Yeah, let's take this off. Okay. So hello everyone. I'm Leijie. I'm a PhD student from University of Washington. Like, even though I'm L Seattle, very close to Vancouver, but because of the visa issues, I cannot come in person, so that's why I'm over the Zoom. But today I would like to glad to present our work about Bluesky custom feeds and titled Understand the Landscape of Custom Feeds on Bluesky. Cool. I'll start with thinking about the platforms you use every day, like YouTube, TikTok, Instagram, or Spotify. Each of them has a recommendation feeds that feels quite personal to you.
But you have no saying how that particular algorithm works or what even optimize for. Like whether you optimize for likes or dislikes, the ranking is just instead decided by those single centralized algorithms controlled by those big companies. And this concentration of power in these days has two consequences. So first, you they have no visibility into why certain content appeal into their feed. The algorithm is just a black box. And we see a lot of complaints about user about them in these days. And second, creators often learn to optimize for that particular algorithm to appease them. And over time we can see the content start to become homogenous everywhere.
Like they are all kinds of clickbait kind of content. And this is what Bluesky does something different here. Like they call algorithmic choice. Bluesky custom feeds allow users to create, subscribe, and share their feed algorithms instead of being stuck without a particular default. And on the consumer side, the field consumers, they can further subscribe into the feeds offered by their communities. And on the creator side, third-party tools make it possible for anyone to build and publish their own feeds. This connects to a very long-standing vision in the research community, the idea of middleware. The core idea is simple.
What if the social media services like recommendation, moderation, or the governance in the communities did not have to be controlled by the platforms? What if they could be provided by this independent middleware services so that you don't have the choice among them? And this is just a theoretical vision, it's already happening at scale. Today, as people already know in the room, like the Bluesky today has 42 million active users, and over 40,000 custom feeds have been created. So this makes custom feeds as the first large-scale instantiation of the middleware version for feed recommendations specifically. But for this particular ecosystem to work, someone has to actually build in their feeds.
And that's the feed creators we focus in this particular work. But we know little about them. Like what motivates them to creating feeds in the first place? And what does it actually take to run a particular custom feed? And what challenges and opportunities do they face? These are the questions driving our work. So in particular, we organize our work around these three research questions. First, what kinds of custom feeds are there? Second, what kinds of how do the feed creators building, maintaining and sustaining their custom feeds ecosystem. And third, what kinds of challenges and opportunities exist for the custom feed ecosystem.
To answer this research questions, we conducted semi-structured interviews with 24 feed creators on Bluesky. And this particular chart shows a range of feeds they created, measured by the likes they received on Bluesky. Some created they are relatively small with under like 100 likes, but there are others who building feeds over 10,000 likes. So we have a good spread in our interviews from the hobbyist builders to some of the most popular feeds on the platform. And here we talk about our findings.
The first is a utility providing feed. They are focused on delivering the content to feed consumers, thinking of them as a well-curated feeder filter. And there's also community building feeds that are very different. They're cultivating the space for the feed consumers to connect with each other. And as we will see, these two orientations shape their motivations, their daily tasks, and how they sustain their feeds in the long term. Specifically for the utility providing feeds, the feed is primarily a content delivery mechanisms for the feed consumers. The creator's job is mostly done as setup. They configure the hashtags, keywords, or that accounts, and just let the feed run on the Bluesky.
On the other hand, the feed consumers are purely receptive or passive. They just follow the feed like they're following a particular channel for a particular block site. They do not actively contribute to what particular content is appears in the feed by themselves. And within the utility feed, we also see two distinct subtypes. The first is what we call setted and forgetted mode. People just cache the relevant post of a given topic just with a very simple list of keywords, such as mushrooms or some fan communities. The creator just sets it up once and just largely forgot it and step away.
The second type is very highly curated feeds. Those use vetted poster lists or custom ranking algorithms to ensure the content they delivered are super high quality. It's almost like an editor's choice or editors picking the newspapers. Unlike the first type, they require very ongoing maintenance. The creators constantly turning and up tuning and updating their algorithms to ensure the content is always always high quality. Community building feeds work very differently from these utility feeds. Here, the field consumers are not just passive recipients. They actively contribute in the content and interact with each other simply by the feed.
Like they were adding a particular keyword to their post so that they're essentially having conversations with each other in the feed. The feed becomes a very shared social place, and the feed creator's role shifts here too. Instead of configuring rules, they tend to the social space, maintaining norms and keeping the community healthy. I will give you two concrete examples of community building feeds. So there's a particular feed we found out in our interviews. They just surface the post from the users below a particular threshold of follower numbers so that they can design it to welcome the newcomers and give the less known voices a chance to be seen in their feed.
