Today, I’m talking to Jay Graber, the CEO of Bluesky Social, which is a decentralized competitor to Twitter, er, X. Bluesky actually started inside of what was then known as Twitter — it was a project from then-CEO Jack Dorsey, who spent his days wandering the earth and saying things like Twitter should be a protocol and not a company. Bluesky was supposed to be that protocol, but Jack spun it out of Twitter in 2021, just before Elon Musk bought the company and renamed it X.

Bluesky is now an independent company with a few dozen employees, and it finds itself in the middle of one of the most chaotic moments in the history of social media. There are a lot of companies and ideas competing for space on the post-Twitter internet, and Jay makes a convincing argument that decentralization — the idea that you should be able to take your username and following to different servers as you wish — is the future.

It’s a powerful concept that’s been kicking around for a long time, but now it feels closer to reality than ever before. You’ve heard us talk about it a lot on Decoder: the core idea is that no single company — or individual billionaire — can amass too much power and control over our social networks and the conversations that happen on them.

Bluesky’s approach to this is something called the AT Protocol, which powers Bluesky’s own platform but which is also a technology that anyone can use right now to host their own servers and, eventually, interoperate with a bunch of other networks. You’ll hear Jay explain how building Bluesky the product alongside AT Protocol the protocol has created a cooperate-compete dynamic that runs throughout the entire company and that also informs how it’s building products and features — not only for its own service but also for developers to build on top of. 

Jay and I also talked about the growth of the Bluesky app, which now has more than 5 million users, and how so many of the company’s early decisions around product design and moderation have shaped the type of organic culture that’s taken hold there. Content moderation is, of course, one of the biggest challenges any platform faces, and Bluesky, in particular, has had its fair share of controversies. But the idea behind AT Protocol and Bluesky is devolving control, so Bluesky users can pick their own moderation systems and recommendation algorithms — a grand experiment that I wanted to know much more about.

Finally, Jay and I had the opportunity to get technical and go deeper on standards and protocols, which are the beating heart of the decentralization movement. Bluesky’s AT Protocol is far from the only protocol in the mix — there’s also ActivityPub, which is what powers Mastodon and, soon, Meta’s Threads. There’s been some real animosity between these camps, and I asked Jay about the differences between the two, the benefits of Bluesky’s approach, and how she sees the two coexisting in the future. 

Okay, Bluesky CEO Jay Graber. Here we go.

This transcript has been lightly edited for length and clarity. 

Jay Graber, you’re the CEO of Bluesky Social. Welcome to Decoder.

I am really excited to talk to you. Bluesky is one of the most exciting new social media apps on the scene. There are exciting ideas inside of it around federation and decentralization, both of which I’m obsessed with. So get ready, but let’s start at the very beginning. What is Bluesky Social?

Yeah, Bluesky is an open social network. It’s open-source. It’s built on an open protocol, but it looks and works very much like Twitter. And so it’s a project that was actually originally founded by Twitter back in 2019. Jack [Dorsey] tweeted that he was going to fund the development of a protocol that Twitter would someday run on. And so I came into the project as someone outside of Twitter that had never worked at Twitter but had experience with decentralized social protocols. So I built out a protocol, and then, through a series of changes of what happened to formerly Twitter, now we’ve ended up building an app that looks a lot like Twitter.

I can’t imagine anybody didn’t figure this out. But Jack is Jack Dorsey.

The former CEO of Twitter.

The once former and then CEO and then now former CEO once again of Twitter. That piece where he thought Twitter should have an underlying decentralized protocol, and I don’t want to talk about Twitter too much, but that was a big deal, right? The idea that Twitter was too important, it was too centralized. I think Jack was uncomfortable with his moderation power. 

Obviously, we went through a big moment around content moderation, but that core idea that one company should not have so much power led to a lot of things, including the creation of Bluesky. What was the pitch there? Did he come and say, “Hey, come run a program inside this company to devolve the power of the company”? 

I actually pitched Twitter a lot on how I thought Bluesky should work.

So the way that this worked is that Jack tweeted that, and they created a Bluesky Twitter account at the end of 2019. I got really excited. I was currently working on a social app of my own that was an alternative to Facebook Events called Happening, and I’d been building on decentralized social protocols, playing around with them, doing a lot of research, published a lot of research on it. 

So then I was one of the experts that they pulled into this small chat room with a group of people who seemed to know what they were talking about. And then, they asked the group, “What is the best way to build this protocol? You guys know about protocols.” And then, in 2020, the pandemic hit, and so this whole project got derailed by the pandemic as Twitter was clearly occupied with other things.

And then mid-2021, Twitter circled back around and started interviewing people, both people within Twitter and experts without, interviewing somebody who could lead the Bluesky project. And so, mid-2021, they interviewed me along with some other folks and then chose me to lead Bluesky based on this vision that I pitched for how we were going to build Bluesky, which is essentially how we’ve built it. And so that’s how I got involved. 

Once I took lead, there were different options as to how we were going to structure Bluesky. As a category, a project within Twitter, it was very much like Birdwatch, which was what built Community Notes, sort of a skunkworks project that was going to go off on its own and do this experimental way of doing things in a more open, crowdsourced way. But I didn’t want to be within Twitter because Twitter had moved very slowly to nominate me as lead in the first place.

And there was just going to be a lot of challenges. 

One of the things that I was hyperaware of was just the risks of centralization because one thing I always like to say is the captain can always sink the ship. Jack was the captain of Twitter at that time, but if the captain changes or if that support left, then I would lose institutional support for Bluesky, and I really wanted this project to exist. So I insisted we spin out as a separate company, and then that took about six months to figure out, but then we got set up as a separate company, and then, actually shortly thereafter, the captain did change. And so then we ended up in this position — we were able to keep going. In a lot of other circumstances, if we were within Twitter, I don’t think the project would be alive anymore.

That part is really interesting, right? The idea that the leadership might change, but the project should remain or the protocol should remain. Right now, Bluesky is a company. The audience heard me introduce you as the CEO of Bluesky Social, which is the company. It’s also the app — it’s not the protocol. Explain how that all works. What is Bluesky Social, specifically?

So Bluesky was originally just the name of this project within Twitter to build this decentralized protocol. Then, the protocol we actually built was called the AT Protocol, or AT Proto for short. And this sort of symbolizes the fact that you can find people at their @ handle username. And so we wanted this to someday be the last social identity you’ll ever have to create because you can move it in between apps and services. You can take your identity and your relationships and your data with you. And so that’s the underlying protocol layer. 

The goal is to make social apps work more like the web itself. And then Bluesky is the app that we’ve built on top of the AT Protocol, both to show how it works and to develop the protocol actually through product-driven protocol development. You run into a lot more things once you’re actually building on a thing that you intended to have developers build on.

So Bluesky, the product on top of AT Protocol. I always call it “A.T.” Protocol. Are you good with “A.T.” or “at”?

We’ve been calling it the [at] Protocol because, yeah, it’s easier and sometimes just “AT Proto” for short.

It stands for something, right? 

Authenticated transfer protocol.

Very good. Was that a bacronym? Did you back into that or did you come at it front ways?

I think we actually came at that front ways because, earlier on, we were calling this — there were a lot of posts I wrote early on that are way more technical than the audience we’re usually talking to these days, but stuff about self-certifying data structure and authenticated transport, and these are all the design philosophy that we pulled into Bluesky, which borrowed from peer-to-peer networks but did this hybrid federated approach. So that’s the wonky parts under the hood that most people don’t think about anymore.

I love that. So you’ve got this protocol. You’re going to build a product on top of it called Bluesky. Bluesky is now open to the public. It’s growing. I think the last number I saw was 5 million users. There’s obviously a moment to capture market share from Twitter. I recently saw a report saying 30 percent of Twitter users have stopped using it in the last year. Is your focus on growing Bluesky the product? Is it on the AT Protocol? How do you shift that attention? They seem like they’re competing a little bit.

Yeah, they both cooperate and compete because part of our philosophy was that developers, we want to get developers to build on the AT Protocol, but devs want there to be users to build it for. And so if you get users into the ecosystem, anyone building on either the interfaces for custom feeds or custom clients on Bluesky can come in and now build for 5 million users. But also, AT Protocol will be able to support other kinds of apps, and then users will be able to move between them with the same identity and relationships, and then that will just make it a lot easier for other people to build. And so that was part of our idea was just build an app, and then we get this positive feedback loop going of developers and users coming into the ecosystem.

So that “cooperate and compete” framework is really interesting. Most standards that have a big product on top of the standards have that kind of relationship, but they’re usually managed by giant companies or at least a handful of companies that are competing themselves. There might be a standards board involved. Right now, for Bluesky and AT Protocol, it’s just Bluesky. How is Bluesky structured, and how do you contend with that cooperate-compete dynamic inside the structure?

Yeah, so we are structured as a public benefit corporation, so that means we have a mission. It’s to “develop and drive adoption of large-scale adoption of technologies for the open and public decentralized conversation.” I think I might’ve mangled that a little bit, but the gist of it is for open public conversations on an open protocol, and that’s the mission that all of our stakeholders know — our investors, our board members — and we’re able to pursue that and put any sort of profits back into that mission rather than having to put fiduciary duty first. But that doesn’t mean we’re prohibited from making a profit. So we are a company.

And the company is responsible for the protocol right now. Inside the company, how is that structured?

Yes, it’s within the company. So basically, we’re all working on it together. Everyone’s very mission-aligned that we’ve hired, and it’s something that, right now, we’re just trying to show the benefits of this kind of an open ecosystem that people can build on and get users in the ecosystem, and then we’re moving toward standardizing pieces of the protocol. So pieces that have become really stable, it’s working well. There hasn’t been any need for big changes recently. That’s something that we’re trying to move toward a standardization process because once something gets standardized, then it’s going to move a bit slower, but then that gives it more resilience.

How big is Bluesky today? How many employees are there?

We’re about 18 across engineering and ops, and then we have about that number on support and moderation.

The moderation side is the part that scales the fastest, right? If I look at Bluesky and AT Protocol and what it’s meant to enable, it’s a way to rethink moderation, probably most of all, right? If it’s as simple as “you don’t like the moderators, and you can take your account and leave” — that’s one idea. You have this idea about composable moderation, where people could write their own moderation stacks or different kinds of filters or different kinds of algorithms. Right now, though, moderation just sort of scales with the user base — you just have to spend more time and money moderating. How are you managing that against Bluesky’s growth?

We just hired Aaron Rodericks, former trust and safety co-lead at Twitter and head of election integrity. He’s come on to bring a lot more experienced leadership to this team and make sure that we’re building a strong foundation for moderation on Bluesky. But what you mentioned there — you definitely read about our approach — it’s a hybrid approach between us running a foundation within the app, making sure that we have this really good service within the app, and having decentralized moderation. The closest analogy here — it actually works a lot like custom feeds — is the custom feeds we build. So we provide some default algorithms, and we develop an in-house algorithm, but we also allow anyone to build an algorithm. 

So there’s about 40,000 custom algorithms out there and custom feeds, and some of them are very simple, some of them are very complex, and users can install them and switch between them. So moderation services are like this, but because Bluesky is also a digital space that we’ve created, you can’t opt out of [our moderation service] and our client, and then that’s what we’re providing all this moderation, in-house moderation, for. And then, actually, by the time that this podcast airs, you’ll be able to go into the Bluesky app and see the composable moderation because that’s coming out this week.

I’m really excited about that. I want to come back to that and talk about it in detail. I want to stick on the Decoder question just for a couple more minutes here. You started this thing inside of Twitter. You fought to leave it. That’s a big decision to make. The entire world of social networks, I think, has flipped over since you spun up Bluesky and began work in earnest on this project. It’s an election year. You have big ideas about moderation, you have a team, you’ve got to raise money. That is an awful lot of decisions to make. How do you make decisions? What’s your framework?

I think I have a very collaborative approach, but as with a lot of things I do, I take a hybrid approach between centralization and decentralization.

Perfectly on theme. Very good.

