configuration editor, which is similar to um, I think uh tools that others have developed. Um but basically on the one hand, you can kind of tune your your Skygex experiment and it'll like responsively show you a preview of what your skygist just will look like.
So one thing we're really excited about here is building good kind of defaults for people to use and then seeing which defaults people gravitate toward, especially kind of like what do people want and what do people end up using here. But then we hope to give sort of more flexibility, including like these follow weights. So like can you talk, uh, can you kind of put different weights or like inclusion on your followers. So for example, if I'm following a bunch of academics, but I want my paper sky just to mostly be specifically the like econ CS academics, then I could like filter down to that.
Um and then different weights, both on in terms of like global likes as well as uh interactions from people I'm following. Um then, oh yeah, this is just showing the time decay uh does things. Um then uh and then also uh filtering down to specific categories of content. So I've here filtered the AI category and um uh these are like trending things uh in the last uh you know 72 hours in in uh in AI archive posts, and um yeah you can imagine swapping these out for like sparse autoencoders or other um other uh topics, like I think um I think the Chive talk had a bunch of like topics and stuff that we could maybe integrate with.
Um then how am I doing for time? Sorry. Two minutes. Two minutes? Okay, amazing. I have time for this. So yeah, so this um last thing is is uh project I'm really excited about that is led by Kenny Pung uh at Cornell Tech. Um and this is uh yeah um yeah just a really cool project. Uh so we all know that algorithmic feeds are flawed in many ways. So um there are filter bubbles, there's Doom scrolling. Um people feel trapped uh often in their in their uh um by their algorithms. And there's kind of extensive research on improving feed algorithms and understanding the ways in which they fail and you know improving fairness, improving diversity.
Um but Kenny's idea is why do we have to be you know restricting ourselves to this linear feed interface. Um so the original ver uh vision of the web was to browse hypertext, like that's where HTTP comes from. And this goes all the way back to like the 1950s, the vision of Van Ever Bush of like um uh you know navigating trails of information and how can we um you know switch between um different topics and make connections. Um and there's a connection here to the work that Semble's doing, um, also building trails. Um and so uh we'll assemble like uh assembles trails using like human curation.
Kenny's idea um uh is is building trails for interacting with social media through uh through new um LLM technology to basically um generate these trails algorithmically. Um he uses sparse autoencoders, which basically generate interpretable labels for different um pieces of content. Um and excitingly, he implemented this on like a week's worth of Bluesky data. Um so you should all check this out. I'll show a demo in a second, but if you want to pull it up, um this is called SkyTrails, and this is um his uh basically uh Bluesky browser. Um if you open SkyTrails, you can see a uh selection, a random selection of the trails that are available.
Um I'm just gonna click into the PG Woodhouse and Jeeves, and here are a bunch of posts related to this topic. And as I'm scrolling, I see, okay, now I'm interested in discussions about ChatGPT, so I click on this trail, and I'm led to the trail about ChatGBT. I can scroll down here, and then oh, this is really interesting and esoteric is seeking or suggesting alternatives. I'm now on this trail, I can scroll down. Oh, stylish and fashionable aesthetics. I scroll here. And so now you can see how I've really navigated the space of content. I've found things that are kind of cross-cutting these concepts that you maybe wouldn't have like expected even are being discussed, or like certainly there's not going to be a custom feed for like you know, what was the last one?
Like seeking or suggesting alternatives. Um, but maybe it's something I'm interested in. And so there's this uh this um ability to like quickly navigate the world of content, find new interests, find new discussions that people are having. Um and so that's the the idea of postlinear feeds. Um so yeah, with that, um thank you to our wonderful set of collaborators, um, including folks in this room, um like Andra. And um yeah, uh thank thank you for listening. And I also want to shout out the STEC labeler from the Social Technologies Lab, which is the account activity labeler, which I think is quite popular and very effective, and I'm a big fan.
So that's from another group at Cornell Tech. So thank you, everyone. Amazing. Thank you, Sophie. Questions, anyone. It's not really a question, sorry. But six, six, six, six, six, six, six, like the I I like the the sky trails. I like I like the the lateral movement through topic space, uh, the memex like um interface. Yeah. Awesome. You mentioned earlier the um possibility of doing science of science using the data uh from your sky just uh tool. I'm curious, are you are you actively working in that area? What do you have plan? Because as someone who's just dipping his toes into Scientometrics, you know, this is very exciting, and I'm wondering what you're thinking of.
Yeah, thank you. I think our interest is like um kind of science of academic communication specifically. So understanding, I think our angle will be like looking at the um these like feed uh like the the trending algorithm or like the different algorithmic choices we make on paper sky just and how those choices, especially kind of interpretable um design decisions impact kind of discourse and engagement and you know who's following who and stuff like that. Um so that's I think the angle that we're interested in, uh especially as well as these kind of like more observational questions of like across Bluesky, like as we're ingesting the fires, what are we seeing about like what people are talking about and um engaging with.
Yeah. Great talk, thanks. Um with custom feeds in the network and kind it and incentives is so you can do these studies where you're like how does it imp impact individuals' behaviors through what feeds they see. But a lot of what feeds are doing is they're setting. They're kind of like the rules of the game, and they end up, you know, certain kinds of content if they're getting a lot of traction or visibility or people are interacting, even like liking or following, right? Getting these social interactions more than just like uh social media interactions that will incentivize people to create different kinds of content and get get this big loop going.
Um so you change what kinds of content is available through that. But that's hard to do with the exciting thing with with Bluesky the app and the protocol is that we can do these people can create new things and do experiments pretty easily for individuals, but it's harder for groups. Have you do you have any ideas on how to suss that out or try to get kind of like signal through like especially if it's only impacting a smaller part of the overall graph or network. Yeah, that's a really interesting question. I guess there's two aspects to my answer though.
The first is that like this is sort of one of the things we're excited about, like with the trending news feed, um, is looking at so we did a pilot experiment like um basically putting a lot of the re uh the weight um on the grade, these graze multipliers on the reposts and seeing how that changed like what content was so we ended up seeing a lot more like um smaller news organizations were um rising at the top, and and you ended up seeing like sort of an amplification of like you you created uh higher weight on reposts and then you saw a lot more reposts um relative to the other feed.
But I think maybe your question is more just like okay, the trending news feed is sort of an outlier in terms of like feed size, and so how can we actually study um like uh how these feeds are are impacting sort of things on a bigger scale when they're small. And I have not actually thought of that question, and I'll think more about it. Thank you. Hi, so apologies in advance if my question is just like completely out of nowhere because I'm really sorry I joined your talk late. Um but I'm wondering, like, especially looking at the kind of like trails of conversations and topics that you know users and people are kind of posting about.
I'm wondering if you've done any exploration.