And another just like the feed consumers curating the content by themselves. So they can just reply into any post on Bluesky with a particular keyword. And in this way they are nominating the post to appear in the feed automatically. And one of our participants captured the creator's role very well. They said, I almost see myself as a librarian. I'm making sure people are not yelling, they are not picking up their charge and putting books back when they are supposed to be. I'm making the place as preventable and welcoming as possible. So they are just tending to that particular social space as a feed creator, which is very different from the utility feed creators.
And after that, let's take a look about how people building their custom feeds. And we found out that most feed creators navigate the ecosystem of third-party tools. They do not have to start from scratch. But for some of them, those tools are not enough for their needs. So they end up building their own infrastructures. So for example, most feed creators they rely on third-party tools, and there's a real diversity of options for them. Brown Sky Feed, Bluesky Feed Creators, and Grays, just to name a few. So the variety matters, it means that the creators can always choose the tool that best fits their needs and their technical levels.
Within these tools, creators try to combine various kinds of building blocks to shape what appears in the feed. Some using this keywords or regular expression matching to catch a post on specific topic. Others try to maintain the vetted account list, hand picking who got included in their feed. And some also use a Bluesky labeler system to help filter the content as part of the content moderation efforts. But the third party do not work for everyone. Some creators have the needs that go beyond what existing tools can support, and they end up building their own infrastructures.
So for instance, one participant build a fully personalized recommendation system on their own because they need a fine-grained interaction data that this third-party tool simply cannot access. And another try to write a custom implementation in Rust because their field logic is often too complex for any block-based tools offered by these third-party tools, and they also need to support much larger scale of use. So in this way, these feed creators are not casual hobbyists who just try it for fun. They are essentially running some independent technical infrastructure through the custom feeds. And our third key finding is about what it takes to maintain a custom feed over time.
And it turned out the maintenance goes well beyond just keeping the algorithm running. Creators are handling moderation, driving discovery, and try to listen into the consumer feedback. And for most of these tasks, they get very little support from the Bluesky platform. So for instance, the utility providing feed, the commun the moderation effort is very different across these different kinds of feeds. But for the utility providing feeds, it's more about larger upfront effort. They set up some strict rules to deliver the content, which means that usually there isn't much toxic content to moderate for after the feed go alive.
But for the community feed, it's often ongoing because the community itself is always evolving. There's new member joins, there's new dynamic emerge, and norms need to be actively maintained. And on top of that, the moderation decision is also much more nuanced. It's not just about the topic being misaligned with their feed topic, but instead it's a social decision. They need to carefully make sure their decision does not hurt the community relationship. And the Bluesky also decentralized the creation of the feed, but they do not decentralize the discovery of the feed. The feed top on the Bluesky is often static, ranked by the popularities, serving the same set of feed that were popular two years ago.
And for the new feeds, especially non-English ones, they have little chance to be seen by the newcomers. And the creators need to take matters into their own time. They have to promote the feeds by themselves on the social media in the communities whenever they can. And we heard this frustrations directed from the creators. One participant told us I have been commenting on every single feed topic post I found, telling them about the feeds and how to set up the Bluesky accounts. And they've sent out thousands of messages. And they concluded that even though the feed is successful, but obviously it's not a sustainable strategy for a lot of feed creators.
And another participant complained that if creating a new account and go to the feed tab, you'll see the same top feed from two years ago. They are all in English and often only about the United States. So it's not interesting for the people outside of the United States. And the centralized algorithms can improve the continuous on these YouTube or TikTok platforms. They listen into your clicks, your skips, and the dwell times and feedback into their recommendation algorithms. But in the Bluesky for the custom feeds, they have not that kind of infrastructure. There are two several kinds of feedback they can tap into.
They get some implicit feedback from the people who likes replies or reports, their feeds from their custom feeds. But they lack the technical tools to actually analyze and learn from this data at scale. They also got a lot of some sort of explicit feedback from their direct messages or some server responses they distributed in their feed. But it's often very rare and only comes from the people who are most motivated to consume their feed content. There are also some third-party tools provide some basic analytics, things like the daily active users or the post volumes. But this only tell the feed creators that the feed is still alive, but not whether it's actually working or not.