It is. I really do. I spent a long time working on decentralized technologies, and I saw their limitations as well, but I also think that there’s a lot of limitations to centralized systems. It just is a system property. It’s true of human systems and technological systems. So, first of all, I think I make decisions trying to base it around our long-term mission. So, as a public benefit corporation, we really take the mission seriously, and we try to make sure that we are making decisions that are going to build something that’s actually a better social ecosystem for people. And then I believe in collaboration, so I try to get input from as many people on the team as possible — people who have the most relevant experience in the decision-making — and a lot of things start out as proposals that we talk through and then write down and then get feedback on.

And we operate roughly with a consensus with qualifications model. So I get input, and then if there’s a clear consensus, we go with that. And if there isn’t, I’m the tiebreaker and make the final call as CEO. And then we try to expand when we have the capacity outside of our internal team and talk to experts in the field. Even before we brought on Aaron, we talked to a lot of trust and safety experts, and then we’ve consulted with users, community members, people who are using the app, people who aren’t using the app. 

The app itself is sort of a fire hose of feedback, so we have tried to find ways to sample from that and as well as reach out and have deeper conversations with people, talk to people who aren’t on the app, and then this is how we’re trying to make decisions that are technically sound but also user-centric, talking to other people in the decentralized social space, too.

A lot of us come from backgrounds where we’ve worked on other projects and so we reach out to people that we know to talk about “what about this protocol decision or this design decision.” And then people have a lot of autonomy on the team, so we try to move pretty fast. A lot of what we do is about experimenting and learning, so we try to make the best decision we can at the time that leaves us room to adapt down the road. And so that’s actually why a lot of things we built are composable because we aren’t assuming that we’re going to get everything right on the first pass.

We have this big vision for how all these pieces can come together to produce a good social experience, but if we get it wrong, we want the pieces individually to be standalone so that someone else can take these ideas, recombine them in a different way, and then maybe arrive at a solution that, in the next iteration, is the right way for social to work in this open, composable way.

Can I ask a really dumb question? I’ve now said composable. You’ve said composable twice. Can you explain to the audience what you mean by composable?

Sure. So, basically, have something that comes apart into lots of pieces, like building blocks, like Lego, and then you can put those together and you have a lot of, maybe you only have 10 little different types of Lego pieces. You have a lot of them, and then you can put them together into all sorts of different shapes. And so that means that you can build very different-looking houses out of the same set of Lego blocks.

A few more questions on the structure and how that works. You said you’re 18 people on engineering, 18 on the other side of the company — 30 to 40 people, I’m guessing. The way you’re describing decisions works really well for that size team. I know that because I’ve long run a team about that size, and I feel like I can have an idea, I can communicate it effectively.

You jump up to the next size — 100 people or 500 people — you have to build a different kind of culture. Are you thinking about that now, or are you hopeful that you can just get to the next size? I’ve met founders and CEOs sort of who take both approaches — that “I would love to have that problem. I’ll deal with it when I get there.” But it seems like you’re thinking a little down the road and especially that having decentralization in your brain, you might be thinking about that now. How do you plan to scale the culture there?

I’m thinking a bit about that now. As you mentioned, we already have a lot of things we’re working on and building toward, so I can’t devote a ton of time to it, but trying to talk to people who’ve been there because I haven’t before and learn from people who’ve worked at different scales of organizations and figure out what works and what doesn’t work. So whenever I get the chance and I run into somebody who’s running a different size of org, I’m very curious at this point, asking them about how they do things, how it works. I found it super helpful to just talk to people who are running orgs at similar stages and at slightly bigger stages and find [out]: how do they run meetings, what is their writing culture, how’s their decision-making work?

And right now, I think, just because we haven’t got there yet, I’m still in a learning phase of figuring out what’s probably going to work best when we get there, but I would like to keep some of the same principles of having collaboration. I mean, the end goal is: how do we get to the best answer and best harness collective intelligence? Like collective intelligence in terms of building an open protocol anyone can build on and collective intelligence within the company of, “Everyone here, we hired them because they’re a smart person. What piece of the problem do they see, and how can we put together their insights into something that gets us to the best decision?”

One person who famously experienced a number of trials and tribulations through that scaling process is Jack Dorsey, who started the project with you. He’s still on your board, I believe. Is he involved? Is he someone that you’re bouncing off of, or is he being Jack Dorsey on a cloud somewhere?

[Laughs] I would say the latter, being Jack Dorsey. And yeah, Jack is, I think, a visionary who really wanted social to be on an open protocol, and he has been interested in lots of different decentralized protocols, and he has sort of a big-picture vision of how he wants this to work. And then the specifics of how we built Bluesky was very much the vision that I pitched and that the team has designed. So, the early team I hired, we were people with different experience in different decentralized social projects that came together and learned from the mistakes we’ve made in the past and built something that synthesized those insights.

Do you get any feedback from Jack at all? Does he send one-word emails or anything at all, or is it just kind of silence?

Some feedback occasionally.

Well, early on, we had a sort of discussion around downvotes and upvotes and downvotes or the like button. And early on, we actually put in downvotes, and then we got so much user pushback that we were like, “Okay.” Also, one of the critical problems we realized with the way that we put in downvotes was it’s all public. So because we have all these composable services, everything has to be public for every service to see it and operate off it. So that meant downvotes were public. People really didn’t like that, and we recognize that, you know, it feels bad to see, “Oh, everyone can see all these people downvoted my thing.”

So I think that that was actually an important insight in the sense that social networks, part of the toxicity, I think, comes from there only being positive feedback cycles. And so that means things only really get amplified, and when you really want to say this is a terrible post, you reply to it or repost it or something, and then that actually spreads it further. Whereas systems where you have downvotes mean there’s this downregulation of things, and you can express your disapproval of this by downvoting it, and then if enough people downvote it, maybe it just disappears. And so that’s actually, I think, an important piece, but it’s important that also, when you downvote it, it’s not this big public statement. So that’s something that we would still like to bring back at some point — this downregulation — but maybe not in that exact form. So, we learned from our users.

That sits right at the intersection of “What does a protocol enable?” and “What does the product expose to users??” Maybe the protocol supports downvotes, but Bluesky, the social network that people are on, never shows it to anybody or doesn’t support that feature of the protocol itself. When you think about, “Okay, we should put this in the protocol. This is something we’d like, but our users on Bluesky don’t like it,” does that impact the protocol more, or are you saying, “Well, we’ll get to it in the protocol someday, but maybe Bluesky won’t have it”?

For different things, there’s a bit of both. That kind of depends per design decision. One example here is that we don’t show a likes tab in the Bluesky app, and that was something where it was an intentional decision early on, and actually now, we show it to you. So, since we don’t have bookmarks, you can kind of use your own likes as a bookmark, but we don’t show it to everybody. So, if I go look at your profile, I don’t see your likes. On the other hand, that is public information, so there’s other sites that do show it, but this was just a little bit of friction that we added to say that even though the protocol supports anyone seeing your likes, we think that there’s some downsides to just having everyone’s likes on the profile. So, we didn’t show it by default. 

Another example is we give you an opt-out of the public web interface in the Bluesky app. So all the data is public on Bluesky, all your posts are public, your likes are public, and anyone can build an app that shows all of that, and anyone can look at the fire hose. But if you don’t want a lot of people to click through and see your post, and let’s say one Bluesky post gets embedded in a news article or something, you can opt out of having it be clicked through if someone’s not signed in. And that’s just some extra friction that we’ve added, and we’ve really learned, through building and running a social app, how much these kinds of little frictions and defaults matter in terms of shaping people’s experience and giving a good experience, even if the protocol really enables something that can be done in a different way.

I think this means I should probably split this up into two parts. I want to ask about the app and where you are with the app, and then I promise you the nerdy questions about the protocol are coming. I warn you. If you’re listening to this, just be warned: they’re coming. I would’ve spent a whole hour doing nerdy protocol questions, but I think we should talk about the app first because the app’s on a roll, right?

It’s growing. You have 5 million users. I’m assuming some of the funding you raised is around the success of the app or at least driven by the success of the app. And then, as you mentioned, developers are attracted to the protocol because there are users there. I would say — and I hope you agree — Bluesky has a very distinct culture from the other kinds of social networks that have sprung up in the sort of wake of the Twitter-to-X situation. I’m not going to get myself into trouble by describing the cultures of Threads or Mastodon, but I will say Bluesky is funnier. It’s a little more chaotic. It’s a little weirder. Why do you think that is?

Early on, we had several waves of users come over that were all really big posters. And most social networks, you have this dynamic where you have 1 percent of people posting and then a smaller percentage of people liking and then a lot of people lurking and just looking at stuff. And early on, we had this crazy ratio of 90 percent posters, and so it was extremely active and tons of people firing off shitposts essentially — really fast, funny takes on things and memes and a lot of stuff. And then, since we’ve grown, I think the majority of people tend to use social in a more consumption mode, and so it has shifted a bit more toward lurkers. But still, relative to other apps, I think there’s a lot of posters — people who like posting and engaging that way — and then that results in a pretty chaotic and playful culture, I think.

Do you think that culture comes from moderation? Is it luck? Is it something that you are trying to preserve? Where do you think it comes from, and how do you keep it going?

This, I think, might actually just be a community element, and partially early on, a lot of the team, we all were posters, we all were social media users, and we engaged a lot. We still post and talk to people and stuff. And not just me — quite a few people on the team. The devs talk to dev users. We have people who’ve worked in journalism, and so we have this culture of just active discussion and using the app. 

And then there’s been lots of interfaces for creativity, both intentional and unintentional. So, intentional ones are the ability for anyone to build a custom feed. We created this nice interface in the app. Anyone can build a feed. And so now there’s really niche feeds. I really like the moss feed that just shows you nice woodland pictures of moss, and then you can find a post from some user with five followers, but they posted a moss picture, and now it shows up in the moss feed.

So this is a way where engagement gets spread out beyond one central algorithm into all these little niche communities. I guess another intentional way that’s creative is there’s all sorts of different clients and experiences you can build. So people have built these different experiences that show you things a different way, show you who’s in your social network and all these little visualizations that people have played around with. 

And then the unintentional ones are for a while, we had actually a bug in the app that caused this very long thread to break, and it became actually one long thread called “HellThread” that people were posting very aggressively into. And that bug has since been fixed, but for an era, there was this crazy culture that formed around this bug that almost became a feature. So that was an example of things just being this kind of playful, chaotic energy taking place through both the intentional interfaces for experimentation, creativity, and the unintentional ones, like the bugs that became features.

That second part — the unintentional bugs became a feature — those emergent properties of social networks or communities, it always feels like the most special thing: that the people who run the platforms don’t really know what’s going to happen. The users decide what’s going to happen in ways that are completely chaotic, and the best platforms sort of reinforce that and build upon it. And that is a really healthy feedback loop. 

I would actually put up Twitter as the best example of this feedback loop ever, right? So many core Twitter features were invented by the users. Is that on your mind? That you need to watch your user base and take the ideas into the platform directly and build upon them? Because I don’t see a lot of other platforms really leaning into that anymore.

Totally. Yeah, we do. There’s things that we’ve had on the road map and longer things that we try to stay focused on, like this vision of algorithmic choice and composable moderation. Those are things we prioritize because, from the start, we were going to do that. And then other things, like hashtags. At first, we thought, “Well, I think the custom feeds will allow you to do hashtags within feeds.” People did do that for a while, but people still wanted standalone hashtags, and we got enough takes from users saying, “Just bring us hashtags,” that eventually, we added them.

Same with mute words, like actually, composable moderation is going to let you do something that is much more granular and higher-powered than just mute words. But people want the simple interfaces, too. That’s something we’ve learned. So we added mute words, and we have mute lists and block lists, and these are simpler building blocks. And now you’re going to have bigger, more complex building blocks — the full dashboard of power tools under the hood.

Do you think you’re going to end up in a place where you have what I will just call the Microsoft Excel problem, where so many people have asked for so many familiar features that it’s actually hard to bring them into a new paradigm, like composable moderation or adjustable filters? Every big software product has this problem where there’s an obviously better way of doing a thing, but then people are familiar with the way they’ve been doing it, and so you end up with both or you never have the social capital to do the better way. Do you think about that problem: if the future is feeds in the protocol but everybody wants the brute force moderation of the past, you’re going to end up in that tension? 

Yeah, we actually do think about that a bit, and this is where having some amount of ecosystem guidance from a team that’s really thought about it might help. But also then, if we get it wrong, the goal of the whole app is open-source, and so people can fork off it and sort things out, add things, take things a different direction. And so maybe if our app gets really bloated or if it’s just not working right anymore, somebody who has a strong idea for how to do it differently or do it better can fork it off and say, “Alright, we’re going to clean it up. This is the way we’re going to do it now.” 

So, a lot of the ways that we try to design things is so that we can provide good leadership in the ecosystem. And centralization has its benefits when someone has a plan and they know where they’re going and it’s a good place and people want to follow. But then, if that leadership fails or gets things wrong or changes, then there’s other options and people can vote with their feet and go over to the better option, and new entrants can come in and provide another example of doing something differently, doing it better.

So I feel this in my soul. This is something that we have long wanted across a number of products. We’ve talked about it at The Verge for a long time. I talked to the CEOs of other companies that are in these kinds of relationships with protocols or standards or open source, and at the end of the day, they often come back to “…but we also have to make money, and the best user experience is often the one that we control. So sometimes the protocol slides by the wayside. Sometimes the commitment to open source gets diminished because we have to make some money here. And if we control the user experience, the user experience gets ahead of the protocol, so be it — we’ve got to eat, we’ve got to pay the team.”

In particular, I’ll give an example of this. I think Twitter had a huge ecosystem of third-party clients. It was a playground for developers. So many cool things happened and were built. So many tools were built that had nothing to do with posting or reading tweets at the end of the day. And [then] they had to monetize, and they realized they couldn’t put the ads in all those clients, and they shut down all those clients. Do you see that problem coming? Is that, “Oh boy, we’re going to have to monetize this thing somehow, and we’re going to have to reclaim some control to do that?” Or do you have another idea for how you’ll manage that tension?

Yeah, we do want to make money, and we are trying to do our best. Part of this is sort of tying yourself to the masts, like Sisyphus-style, of “this is the ways that we’ve committed to openness.” We’ve already open-sourced the app, and we’ve open-sourced everything you need to run your own version of Bluesky on the app protocol. And we have built into the protocol ways for users to move all their identity and relationships and data around. And so the idea is that each user has all their data as this package. We actually call it a repository. If you’re familiar with Git and GitHub, it’s like your data repository is like your GitHub repository. And then a lot of people use GitHub, and that’s a centralized site, but then they can take their Git repository and move it to GitLab or some other site.

And so, that’s the built-in “vote with your feet” piece, and you can take your stuff and move it. And then federation is an important piece that we just started for self-hosting federation a few weeks ago, and that means that you can run your own server, host your own data, and then if we try to close things down or whatever, you can just say, “Well, now I’m just going to use a different service because I control my own data and it’s on my own server. I control my identity, my data, and where it’s hosted.” And then that means if a lot of people are federating and self-hosting, that ecosystem can just shift around piece by piece to the services that haven’t shut down their APIs and that haven’t let users feel betrayed.

Those are the things that keep you from making money in the bad ways. What’s the plan to make money?

We’ve been building marketplaces within the app, essentially. So, we’ve got information marketplaces, moderation marketplaces. This is a direction that we’re going to lean into. We’re also providing a service, and we’re already making a small amount of money through one step we’ve made into custom domains. So, the idea here is you can use your website that you own, like jaygraber.com, as your user handle, but you have to be technical to set that up. You have to go buy your domain. You have to change the TXT record of your DNS settings, and a lot of people don’t want to do that or don’t know how. And so, then we sell you a domain through us. We’ve partnered with Namecheap to do that, and then that makes us a little bit of money, and then that’s an example of the kinds of services we’re going to expand into overtime.

Do you think that that is infinitely scalable? I have a quote here that one of our community people at Vox Media gave me. Mark Zuckerberg talks about Facebook being a “community of over a billion people,” and my colleague said, “That just makes me cringe. There’s no such thing as a community of over a billion people.” Is that something you think about: “We should have these smaller communities, or we can preserve the vibe and the funniness of Bluesky now even as we scale because it will splinter naturally with federation and decentralization”?

Yeah, there’s lots of joints on which it can splinter. There’s already feeds, and there’s going to be the moderation services. These will all create different experiences just within even the Bluesky app and are already creating different little corners. Then there will be different servers. There will be different apps running different entire pieces of the network. So, that will, I think, cause a lot more variety in the ecosystem. Part of the idea of having an underlying protocol is that, actually, social networks systems tend to go through periods of consolidation and then fragmentation, and a protocol that connects them all lets this happen naturally.

So, the example is the web itself is a protocol. Websites come and go. Some of them consolidate a lot of users, and then over time, they get worse and worse, or they shut down and then users go somewhere else. And that consolidation and fragmentation happens on the web. So this is trying to do that for social, so sites like Bluesky can come and go and grow, and then if we someday aren’t doing the best anymore, something else can come along, and users can fragment out and then maybe re-cohere around someone else who’s doing things really well.

You mentioned the plan for monetization is around the marketplaces for algorithms and feeds. I assume that will take place inside the Bluesky app, the one that you control, right? That makes the most sense. “Here’s all these users. You can buy an algorithm from us and see what you want. Maybe it’s about moss. Maybe it’s something else.” Is the plan to make money inside the app, or is the plan to make money with the protocol?

There’s ways to explore both. Right now, we have a lot going on in the app that we’re really focusing on, but also, I think one insight is we’re trying to make social more like the web. And early on, people didn’t know how the web was going to monetize, but it definitely did. Services definitely monetized on the web — tons of services that made lots of money built on open protocols and nothing that’s gotten to massive scale on the internet has ever just totally failed to monetize.

So, that’s something that we just think we need to show people, first of all, that this vision is something that can work, that is fun, useful, good, better than the alternatives, and also good for society, creating healthier social media ecosystems. And then we really believe that money follows value. If we can demonstrate this value and people really want this, then it’s something where we’ll be able to make money offering services in this ecosystem.

Services on top of the protocol, though, not the protocol itself?

The protocol itself is not a blockchain, not something with any built-in monetization mechanism. It’s like SMTP for email or HTTP for the web. It’s just a way for computers to talk.

That’s a great preview of my next question. Well done. Those protocols — HTTP [and] SMTP — they’re maintained by standards bodies. Some big companies volunteer some people. They don’t make money themselves — they’re maintained by a bunch of people who are invested in those protocols existing and growing so they can put services on top. That’s more or less what you’re describing. Would you ever turn over AT Protocol to a standards body?

Yeah, and we’ve already begun talking to some standardization bodies — like starting the very early stages of that work, socializing the idea, taking on the pieces that are relatively more solid, as I mentioned earlier. The AT Protocol is actually made up of several pieces that the way identity works, the way data works, and these pieces, we aim to get standardized and then stewarded by governance bodies.

Alright, get ready. It’s time for some extremely deep standards drama around protocols. I told everyone it was going to happen, and now it’s here. So, you’ve got a protocol: AT Protocol. There’s a competitor protocol called ActivityPub. Mastodon runs on ActivityPub. Threads is going to federate using ActivityPub. What, in your mind, are the pros and cons? What are the differences?

So, ActivityPub was around when we got started. I did an ecosystem review of all the centralized protocols that existed in 2019, including ActivityPub. And we looked at it and decided that we needed to build something different because there were some critical pieces that we thought were missing. So, one of them was around the composability and the interfaces for composability we’ve designed. Like the way that we do custom feeds and all these moderation labelers, that’s really not possible with the way that things are very server-centric in ActivityPub right now. Your server is very much your community where all of this happens, and we have split things up into these microservices on the backend, mirroring a bit more of how a large-scale social network works with the global feed. 

The other thing was the global feed, like having global search and discovery be a first-class thing that we were building for. And also, having all of our users at the start know that this is public data and it will be remixed in all sorts of ways by global feeds is something that, both technically and culturally, we had to design for. Because ActivityPub has clustered more around servers that they federate and they talk to each other, but there isn’t a service that scrapes all of it and gives you this big fire hose, and even if it’s technically possible to build, there’s been community resistance to people doing that. Even someone in the community who was building a bridge between AT Protocol and ActivityPub recently got a lot of pushback from the ActivityPub community on not wanting that bridge. So, that’s just sort of both cultural and technical. 

Then another thing was we really wanted to get account portability. So, this ability to leave with your identity and your data and have fallbacks with the way that we’ve designed your repo, you can even back up all your posts on your phone or back it up on your server that you control, and then you don’t have to have any sort of friction when you want to move. So, you can move between services in ActivityPub. But if… for example, Queer.af recently, their .af domain was seized by Afghanistan, and then people were stuck because there was no warning, and then they have to rely on their old server to help forward their stuff over to a new place. So, we wanted to get around that problem and make sure people always had the ability to move. 

Then, we wanted to have good UX. There were just a lot of complaints with users around the UX of Mastodon, and we wanted to provide something that was more just tailored for the mainstream user who was used to Twitter, who could come on and not have to worry about, “Oh, what server do I pick when I sign up? What does it mean? Is this going to shape my experience forever?” On Mastodon, you have to know that right when you sign up because it’s going to shape your experience. Here, you can sign up on our server, which is the default lobby or gateway into the ecosystem, and then you can move to another server and shift to your own. So, when we opened up federation, several of our team members moved their main accounts off onto their own server, and it was pretty seamless. Nobody notices that people are running on their own servers, and it’s still just all one experience in the app.

Let me be very reductive here and see if I got this right. It feels like the first order bit for Mastodon [and] ActivityPub is the server, and you’re saying your first order bit is the user, right? You have a user account; it can kind of live anywhere, and that’s the thing you’re in control of. On Mastodon, there are servers, and the server administrators are more in control. Is that too reductive? Is that accurate? Is that a good way to think about it?

That’s actually a pretty good way to think about it. And this is where, earlier on, since we’re getting technical now, I said we were borrowing from peer-to-peer systems. Peer-to-peer systems really try to do everything at entirely the user level with no servers. And that actually creates a complicated user experience, but that means that every user’s fully in control of all their data all the time. We have the ability for users to do that, but we have servers to make your life easier, so you don’t have to have your phone be trying to directly communicate with 5 million other phones or a computer trying to do that. But then you can always move off of our server and move off of these services.

So, the property that we really tried to get out of this piece of decentralization is the right to leave. And so, you can use a service and can use even a big service, but you always have the right to leave built in because we’ve designed around the user. And the other way we’ve designed around the user is not just at the technical level but at the UX level. So, try to design for: what do users really want? How do we get that sort of convenience and ease of use for the user? And thinking from a UX as a first-class thing that we’re designing for.

There’s some tradeoffs there. Some are really obvious, right? Mastodon server administrators often find that they’re running a server and they have software costs and overhead, and they have to maintain a server for a bunch of users. At the same time, they’re in charge, for better or worse. If something goes wrong, they can shut it down or they can delete stuff. They are in control of the content. They can set different kinds of content moderation policies. There’s a decentralization of authority in that way. It feels like AT Protocol is much more individualistic in that way.

ActivityPub points you toward groups of collectives that may wish to interact with each other. And AT Protocol and Bluesky are much more like, “You’re on your own. If you like it, you can leave, and there’ll be a larger market of individuals.” How do you think about, “Okay, if someone doesn’t like the server they’re on, they’re going to leave”? They might not actually be sending any signal at all that the server administrator was doing a bad job because you can’t tell. Whereas on, I think, Mastodon, it’s like, “Oh, somebody just left this account behind.”