The result is that the feed creators do not get a lot of feedback to continue improving their feed. And in the long term, they just lose the motivation to do that. And our fourth and final key finding is about sustaining the custom feeds. The custom feed ecosystem, over 40,000 feeds run almost entirely of volunteer labels. Creators need to absorb the cost of running the feed themselves with very little to no platform support. And what is sustain them isn't their financial incentives or the platform recognitions. It's often just intrinsic motivation. And if we take a step back, there's some cost of sustaining a feed.
And on the infrastructure side, the creators need to pay out of the pocket for third-party subscriptions to keep their feed running. And beyond the money, there's also time commitment. They have to do the regular maintenance, moderation, and responding to user feedback. But what keeps the creator going, we call it the hobbyist model. The motivation is the only feed. For the utility providing creators, seeing their subscriber account grow is a key signal. It tells them the feed is still useful and worth maintaining. But that also saying that without that particular signal, feed creators are more likely to lose motivations.
And that's why we observe there's a bunch of certain and forgetted feeds for the utility part. For the community building creators, the motivation comes from something very different. They are seeing the people actually connecting with each other through their feed. And that makes them happy about their job. But this mirrors what we see in those volunteer ecosystems like Reddit and Discord, where the community itself is just the reward. We also asked about feed creators about monetarization. And the most felt that not all the feeds can justify it. The key factor is that where the feed is built on dedicated, irreplaceable effort.
So for example, the set and forget feeds, they are very easy to replicate. If someone charges for it, users can just build the feeds by themselves elsewhere. But for the highly curative feeds, they are very different. The creators have the ongoing effort and judgment to make the feed valuable to their feed consumers. And we do see some successful examples of monitorization in that space in our interviews. Creators volunteer their efforts very willingly, but they also think the infrastructure costs deserve some financial support. One participant very simply that it's not about getting rich. It's just about getting covering expense.
So I'm not going to pay out of the pocket to help the community. And another frame and even more positive, like they charging a small amount of dollars isn't just fair to the creators. It's creating a positive loop of dollars going through the communities so that the whole Bluesky custom feeds can go more go more alive. So with all these findings, we have these two design implications to see how we can further improve the Bluesky custom feeds as a community and also as researchers. We see there's two key directions to better support feed creators beyond the initial building process.
First, we need to help creators to better make sense of the implicit feedback they get from their feed creators. Feed consumers like likes, reposs, and replies exist. But creators need to better choose, to better turn into that noisy data into some actionable insights they can use to improve their feeds. Second, we also need to rethink about the feed discovery. Right now it's still platform controlled, even though it's very fair control, like it's always ranked by the popularities. But enabling more personalized feed discovery beyond just the flat platform control ranking will give new users and not English feed a real chance to reach to the real right users.
And we also want to say something about the sustainability. If creators burn out and cannot cover their cost, feeds just will stop being maintained, and the communities building around them will slowly go away. From our interviews, we found two potential passes that the feed creators themselves are already experienced with. One is that donations, they will solicit some donations from their feed consumers. This is similar to how the Wikipedia sustain itself through those volunteer like community donations. And we also see those popular feeds doing transparent advertising. They have some ads within their feeds, but the ads is opting and the community scope to ads so that they can keep the revenue but do not deliver like very irrelevant ads to their consumers.
Although neither of them is a perfect solution, but they both point out towards a model where the community of the custom feeds can support the infrastructure and support itself in the long term. Here's all of our presentations and uh thank you for everyone listening, even though it's over Zoom. And do you have any questions? And here's my contact. Do we have any questions, comments, thoughts? Nope. What'd you got? If you were managing Bluesky the app, what would you do? The question was if you were managing Bluesky the app, what would you do? Um I think as I can see in the design implications, I will at least try to improve the feed discovery feature.
Like I think people do want that part to be like it's inevitably going to be some sort of centralized in the platform by the Bluesky platform to design how to ranking these different kinds of feeds. But obviously, I think one single feature you can see just to recommend relevant feeds for people to explore so that they are not being stuck in those popularity ranking feeds, which is always just the same set of feeds all over the years. And I think another important part people constantly want is about a better guidance about how to do the advertisement in their feeds.
Like people are kind of try to figure it out by themselves and try to test that in their own population, the consumer's populations. But it would be beneficial for them to see like what kinds of ads are tolerable in the feeds so that they can be more actionable and more comfortably place ads in their feeds so that they're not going to do the free label for the whole ecosystem. Great. Well, I think that's it. Thank you very much. Yeah. Thank you everyone. Cheers.