I think, actually, collectives are going to emerge, and they already have in some ways. Collectives aren’t formed at the server infrastructure level. So, one of the ideas that we had was, on Mastodon, your moderation is very much tied to who runs your server, but often, the type of person who wants to do DevOps essentially and manage a service online is different from the type of person who’s a community builder and wants to organize a group of people together to set a different set of community norms and moderate. You can get those people together to run a Mastodon instance, but it has to be tied around the instance, so we separated those pieces. So, now you can get a group of people together and run a moderation service. We give you software to manage reports and stuff like admin tooling, and you can do this with the whole team, and then you can get a bunch of people to use it, and then you can get a whole community based around a moderation culture that you’ve created.

Then, you can do that with the feed as well. And then all these things, as I mentioned in being composable, can be bundled and unbundled in different ways, so you could recreate the Mastodon “everything is tied to your server” experience by setting defaults on your server: this is the moderation regimes we’re going to be using, here’s the labelers we’re using, here’s the feeds we’re using, etc. And a server operator still has power at the end of the day because they control where your stuff is literally hosted. But then there’s all these other interfaces beyond that where control and agency can happen. So, any service in the network has the ability to set rules, and then users have these interfaces where even if they’re non-technical, they can go in and create rule sets.

You mentioned the controversy around the bridge software. It was called Bridgy Fed. This was some of the most intense GitHub comments I’ve ever seen in my entire life — truly out of control for basically a piece of middleware. Why do you think that happened? That the idea of bridging AT Protocol onto ActivityPub led to one of the most intense recent developer flame wars that I can think of? Is it the cultures of the protocols, the culture of Mastodon versus Bluesky? What, in your mind, led to that conflagration?

I think it’s a big part of the culture of Mastodon. This was, as I mentioned, one of the reasons that we didn’t try to get ActivityPub to change toward the direction of what we wanted to build because not just the technical primitives being different, there’s also this culture of resistance to global feeds and global algorithms. And that means that people who had tried to do a search engine for all of Mastodon in the past or things like this had gotten shut down even before Bridgy Fed in the past. I had seen that even back in 2020, 2021.

I think it’s just a continuation of that culture. And the key thing is even if you have a protocol that lets servers talk to each other and federate, part of it is human governance, and if people who run that server don’t want to talk to you, then even if the computers can talk, they don’t have to talk. And so that’s essentially people in Mastodon expressing their preferences of the kind of communities they want, which is, they don’t want to talk to Bluesky users, so they can just not federate. And then there was a lot of discussion that was around opt-in versus opt-out. So, it’s discussions around what kind of governance norms do we want to set as protocol communities.

Part of the argument in favor of the Mastodon culture that I’ve heard over and over again is it allows safer spaces to form for certain kinds of groups — that the servers can be closed. It is mostly opt-in that people defederate all the time in the Mastodon community. The number of servers that have defederated from Threads or have sworn a blood oath to never federate with Threads, it’s very high. Do you think about Bluesky as providing the same kinds of tools that make people feel like the spaces are safer?

That has been our goal: to give users the tools to build spaces that are as safe as they want. And I think one thing that people have said the protocol is missing right now is private accounts. That’s something we’d like to get there, but so far, in our mission statement, we’re focused on public data and the open Twitter model, and there’s other protocols out there for private communication. Like there’s Matrix, Signal, and other protocols that do private communication. So, we focused on the hard problem of global public conversations and algorithmic choice. But even within that, we want people to have the ability to feel that in Bluesky right now, they’re participating in public data, public posting, but you want to not interact with people who are going to harass you. You want to not see stuff that’s going to be damaging. 

So, you want to be able to partition off your little piece of the network. And so, we’ve tried to give people all the tools to do that, not just at the architectural level of here’s servers that you can run and here’s services you can host but also at more user-friendly levels, like the user lists and block lists and the ability to run your own labeling service where someone non-technical can do that through software that we built for you to go in and just say, “Alright, I want to start setting rules and filtering out this kind of content for me and my community.”

We’ve mentioned Threads several times now. Obviously, Meta is starting Threads. When you think about onboarding the non-technical user onto decentralized social media, Threads seems like it’s way ahead of the curve. It has 100 million users. It’s going to use ActivityPub, right? Meta has already started testing it. I think Adam Mosseri’s account can be followed on Mastodon through ActivityPub. Does that represent a competitive threat to you? Is that something that you’re watching, or are you trying to build something else?

I think it’s really interesting that the ecosystem is moving this direction overall. I also, at the end of the day, want everything to be on an open protocol, and I think that our protocol is the most resilient long-term and is going to give the most flexibility and guarantees of this being a healthy ecosystem where people can move around. But also, I’m just excited to see things starting to move this direction. Mike Masnick, who wrote “Protocols, Not Platforms” as a paper, said that he shopped this around to all the big tech companies at the time, and this is one of the things that convinced Jack to do Bluesky. So, Twitter listened, but Meta was like, “No, this is never going to happen.” But since then, they seemed to have moved on from that position, and maybe we were one of the things that have started moving things.

I think that overall, it’s a healthier direction to go toward open protocols. One of my concerns about ActivityPub and the Threads model is, because it’s so server-based, if you have one massive server come on, like the Threads server, things really centralize around that. And so, it might functionally not be very open because people are tied to the server and it’s hard to move, and it’s going to be really up to Threads whether they ever become fully protocol-compliant and let people move off Threads easily. Then, is there going to be that composability there? Are users really going to have that much control? Those are open questions, and we’ll see where Threads takes things. 

But I think that we’ve built that in from the start, this openness and composability. We’re also open-source, and so that’s the other guarantee. Our code is open-source, all these interfaces we’ve built. That’s not the case at Threads. It’s going to be something that, if it becomes fully protocol-compliant, I think there will still be risks to users.

One of the reasons you can see why Threads might prefer ActivityPub, just based on this conversation, is it’s very server-based. I’m guessing Meta loves the idea of running the big server. There’s some benefits to that, too. Meta has a giant compliance department and a huge content moderation team. There are parts of the Bluesky approach that seem like it’ll be very challenging to run at scale. Just really dumb — the data is everywhere. It seems undeletable because it’s so public. If there’s a copyright claim, can you actually pull something down? Can you get rid of it?

Ultimately, data has a place where it’s being hosted, so whatever server that you’re running, the server runs in the jurisdiction where you posted it, and whoever’s running that server is going to have to be responsive to their legal jurisdiction. So, there ultimately is a host for data. It just moves around more.

Are you thinking about that stuff, like, “Boy, we’re going to run in a bunch of jurisdictions, and the Indian government wants Bluesky to have an actual person in India that it can arrest if they don’t get the moderation rules they want”? Those seem like very thorny issues for a company that’s very early, especially one that has to explain, “Hey, maybe we don’t have all the controls that you’re used to from a TikTok or a YouTube.”

This is one of those things where I’m really glad we just hired Aaron because he has experience–

… with platforms at scale and government requests, and our goal right now is to grow and get to that kinds of scale where these will be a good set of problems to have, but bringing in the expertise early so that we can be prepared to deal with it and someone who really understands our whole architecture and can help us navigate those questions is where we are right now, in terms of preparing for that.

One of the things you talked about rolling out soon is some of the composable moderation. I read your blog post about it. Again, just to be reductive — and tell me if I’m wrong — it seems like the core idea here is more people should be able to label more things in more ways. We’re going to allow the users to say, “Here are the labels, and we’ll get the labels out there.” It’s an election year. I feel like the election-year experiments with labels or the covid-year experiments with labels have taught us all a lot. What are you pulling from all of those previous experiences with big platforms doing labels to say, “Okay, our big idea is labels, but decentralized”?

I think Birdwatch was this project that we thought was really cool on Twitter, and seeing Community Notes come out of it was a success in terms of them pioneering an algorithm that could usefully augment and annotate information out there. And, basically, the interface we built is for a thousand algorithms like this to be experimented with, and for some of them to be complex algorithms that were very specialized algorithms like the Birdwatch algorithm, Community Notes algorithm. 

Or some of them to be very manual, like people are like, “I’m an expert in this thing. I’m just going to label for this thing.” And so then that allows all these things to come together to an experience that, again, it’s about harnessing collective intelligence, letting experts, people with local context, move faster on these things than the company whose interests might not be entirely aligned in solving this niche problem or this jurisdiction or that’s just too slow because it’s a big company or it’s a small team or whatever.

So, our goal is to let that whole ecosystem just iterate and experiment, and then we try to have some amount of leadership in terms of what we’re encouraging people to build, how we’re creating and surfacing the best stuff that gets built and bringing it to user’s attention and helping them install it. I think that it is going to produce a better system, and there’s some ideas right now in the academic world of this concept of moderation middleware that they’ve been designing and thinking around doing this within the centralized platform paradigm of offering, “Hey, instead of the only way to make change at these companies is to have nonprofits, interest groups, governments lobbying the company for change, maybe you could create a middleware interface where people could directly add some of their input into things.” Composable labeling is like middleware but as a first-class thing built in where anyone can build something like this, and because it’s fundamentally open, it’s taking that from the fundamental premise rather than trying to add that as an intervention into a centralized system not designed for it.

Can I ask you a question that I always ask the AR companies but, I think, plays in the world where we have different kinds of labeling systems? I can imagine a world where someone posts a picture of the US Capitol, and for an AR company, this is very difficult, right? You’re looking at the United States Capitol, and your Apple Vision Pro has to tell you what happened there, and you can either put January 6th or you can put the passage of Obamacare — two things that would enrage different people in very different ways.

If you have composable moderation, you have different kinds of algorithms, appending different kinds of labels. You still might break that shared reality. Someone might post a picture of the US Capitol, and one person might tell you, “This is where Donald Trump did a coup,” and someone else might say, “This is where the election was stolen from Donald Trump.” Are you worried about that sort of breaking of reality, or do you think the market will actually converge on the truth?

Both. I think that basically — to go very meta with this for one moment — I think that every new form of technology that comes along causes a lot of disruption, particularly information technology. So, when the printing press happened, people said, “This is going [out] in the world. We’re going to have heresies like revolutions.” All those things did happen, and there was a lot of chaos for a while because the shared truth of the old order broke down. But then eventually, we developed new institutions that built around the printing press, and we developed new ways of understanding the written word and arriving at truth. We developed encyclopedias. We developed the academic citation system. We developed a legal system that’s very text-based and builds off previous texts, and so we developed all these systems that helped us create a society that works.

I think the information era has once again totally fragmented shared reality that the printing press had established and the written word and then broadcast media, and so now everyone can broadcast on social media. And how do we develop new kinds of systems where we can come to shared truth again? We’ll have to develop systems that [have] new algorithms like community notes, that kind of approximate shared truth from all these people broadcasting into the web.

And how do we do that? Is it going to be a centralized effort where one company just funds and develops that and gets the right answer? Or are we going to open it up to experimentation and then maybe lots of people going at the problem start to arrive at the right answer? And our approach is the latter, which is opening up the system, having lots of people experiment on building new ways of curating, like arriving at approximating truth, pulling lenses apart to look at this shared information and then combining them again will approximate at some point. People will start to converge around new ways of doing things. 

So, we also, as a company in the ecosystem, will be trying to approximate: what is a good way to do this? What are the best labels out there? How do we bring those in? Maybe if we’ve talked to a reputable fact-checking org and they’re running a really good fact-checking service, we’ll promote this in the app, but then someone else could build a different app, and then how is the ecosystem going to evolve? Well, it’s an open system, and eventually, at some point, we’ll probably reach a point where we know what the good sorts of ways of arriving at truth are and the good institutions, like providing objective stuff or stuff that gets things right most of the time, and we’ll converge around that.

That’s a very high-minded idea. You have to fund it. It sounds like your plan to fund it and be sustainable is to operate those marketplaces to take a cut. So, I buy an algorithm from you, and you’ll take a piece of that transaction.

Yeah. As a company, but then also because this is an open ecosystem, it’s not just limited to our funding model. We are in the early stages of building out this vision, and so we need to keep going, prove out the vision, show this, make money as a company, create a marketplace, but then also somebody running one of these services on its own, already people building clients like other services in the ecosystem are… some of them are charging, some of them are taking donations. I think some things will end up nonprofit-funded. If you look at things like email, you have big companies that run email services. You also have nonprofits that run their own email servers, and you have institutions like your university might give you an email, or your company might provide you an email.

And so, this is something where if we show value, once again, money follows value; we show enough value in this ecosystem. Some of these might become institutions funded by nonprofits who are interested in a better information ecosystem. They might fund moderation services, might fund feeds, things like this. Institutions might run them. Communities might get together and self-fund. Some of them might be paid services that payments happen outside of us. Others might be paid services where the payments happen through us.

Do you think any of them will be businesses unto themselves? I’m reminded of this Bill Gates quote about Windows as a platform, where he is like, “A platform is when the economic value of everybody that uses it exceeds the value of the company that made the platform.” So, all the people who built companies on Windows, you add them all up, they exceed the value of Microsoft. You’re kind of talking like that, right? You’re going to build a platform. You’re going to let people buy and sell on it and build products on top of it. Do you see someone saying, “Okay, there’s an opportunity to build a great content moderation company that runs on top of AT Protocol that Bluesky users can buy”?

We’ve already seen some people start to move in that direction. I think that is something that’s going to happen.

How long until that happens?

I wish I could see the future, but I don’t actually have that ability.

You’re probably better at seeing it than I am. I’m very curious. This feels like there’s some inkling of “this is how it should work,” but you’ve got to train a whole bunch of people to think differently, and they’re right up next to Elon Musk being like, “I’ll just do it for you. Just show up and tweet.” That seems like the hard part, right?

Right now, a lot of what we’re doing is trying to communicate the vision, like talking to you. Earlier on, before we built stuff out, yeah, like last year, I was turning down a lot of media because we had to build this out, and we really want to show, not tell. Now we have examples that we can point to, and it makes it a lot more real rather than me just sitting here expounding on the printing press. It’s very high-minded and not concrete. But now we actually have examples of many different feeds and many different services that have been built in this ecosystem, and so one of the things we’re starting to do is to fund the developer ecosystem.

We’ve started a grant system, and we’re going to start giving money to some projects just to help boost this ecosystem, but I think over time, these funding models are going to shift and change, and we’re going to experiment with a lot of them because a whole ecosystem that’s like the web means a whole variety of experiments like the web itself kicked off, and I find that really exciting because social has been in a period of stagnation because it’s consolidated in a few companies, and early on, people built all sorts of Twitter clients and experiments on social, and even Facebook was more open back in the day, and then that whole ecosystem got closed down when these companies shut down their APIs.

But if we open that up again, I think social will be [in] another period of experimentation. There will be lots of companies and hobbyists, just people trying things and playing around again. It’ll be like a playground era of social but in a mixed iteration where we’re aware of a lot of the problems that happen through platforms and mass broadcast at scale. And we have a lot of new tooling, like machine learning has moved on a lot since the last time we had a lot of growth in social, so I think a lot of interesting things are going to come out of this era of building.

Let me end by bringing us all the way down on the ground. You have a lot of big ideas about protocols and platforms. (I’m obviously obsessed with them.) But you run a social network that’s growing in an election year. You’ve had to deal with some very real content moderation controversies on Bluesky already. There’s been some real problems with racism. There’s been problems with death threats, and you’ve just had to do some direct moderation, just take some direct moderation decisions. 

It’s an election year. It’s going to be weird. There’s going to be a bunch of AI deepfakes of both candidates coming. Do you think your new tools are going to let you address that stuff in a new way, or are you going to have to go back to the old ways and just take some direct moderation decisions as you’ve had to in the past?

I think the new tools will help a lot, and we’re also not relying entirely on them. As with everything we’ve done, we’re doing a hybrid approach where we are running our own moderation team with experienced leadership and we are going to do our best, and then we also have all these open interfaces where other people who are more experts or people who think we’re not doing it well come in and do it themselves, and then we’ll stitch a hole together out of both all the open experimentation and the decisions that we’re directly making and what we decide to carve out as, like, this is stuff we’re doing really well and that we think the foundation we need to set for this to be a good experience.

Are you worried about the flood of AI-generated garbage on every platform? Is that coming for Bluesky right now?

Not quite yet. I mean, we’ve been dealing with spam and all the other problems that networks run into, but we’ve also been building our own automated tooling, and it’s a constant cat-and-mouse game, but we’re trying to stay cutting edge in terms of how we address these things using a combination of human moderation and automated tooling to catch stuff. And then also, part of the open interface is we’ve already seen some cool experimentation going on. We had a hackathon a few weeks ago at [Y Combinator], and there were some groups that made deepfake detectors using state-of-the-art models that do deepfake detection, and so it’s this cat-and-mouse game of AI is being built to create deepfakes. Also, there’s deepfake detectors now, and so those things can come together into things that, if we don’t pull them in directly, they can run as independent services in this composable moderation thing, and if they’re running really well and super critical to the ecosystem, then we’ll find ways to integrate them more into the app experience.

Are you seeing that already taking place on Bluesky, or are you just watching this sort of Taylor Swift deepfake problem on X and hoping it doesn’t hit you before that stuff is ready?

We haven’t seen a lot of it yet, but we’re preparing to deal with it through both our policies and training, making sure everyone is aware and then also making sure that we’ve gotten this moderation tooling built out, which you’ll be able to see in action soon, that’s going to let people start to make these interventions.

Alright, last question. It feels like the people who run the more centralized platforms have lately become utterly beholden to their own posters. Like, Elon Musk seems very responsive to the posters on Twitter. Adam Mosseri, in the middle of the night, is like, “I don’t know. I’ll look into it.” You see that reaction, like, “I’ll be the face of the thing.” To some extent, you’re the face of Bluesky. You’re here doing this interview, but you have an out, right? It’s come up several times in this conversation, like, “I don’t know. We’ll see what the marketplace does. We’ll see what the community does.” Do you feel that pressure to be the face and to be responsive, or do you think that that out will actually take hold?

Yeah, I do, and I’ve accepted it to some extent, and I’ve also told people I am sort of the steward of this ecosystem right now and the creator of… let’s think of it as a city-state in an ecosystem we’re trying to build. My bio on Bluesky is, “Let’s build a federated republic starting with this server.” And so, there’ll be lots of other federations out there. I encourage you to start one, and I will do my best on this one. 

And then, I also am not… something we joke about on the app is, like, “posters madness” is a real thing that people can get on social media. We try to do some internal awareness on becoming audience captured by our own audience and stay aware of that. And then also, I try to do my best to — I’m a poster — engage with the community but not be the main character all the time. There was a joke at Bluesky I found very funny, which is, “The one thing I love most about this app is I don’t know who the CEO is.” Just like, great. If it’s a thriving enough community where it doesn’t even need me as a main character, that’s great.

Well, Jay, that is an incredible place to end it. You’re a great main character, by the way. Thank you so much for being here.

Today, I’m talking to Jay Graber, the CEO of Bluesky Social, which is a decentralized competitor to Twitter, er, X. Bluesky actually started inside of what was then known as Twitter — it was a project from then-CEO Jack Dorsey, who spent his days wandering the earth and saying things like Twitter should be a protocol and not a company. Bluesky was supposed to be that protocol, but Jack spun it out of Twitter in 2021, just before Elon Musk bought the company and renamed it X.

Bluesky is now an independent company with a few dozen employees, and it finds itself in the middle of one of the most chaotic moments in the history of social media. There are a lot of companies and ideas competing for space on the post-Twitter internet, and Jay makes a convincing argument that decentralization — the idea that you should be able to take your username and following to different servers as you wish — is the future.

It’s a powerful concept that’s been kicking around for a long time, but now it feels closer to reality than ever before. You’ve heard us talk about it a lot on Decoder: the core idea is that no single company — or individual billionaire — can amass too much power and control over our social networks and the conversations that happen on them.

Bluesky’s approach to this is something called the AT Protocol, which powers Bluesky’s own platform but which is also a technology that anyone can use right now to host their own servers and, eventually, interoperate with a bunch of other networks. You’ll hear Jay explain how building Bluesky the product alongside AT Protocol the protocol has created a cooperate-compete dynamic that runs throughout the entire company and that also informs how it’s building products and features — not only for its own service but also for developers to build on top of. 

Jay and I also talked about the growth of the Bluesky app, which now has more than 5 million users, and how so many of the company’s early decisions around product design and moderation have shaped the type of organic culture that’s taken hold there. Content moderation is, of course, one of the biggest challenges any platform faces, and Bluesky, in particular, has had its fair share of controversies. But the idea behind AT Protocol and Bluesky is devolving control, so Bluesky users can pick their own moderation systems and recommendation algorithms — a grand experiment that I wanted to know much more about.

Finally, Jay and I had the opportunity to get technical and go deeper on standards and protocols, which are the beating heart of the decentralization movement. Bluesky’s AT Protocol is far from the only protocol in the mix — there’s also ActivityPub, which is what powers Mastodon and, soon, Meta’s Threads. There’s been some real animosity between these camps, and I asked Jay about the differences between the two, the benefits of Bluesky’s approach, and how she sees the two coexisting in the future. 

Okay, Bluesky CEO Jay Graber. Here we go.

This transcript has been lightly edited for length and clarity. 

Jay Graber, you’re the CEO of Bluesky Social. Welcome to Decoder.

I am really excited to talk to you. Bluesky is one of the most exciting new social media apps on the scene. There are exciting ideas inside of it around federation and decentralization, both of which I’m obsessed with. So get ready, but let’s start at the very beginning. What is Bluesky Social?

Yeah, Bluesky is an open social network. It’s open-source. It’s built on an open protocol, but it looks and works very much like Twitter. And so it’s a project that was actually originally founded by Twitter back in 2019. Jack [Dorsey] tweeted that he was going to fund the development of a protocol that Twitter would someday run on. And so I came into the project as someone outside of Twitter that had never worked at Twitter but had experience with decentralized social protocols. So I built out a protocol, and then, through a series of changes of what happened to formerly Twitter, now we’ve ended up building an app that looks a lot like Twitter.

I can’t imagine anybody didn’t figure this out. But Jack is Jack Dorsey.

The former CEO of Twitter.

The once former and then CEO and then now former CEO once again of Twitter. That piece where he thought Twitter should have an underlying decentralized protocol, and I don’t want to talk about Twitter too much, but that was a big deal, right? The idea that Twitter was too important, it was too centralized. I think Jack was uncomfortable with his moderation power. 

Obviously, we went through a big moment around content moderation, but that core idea that one company should not have so much power led to a lot of things, including the creation of Bluesky. What was the pitch there? Did he come and say, “Hey, come run a program inside this company to devolve the power of the company”? 

I actually pitched Twitter a lot on how I thought Bluesky should work.

So the way that this worked is that Jack tweeted that, and they created a Bluesky Twitter account at the end of 2019. I got really excited. I was currently working on a social app of my own that was an alternative to Facebook Events called Happening, and I’d been building on decentralized social protocols, playing around with them, doing a lot of research, published a lot of research on it. 

So then I was one of the experts that they pulled into this small chat room with a group of people who seemed to know what they were talking about. And then, they asked the group, “What is the best way to build this protocol? You guys know about protocols.” And then, in 2020, the pandemic hit, and so this whole project got derailed by the pandemic as Twitter was clearly occupied with other things.

And then mid-2021, Twitter circled back around and started interviewing people, both people within Twitter and experts without, interviewing somebody who could lead the Bluesky project. And so, mid-2021, they interviewed me along with some other folks and then chose me to lead Bluesky based on this vision that I pitched for how we were going to build Bluesky, which is essentially how we’ve built it. And so that’s how I got involved. 

Once I took lead, there were different options as to how we were going to structure Bluesky. As a category, a project within Twitter, it was very much like Birdwatch, which was what built Community Notes, sort of a skunkworks project that was going to go off on its own and do this experimental way of doing things in a more open, crowdsourced way. But I didn’t want to be within Twitter because Twitter had moved very slowly to nominate me as lead in the first place.

And there was just going to be a lot of challenges. 

One of the things that I was hyperaware of was just the risks of centralization because one thing I always like to say is the captain can always sink the ship. Jack was the captain of Twitter at that time, but if the captain changes or if that support left, then I would lose institutional support for Bluesky, and I really wanted this project to exist. So I insisted we spin out as a separate company, and then that took about six months to figure out, but then we got set up as a separate company, and then, actually shortly thereafter, the captain did change. And so then we ended up in this position — we were able to keep going. In a lot of other circumstances, if we were within Twitter, I don’t think the project would be alive anymore.

That part is really interesting, right? The idea that the leadership might change, but the project should remain or the protocol should remain. Right now, Bluesky is a company. The audience heard me introduce you as the CEO of Bluesky Social, which is the company. It’s also the app — it’s not the protocol. Explain how that all works. What is Bluesky Social, specifically?

So Bluesky was originally just the name of this project within Twitter to build this decentralized protocol. Then, the protocol we actually built was called the AT Protocol, or AT Proto for short. And this sort of symbolizes the fact that you can find people at their @ handle username. And so we wanted this to someday be the last social identity you’ll ever have to create because you can move it in between apps and services. You can take your identity and your relationships and your data with you. And so that’s the underlying protocol layer. 

The goal is to make social apps work more like the web itself. And then Bluesky is the app that we’ve built on top of the AT Protocol, both to show how it works and to develop the protocol actually through product-driven protocol development. You run into a lot more things once you’re actually building on a thing that you intended to have developers build on.

So Bluesky, the product on top of AT Protocol. I always call it “A.T.” Protocol. Are you good with “A.T.” or “at”?

We’ve been calling it the [at] Protocol because, yeah, it’s easier and sometimes just “AT Proto” for short.

It stands for something, right? 

Authenticated transfer protocol.

Very good. Was that a bacronym? Did you back into that or did you come at it front ways?

I think we actually came at that front ways because, earlier on, we were calling this — there were a lot of posts I wrote early on that are way more technical than the audience we’re usually talking to these days, but stuff about self-certifying data structure and authenticated transport, and these are all the design philosophy that we pulled into Bluesky, which borrowed from peer-to-peer networks but did this hybrid federated approach. So that’s the wonky parts under the hood that most people don’t think about anymore.

I love that. So you’ve got this protocol. You’re going to build a product on top of it called Bluesky. Bluesky is now open to the public. It’s growing. I think the last number I saw was 5 million users. There’s obviously a moment to capture market share from Twitter. I recently saw a report saying 30 percent of Twitter users have stopped using it in the last year. Is your focus on growing Bluesky the product? Is it on the AT Protocol? How do you shift that attention? They seem like they’re competing a little bit.

Yeah, they both cooperate and compete because part of our philosophy was that developers, we want to get developers to build on the AT Protocol, but devs want there to be users to build it for. And so if you get users into the ecosystem, anyone building on either the interfaces for custom feeds or custom clients on Bluesky can come in and now build for 5 million users. But also, AT Protocol will be able to support other kinds of apps, and then users will be able to move between them with the same identity and relationships, and then that will just make it a lot easier for other people to build. And so that was part of our idea was just build an app, and then we get this positive feedback loop going of developers and users coming into the ecosystem.

So that “cooperate and compete” framework is really interesting. Most standards that have a big product on top of the standards have that kind of relationship, but they’re usually managed by giant companies or at least a handful of companies that are competing themselves. There might be a standards board involved. Right now, for Bluesky and AT Protocol, it’s just Bluesky. How is Bluesky structured, and how do you contend with that cooperate-compete dynamic inside the structure?

Yeah, so we are structured as a public benefit corporation, so that means we have a mission. It’s to “develop and drive adoption of large-scale adoption of technologies for the open and public decentralized conversation.” I think I might’ve mangled that a little bit, but the gist of it is for open public conversations on an open protocol, and that’s the mission that all of our stakeholders know — our investors, our board members — and we’re able to pursue that and put any sort of profits back into that mission rather than having to put fiduciary duty first. But that doesn’t mean we’re prohibited from making a profit. So we are a company.

And the company is responsible for the protocol right now. Inside the company, how is that structured?

Yes, it’s within the company. So basically, we’re all working on it together. Everyone’s very mission-aligned that we’ve hired, and it’s something that, right now, we’re just trying to show the benefits of this kind of an open ecosystem that people can build on and get users in the ecosystem, and then we’re moving toward standardizing pieces of the protocol. So pieces that have become really stable, it’s working well. There hasn’t been any need for big changes recently. That’s something that we’re trying to move toward a standardization process because once something gets standardized, then it’s going to move a bit slower, but then that gives it more resilience.

How big is Bluesky today? How many employees are there?

We’re about 18 across engineering and ops, and then we have about that number on support and moderation.

The moderation side is the part that scales the fastest, right? If I look at Bluesky and AT Protocol and what it’s meant to enable, it’s a way to rethink moderation, probably most of all, right? If it’s as simple as “you don’t like the moderators, and you can take your account and leave” — that’s one idea. You have this idea about composable moderation, where people could write their own moderation stacks or different kinds of filters or different kinds of algorithms. Right now, though, moderation just sort of scales with the user base — you just have to spend more time and money moderating. How are you managing that against Bluesky’s growth?

We just hired Aaron Rodericks, former trust and safety co-lead at Twitter and head of election integrity. He’s come on to bring a lot more experienced leadership to this team and make sure that we’re building a strong foundation for moderation on Bluesky. But what you mentioned there — you definitely read about our approach — it’s a hybrid approach between us running a foundation within the app, making sure that we have this really good service within the app, and having decentralized moderation. The closest analogy here — it actually works a lot like custom feeds — is the custom feeds we build. So we provide some default algorithms, and we develop an in-house algorithm, but we also allow anyone to build an algorithm. 

So there’s about 40,000 custom algorithms out there and custom feeds, and some of them are very simple, some of them are very complex, and users can install them and switch between them. So moderation services are like this, but because Bluesky is also a digital space that we’ve created, you can’t opt out of [our moderation service] and our client, and then that’s what we’re providing all this moderation, in-house moderation, for. And then, actually, by the time that this podcast airs, you’ll be able to go into the Bluesky app and see the composable moderation because that’s coming out this week.

I’m really excited about that. I want to come back to that and talk about it in detail. I want to stick on the Decoder question just for a couple more minutes here. You started this thing inside of Twitter. You fought to leave it. That’s a big decision to make. The entire world of social networks, I think, has flipped over since you spun up Bluesky and began work in earnest on this project. It’s an election year. You have big ideas about moderation, you have a team, you’ve got to raise money. That is an awful lot of decisions to make. How do you make decisions? What’s your framework?

I think I have a very collaborative approach, but as with a lot of things I do, I take a hybrid approach between centralization and decentralization.

Perfectly on theme. Very good.

It is. I really do. I spent a long time working on decentralized technologies, and I saw their limitations as well, but I also think that there’s a lot of limitations to centralized systems. It just is a system property. It’s true of human systems and technological systems. So, first of all, I think I make decisions trying to base it around our long-term mission. So, as a public benefit corporation, we really take the mission seriously, and we try to make sure that we are making decisions that are going to build something that’s actually a better social ecosystem for people. And then I believe in collaboration, so I try to get input from as many people on the team as possible — people who have the most relevant experience in the decision-making — and a lot of things start out as proposals that we talk through and then write down and then get feedback on.

And we operate roughly with a consensus with qualifications model. So I get input, and then if there’s a clear consensus, we go with that. And if there isn’t, I’m the tiebreaker and make the final call as CEO. And then we try to expand when we have the capacity outside of our internal team and talk to experts in the field. Even before we brought on Aaron, we talked to a lot of trust and safety experts, and then we’ve consulted with users, community members, people who are using the app, people who aren’t using the app. 

The app itself is sort of a fire hose of feedback, so we have tried to find ways to sample from that and as well as reach out and have deeper conversations with people, talk to people who aren’t on the app, and then this is how we’re trying to make decisions that are technically sound but also user-centric, talking to other people in the decentralized social space, too.

A lot of us come from backgrounds where we’ve worked on other projects and so we reach out to people that we know to talk about “what about this protocol decision or this design decision.” And then people have a lot of autonomy on the team, so we try to move pretty fast. A lot of what we do is about experimenting and learning, so we try to make the best decision we can at the time that leaves us room to adapt down the road. And so that’s actually why a lot of things we built are composable because we aren’t assuming that we’re going to get everything right on the first pass.

We have this big vision for how all these pieces can come together to produce a good social experience, but if we get it wrong, we want the pieces individually to be standalone so that someone else can take these ideas, recombine them in a different way, and then maybe arrive at a solution that, in the next iteration, is the right way for social to work in this open, composable way.

Can I ask a really dumb question? I’ve now said composable. You’ve said composable twice. Can you explain to the audience what you mean by composable?

Sure. So, basically, have something that comes apart into lots of pieces, like building blocks, like Lego, and then you can put those together and you have a lot of, maybe you only have 10 little different types of Lego pieces. You have a lot of them, and then you can put them together into all sorts of different shapes. And so that means that you can build very different-looking houses out of the same set of Lego blocks.

A few more questions on the structure and how that works. You said you’re 18 people on engineering, 18 on the other side of the company — 30 to 40 people, I’m guessing. The way you’re describing decisions works really well for that size team. I know that because I’ve long run a team about that size, and I feel like I can have an idea, I can communicate it effectively.

You jump up to the next size — 100 people or 500 people — you have to build a different kind of culture. Are you thinking about that now, or are you hopeful that you can just get to the next size? I’ve met founders and CEOs sort of who take both approaches — that “I would love to have that problem. I’ll deal with it when I get there.” But it seems like you’re thinking a little down the road and especially that having decentralization in your brain, you might be thinking about that now. How do you plan to scale the culture there?

I’m thinking a bit about that now. As you mentioned, we already have a lot of things we’re working on and building toward, so I can’t devote a ton of time to it, but trying to talk to people who’ve been there because I haven’t before and learn from people who’ve worked at different scales of organizations and figure out what works and what doesn’t work. So whenever I get the chance and I run into somebody who’s running a different size of org, I’m very curious at this point, asking them about how they do things, how it works. I found it super helpful to just talk to people who are running orgs at similar stages and at slightly bigger stages and find [out]: how do they run meetings, what is their writing culture, how’s their decision-making work?

And right now, I think, just because we haven’t got there yet, I’m still in a learning phase of figuring out what’s probably going to work best when we get there, but I would like to keep some of the same principles of having collaboration. I mean, the end goal is: how do we get to the best answer and best harness collective intelligence? Like collective intelligence in terms of building an open protocol anyone can build on and collective intelligence within the company of, “Everyone here, we hired them because they’re a smart person. What piece of the problem do they see, and how can we put together their insights into something that gets us to the best decision?”

One person who famously experienced a number of trials and tribulations through that scaling process is Jack Dorsey, who started the project with you. He’s still on your board, I believe. Is he involved? Is he someone that you’re bouncing off of, or is he being Jack Dorsey on a cloud somewhere?

[Laughs] I would say the latter, being Jack Dorsey. And yeah, Jack is, I think, a visionary who really wanted social to be on an open protocol, and he has been interested in lots of different decentralized protocols, and he has sort of a big-picture vision of how he wants this to work. And then the specifics of how we built Bluesky was very much the vision that I pitched and that the team has designed. So, the early team I hired, we were people with different experience in different decentralized social projects that came together and learned from the mistakes we’ve made in the past and built something that synthesized those insights.

Do you get any feedback from Jack at all? Does he send one-word emails or anything at all, or is it just kind of silence?

Some feedback occasionally.

Well, early on, we had a sort of discussion around downvotes and upvotes and downvotes or the like button. And early on, we actually put in downvotes, and then we got so much user pushback that we were like, “Okay.” Also, one of the critical problems we realized with the way that we put in downvotes was it’s all public. So because we have all these composable services, everything has to be public for every service to see it and operate off it. So that meant downvotes were public. People really didn’t like that, and we recognize that, you know, it feels bad to see, “Oh, everyone can see all these people downvoted my thing.”

So I think that that was actually an important insight in the sense that social networks, part of the toxicity, I think, comes from there only being positive feedback cycles. And so that means things only really get amplified, and when you really want to say this is a terrible post, you reply to it or repost it or something, and then that actually spreads it further. Whereas systems where you have downvotes mean there’s this downregulation of things, and you can express your disapproval of this by downvoting it, and then if enough people downvote it, maybe it just disappears. And so that’s actually, I think, an important piece, but it’s important that also, when you downvote it, it’s not this big public statement. So that’s something that we would still like to bring back at some point — this downregulation — but maybe not in that exact form. So, we learned from our users.

That sits right at the intersection of “What does a protocol enable?” and “What does the product expose to users??” Maybe the protocol supports downvotes, but Bluesky, the social network that people are on, never shows it to anybody or doesn’t support that feature of the protocol itself. When you think about, “Okay, we should put this in the protocol. This is something we’d like, but our users on Bluesky don’t like it,” does that impact the protocol more, or are you saying, “Well, we’ll get to it in the protocol someday, but maybe Bluesky won’t have it”?

For different things, there’s a bit of both. That kind of depends per design decision. One example here is that we don’t show a likes tab in the Bluesky app, and that was something where it was an intentional decision early on, and actually now, we show it to you. So, since we don’t have bookmarks, you can kind of use your own likes as a bookmark, but we don’t show it to everybody. So, if I go look at your profile, I don’t see your likes. On the other hand, that is public information, so there’s other sites that do show it, but this was just a little bit of friction that we added to say that even though the protocol supports anyone seeing your likes, we think that there’s some downsides to just having everyone’s likes on the profile. So, we didn’t show it by default. 

Another example is we give you an opt-out of the public web interface in the Bluesky app. So all the data is public on Bluesky, all your posts are public, your likes are public, and anyone can build an app that shows all of that, and anyone can look at the fire hose. But if you don’t want a lot of people to click through and see your post, and let’s say one Bluesky post gets embedded in a news article or something, you can opt out of having it be clicked through if someone’s not signed in. And that’s just some extra friction that we’ve added, and we’ve really learned, through building and running a social app, how much these kinds of little frictions and defaults matter in terms of shaping people’s experience and giving a good experience, even if the protocol really enables something that can be done in a different way.

I think this means I should probably split this up into two parts. I want to ask about the app and where you are with the app, and then I promise you the nerdy questions about the protocol are coming. I warn you. If you’re listening to this, just be warned: they’re coming. I would’ve spent a whole hour doing nerdy protocol questions, but I think we should talk about the app first because the app’s on a roll, right?

It’s growing. You have 5 million users. I’m assuming some of the funding you raised is around the success of the app or at least driven by the success of the app. And then, as you mentioned, developers are attracted to the protocol because there are users there. I would say — and I hope you agree — Bluesky has a very distinct culture from the other kinds of social networks that have sprung up in the sort of wake of the Twitter-to-X situation. I’m not going to get myself into trouble by describing the cultures of Threads or Mastodon, but I will say Bluesky is funnier. It’s a little more chaotic. It’s a little weirder. Why do you think that is?

Early on, we had several waves of users come over that were all really big posters. And most social networks, you have this dynamic where you have 1 percent of people posting and then a smaller percentage of people liking and then a lot of people lurking and just looking at stuff. And early on, we had this crazy ratio of 90 percent posters, and so it was extremely active and tons of people firing off shitposts essentially — really fast, funny takes on things and memes and a lot of stuff. And then, since we’ve grown, I think the majority of people tend to use social in a more consumption mode, and so it has shifted a bit more toward lurkers. But still, relative to other apps, I think there’s a lot of posters — people who like posting and engaging that way — and then that results in a pretty chaotic and playful culture, I think.

Do you think that culture comes from moderation? Is it luck? Is it something that you are trying to preserve? Where do you think it comes from, and how do you keep it going?

This, I think, might actually just be a community element, and partially early on, a lot of the team, we all were posters, we all were social media users, and we engaged a lot. We still post and talk to people and stuff. And not just me — quite a few people on the team. The devs talk to dev users. We have people who’ve worked in journalism, and so we have this culture of just active discussion and using the app. 

And then there’s been lots of interfaces for creativity, both intentional and unintentional. So, intentional ones are the ability for anyone to build a custom feed. We created this nice interface in the app. Anyone can build a feed. And so now there’s really niche feeds. I really like the moss feed that just shows you nice woodland pictures of moss, and then you can find a post from some user with five followers, but they posted a moss picture, and now it shows up in the moss feed.

So this is a way where engagement gets spread out beyond one central algorithm into all these little niche communities. I guess another intentional way that’s creative is there’s all sorts of different clients and experiences you can build. So people have built these different experiences that show you things a different way, show you who’s in your social network and all these little visualizations that people have played around with. 

And then the unintentional ones are for a while, we had actually a bug in the app that caused this very long thread to break, and it became actually one long thread called “HellThread” that people were posting very aggressively into. And that bug has since been fixed, but for an era, there was this crazy culture that formed around this bug that almost became a feature. So that was an example of things just being this kind of playful, chaotic energy taking place through both the intentional interfaces for experimentation, creativity, and the unintentional ones, like the bugs that became features.

That second part — the unintentional bugs became a feature — those emergent properties of social networks or communities, it always feels like the most special thing: that the people who run the platforms don’t really know what’s going to happen. The users decide what’s going to happen in ways that are completely chaotic, and the best platforms sort of reinforce that and build upon it. And that is a really healthy feedback loop. 

I would actually put up Twitter as the best example of this feedback loop ever, right? So many core Twitter features were invented by the users. Is that on your mind? That you need to watch your user base and take the ideas into the platform directly and build upon them? Because I don’t see a lot of other platforms really leaning into that anymore.

Totally. Yeah, we do. There’s things that we’ve had on the road map and longer things that we try to stay focused on, like this vision of algorithmic choice and composable moderation. Those are things we prioritize because, from the start, we were going to do that. And then other things, like hashtags. At first, we thought, “Well, I think the custom feeds will allow you to do hashtags within feeds.” People did do that for a while, but people still wanted standalone hashtags, and we got enough takes from users saying, “Just bring us hashtags,” that eventually, we added them.

Same with mute words, like actually, composable moderation is going to let you do something that is much more granular and higher-powered than just mute words. But people want the simple interfaces, too. That’s something we’ve learned. So we added mute words, and we have mute lists and block lists, and these are simpler building blocks. And now you’re going to have bigger, more complex building blocks — the full dashboard of power tools under the hood.

Do you think you’re going to end up in a place where you have what I will just call the Microsoft Excel problem, where so many people have asked for so many familiar features that it’s actually hard to bring them into a new paradigm, like composable moderation or adjustable filters? Every big software product has this problem where there’s an obviously better way of doing a thing, but then people are familiar with the way they’ve been doing it, and so you end up with both or you never have the social capital to do the better way. Do you think about that problem: if the future is feeds in the protocol but everybody wants the brute force moderation of the past, you’re going to end up in that tension? 

Yeah, we actually do think about that a bit, and this is where having some amount of ecosystem guidance from a team that’s really thought about it might help. But also then, if we get it wrong, the goal of the whole app is open-source, and so people can fork off it and sort things out, add things, take things a different direction. And so maybe if our app gets really bloated or if it’s just not working right anymore, somebody who has a strong idea for how to do it differently or do it better can fork it off and say, “Alright, we’re going to clean it up. This is the way we’re going to do it now.” 

So, a lot of the ways that we try to design things is so that we can provide good leadership in the ecosystem. And centralization has its benefits when someone has a plan and they know where they’re going and it’s a good place and people want to follow. But then, if that leadership fails or gets things wrong or changes, then there’s other options and people can vote with their feet and go over to the better option, and new entrants can come in and provide another example of doing something differently, doing it better.

So I feel this in my soul. This is something that we have long wanted across a number of products. We’ve talked about it at The Verge for a long time. I talked to the CEOs of other companies that are in these kinds of relationships with protocols or standards or open source, and at the end of the day, they often come back to “…but we also have to make money, and the best user experience is often the one that we control. So sometimes the protocol slides by the wayside. Sometimes the commitment to open source gets diminished because we have to make some money here. And if we control the user experience, the user experience gets ahead of the protocol, so be it — we’ve got to eat, we’ve got to pay the team.”

In particular, I’ll give an example of this. I think Twitter had a huge ecosystem of third-party clients. It was a playground for developers. So many cool things happened and were built. So many tools were built that had nothing to do with posting or reading tweets at the end of the day. And [then] they had to monetize, and they realized they couldn’t put the ads in all those clients, and they shut down all those clients. Do you see that problem coming? Is that, “Oh boy, we’re going to have to monetize this thing somehow, and we’re going to have to reclaim some control to do that?” Or do you have another idea for how you’ll manage that tension?

Yeah, we do want to make money, and we are trying to do our best. Part of this is sort of tying yourself to the masts, like Sisyphus-style, of “this is the ways that we’ve committed to openness.” We’ve already open-sourced the app, and we’ve open-sourced everything you need to run your own version of Bluesky on the app protocol. And we have built into the protocol ways for users to move all their identity and relationships and data around. And so the idea is that each user has all their data as this package. We actually call it a repository. If you’re familiar with Git and GitHub, it’s like your data repository is like your GitHub repository. And then a lot of people use GitHub, and that’s a centralized site, but then they can take their Git repository and move it to GitLab or some other site.

And so, that’s the built-in “vote with your feet” piece, and you can take your stuff and move it. And then federation is an important piece that we just started for self-hosting federation a few weeks ago, and that means that you can run your own server, host your own data, and then if we try to close things down or whatever, you can just say, “Well, now I’m just going to use a different service because I control my own data and it’s on my own server. I control my identity, my data, and where it’s hosted.” And then that means if a lot of people are federating and self-hosting, that ecosystem can just shift around piece by piece to the services that haven’t shut down their APIs and that haven’t let users feel betrayed.

Those are the things that keep you from making money in the bad ways. What’s the plan to make money?

We’ve been building marketplaces within the app, essentially. So, we’ve got information marketplaces, moderation marketplaces. This is a direction that we’re going to lean into. We’re also providing a service, and we’re already making a small amount of money through one step we’ve made into custom domains. So, the idea here is you can use your website that you own, like jaygraber.com, as your user handle, but you have to be technical to set that up. You have to go buy your domain. You have to change the TXT record of your DNS settings, and a lot of people don’t want to do that or don’t know how. And so, then we sell you a domain through us. We’ve partnered with Namecheap to do that, and then that makes us a little bit of money, and then that’s an example of the kinds of services we’re going to expand into overtime.

Do you think that that is infinitely scalable? I have a quote here that one of our community people at Vox Media gave me. Mark Zuckerberg talks about Facebook being a “community of over a billion people,” and my colleague said, “That just makes me cringe. There’s no such thing as a community of over a billion people.” Is that something you think about: “We should have these smaller communities, or we can preserve the vibe and the funniness of Bluesky now even as we scale because it will splinter naturally with federation and decentralization”?

Yeah, there’s lots of joints on which it can splinter. There’s already feeds, and there’s going to be the moderation services. These will all create different experiences just within even the Bluesky app and are already creating different little corners. Then there will be different servers. There will be different apps running different entire pieces of the network. So, that will, I think, cause a lot more variety in the ecosystem. Part of the idea of having an underlying protocol is that, actually, social networks systems tend to go through periods of consolidation and then fragmentation, and a protocol that connects them all lets this happen naturally.

So, the example is the web itself is a protocol. Websites come and go. Some of them consolidate a lot of users, and then over time, they get worse and worse, or they shut down and then users go somewhere else. And that consolidation and fragmentation happens on the web. So this is trying to do that for social, so sites like Bluesky can come and go and grow, and then if we someday aren’t doing the best anymore, something else can come along, and users can fragment out and then maybe re-cohere around someone else who’s doing things really well.

You mentioned the plan for monetization is around the marketplaces for algorithms and feeds. I assume that will take place inside the Bluesky app, the one that you control, right? That makes the most sense. “Here’s all these users. You can buy an algorithm from us and see what you want. Maybe it’s about moss. Maybe it’s something else.” Is the plan to make money inside the app, or is the plan to make money with the protocol?

There’s ways to explore both. Right now, we have a lot going on in the app that we’re really focusing on, but also, I think one insight is we’re trying to make social more like the web. And early on, people didn’t know how the web was going to monetize, but it definitely did. Services definitely monetized on the web — tons of services that made lots of money built on open protocols and nothing that’s gotten to massive scale on the internet has ever just totally failed to monetize.

So, that’s something that we just think we need to show people, first of all, that this vision is something that can work, that is fun, useful, good, better than the alternatives, and also good for society, creating healthier social media ecosystems. And then we really believe that money follows value. If we can demonstrate this value and people really want this, then it’s something where we’ll be able to make money offering services in this ecosystem.

Services on top of the protocol, though, not the protocol itself?

The protocol itself is not a blockchain, not something with any built-in monetization mechanism. It’s like SMTP for email or HTTP for the web. It’s just a way for computers to talk.

That’s a great preview of my next question. Well done. Those protocols — HTTP [and] SMTP — they’re maintained by standards bodies. Some big companies volunteer some people. They don’t make money themselves — they’re maintained by a bunch of people who are invested in those protocols existing and growing so they can put services on top. That’s more or less what you’re describing. Would you ever turn over AT Protocol to a standards body?

Yeah, and we’ve already begun talking to some standardization bodies — like starting the very early stages of that work, socializing the idea, taking on the pieces that are relatively more solid, as I mentioned earlier. The AT Protocol is actually made up of several pieces that the way identity works, the way data works, and these pieces, we aim to get standardized and then stewarded by governance bodies.

Alright, get ready. It’s time for some extremely deep standards drama around protocols. I told everyone it was going to happen, and now it’s here. So, you’ve got a protocol: AT Protocol. There’s a competitor protocol called ActivityPub. Mastodon runs on ActivityPub. Threads is going to federate using ActivityPub. What, in your mind, are the pros and cons? What are the differences?

So, ActivityPub was around when we got started. I did an ecosystem review of all the centralized protocols that existed in 2019, including ActivityPub. And we looked at it and decided that we needed to build something different because there were some critical pieces that we thought were missing. So, one of them was around the composability and the interfaces for composability we’ve designed. Like the way that we do custom feeds and all these moderation labelers, that’s really not possible with the way that things are very server-centric in ActivityPub right now. Your server is very much your community where all of this happens, and we have split things up into these microservices on the backend, mirroring a bit more of how a large-scale social network works with the global feed. 

The other thing was the global feed, like having global search and discovery be a first-class thing that we were building for. And also, having all of our users at the start know that this is public data and it will be remixed in all sorts of ways by global feeds is something that, both technically and culturally, we had to design for. Because ActivityPub has clustered more around servers that they federate and they talk to each other, but there isn’t a service that scrapes all of it and gives you this big fire hose, and even if it’s technically possible to build, there’s been community resistance to people doing that. Even someone in the community who was building a bridge between AT Protocol and ActivityPub recently got a lot of pushback from the ActivityPub community on not wanting that bridge. So, that’s just sort of both cultural and technical. 

Then another thing was we really wanted to get account portability. So, this ability to leave with your identity and your data and have fallbacks with the way that we’ve designed your repo, you can even back up all your posts on your phone or back it up on your server that you control, and then you don’t have to have any sort of friction when you want to move. So, you can move between services in ActivityPub. But if… for example, Queer.af recently, their .af domain was seized by Afghanistan, and then people were stuck because there was no warning, and then they have to rely on their old server to help forward their stuff over to a new place. So, we wanted to get around that problem and make sure people always had the ability to move. 

Then, we wanted to have good UX. There were just a lot of complaints with users around the UX of Mastodon, and we wanted to provide something that was more just tailored for the mainstream user who was used to Twitter, who could come on and not have to worry about, “Oh, what server do I pick when I sign up? What does it mean? Is this going to shape my experience forever?” On Mastodon, you have to know that right when you sign up because it’s going to shape your experience. Here, you can sign up on our server, which is the default lobby or gateway into the ecosystem, and then you can move to another server and shift to your own. So, when we opened up federation, several of our team members moved their main accounts off onto their own server, and it was pretty seamless. Nobody notices that people are running on their own servers, and it’s still just all one experience in the app.

Let me be very reductive here and see if I got this right. It feels like the first order bit for Mastodon [and] ActivityPub is the server, and you’re saying your first order bit is the user, right? You have a user account; it can kind of live anywhere, and that’s the thing you’re in control of. On Mastodon, there are servers, and the server administrators are more in control. Is that too reductive? Is that accurate? Is that a good way to think about it?

That’s actually a pretty good way to think about it. And this is where, earlier on, since we’re getting technical now, I said we were borrowing from peer-to-peer systems. Peer-to-peer systems really try to do everything at entirely the user level with no servers. And that actually creates a complicated user experience, but that means that every user’s fully in control of all their data all the time. We have the ability for users to do that, but we have servers to make your life easier, so you don’t have to have your phone be trying to directly communicate with 5 million other phones or a computer trying to do that. But then you can always move off of our server and move off of these services.

So, the property that we really tried to get out of this piece of decentralization is the right to leave. And so, you can use a service and can use even a big service, but you always have the right to leave built in because we’ve designed around the user. And the other way we’ve designed around the user is not just at the technical level but at the UX level. So, try to design for: what do users really want? How do we get that sort of convenience and ease of use for the user? And thinking from a UX as a first-class thing that we’re designing for.

There’s some tradeoffs there. Some are really obvious, right? Mastodon server administrators often find that they’re running a server and they have software costs and overhead, and they have to maintain a server for a bunch of users. At the same time, they’re in charge, for better or worse. If something goes wrong, they can shut it down or they can delete stuff. They are in control of the content. They can set different kinds of content moderation policies. There’s a decentralization of authority in that way. It feels like AT Protocol is much more individualistic in that way.

ActivityPub points you toward groups of collectives that may wish to interact with each other. And AT Protocol and Bluesky are much more like, “You’re on your own. If you like it, you can leave, and there’ll be a larger market of individuals.” How do you think about, “Okay, if someone doesn’t like the server they’re on, they’re going to leave”? They might not actually be sending any signal at all that the server administrator was doing a bad job because you can’t tell. Whereas on, I think, Mastodon, it’s like, “Oh, somebody just left this account behind.”

I think, actually, collectives are going to emerge, and they already have in some ways. Collectives aren’t formed at the server infrastructure level. So, one of the ideas that we had was, on Mastodon, your moderation is very much tied to who runs your server, but often, the type of person who wants to do DevOps essentially and manage a service online is different from the type of person who’s a community builder and wants to organize a group of people together to set a different set of community norms and moderate. You can get those people together to run a Mastodon instance, but it has to be tied around the instance, so we separated those pieces. So, now you can get a group of people together and run a moderation service. We give you software to manage reports and stuff like admin tooling, and you can do this with the whole team, and then you can get a bunch of people to use it, and then you can get a whole community based around a moderation culture that you’ve created.

Then, you can do that with the feed as well. And then all these things, as I mentioned in being composable, can be bundled and unbundled in different ways, so you could recreate the Mastodon “everything is tied to your server” experience by setting defaults on your server: this is the moderation regimes we’re going to be using, here’s the labelers we’re using, here’s the feeds we’re using, etc. And a server operator still has power at the end of the day because they control where your stuff is literally hosted. But then there’s all these other interfaces beyond that where control and agency can happen. So, any service in the network has the ability to set rules, and then users have these interfaces where even if they’re non-technical, they can go in and create rule sets.

You mentioned the controversy around the bridge software. It was called Bridgy Fed. This was some of the most intense GitHub comments I’ve ever seen in my entire life — truly out of control for basically a piece of middleware. Why do you think that happened? That the idea of bridging AT Protocol onto ActivityPub led to one of the most intense recent developer flame wars that I can think of? Is it the cultures of the protocols, the culture of Mastodon versus Bluesky? What, in your mind, led to that conflagration?

I think it’s a big part of the culture of Mastodon. This was, as I mentioned, one of the reasons that we didn’t try to get ActivityPub to change toward the direction of what we wanted to build because not just the technical primitives being different, there’s also this culture of resistance to global feeds and global algorithms. And that means that people who had tried to do a search engine for all of Mastodon in the past or things like this had gotten shut down even before Bridgy Fed in the past. I had seen that even back in 2020, 2021.

I think it’s just a continuation of that culture. And the key thing is even if you have a protocol that lets servers talk to each other and federate, part of it is human governance, and if people who run that server don’t want to talk to you, then even if the computers can talk, they don’t have to talk. And so that’s essentially people in Mastodon expressing their preferences of the kind of communities they want, which is, they don’t want to talk to Bluesky users, so they can just not federate. And then there was a lot of discussion that was around opt-in versus opt-out. So, it’s discussions around what kind of governance norms do we want to set as protocol communities.

Part of the argument in favor of the Mastodon culture that I’ve heard over and over again is it allows safer spaces to form for certain kinds of groups — that the servers can be closed. It is mostly opt-in that people defederate all the time in the Mastodon community. The number of servers that have defederated from Threads or have sworn a blood oath to never federate with Threads, it’s very high. Do you think about Bluesky as providing the same kinds of tools that make people feel like the spaces are safer?

That has been our goal: to give users the tools to build spaces that are as safe as they want. And I think one thing that people have said the protocol is missing right now is private accounts. That’s something we’d like to get there, but so far, in our mission statement, we’re focused on public data and the open Twitter model, and there’s other protocols out there for private communication. Like there’s Matrix, Signal, and other protocols that do private communication. So, we focused on the hard problem of global public conversations and algorithmic choice. But even within that, we want people to have the ability to feel that in Bluesky right now, they’re participating in public data, public posting, but you want to not interact with people who are going to harass you. You want to not see stuff that’s going to be damaging. 

So, you want to be able to partition off your little piece of the network. And so, we’ve tried to give people all the tools to do that, not just at the architectural level of here’s servers that you can run and here’s services you can host but also at more user-friendly levels, like the user lists and block lists and the ability to run your own labeling service where someone non-technical can do that through software that we built for you to go in and just say, “Alright, I want to start setting rules and filtering out this kind of content for me and my community.”

We’ve mentioned Threads several times now. Obviously, Meta is starting Threads. When you think about onboarding the non-technical user onto decentralized social media, Threads seems like it’s way ahead of the curve. It has 100 million users. It’s going to use ActivityPub, right? Meta has already started testing it. I think Adam Mosseri’s account can be followed on Mastodon through ActivityPub. Does that represent a competitive threat to you? Is that something that you’re watching, or are you trying to build something else?

I think it’s really interesting that the ecosystem is moving this direction overall. I also, at the end of the day, want everything to be on an open protocol, and I think that our protocol is the most resilient long-term and is going to give the most flexibility and guarantees of this being a healthy ecosystem where people can move around. But also, I’m just excited to see things starting to move this direction. Mike Masnick, who wrote “Protocols, Not Platforms” as a paper, said that he shopped this around to all the big tech companies at the time, and this is one of the things that convinced Jack to do Bluesky. So, Twitter listened, but Meta was like, “No, this is never going to happen.” But since then, they seemed to have moved on from that position, and maybe we were one of the things that have started moving things.

I think that overall, it’s a healthier direction to go toward open protocols. One of my concerns about ActivityPub and the Threads model is, because it’s so server-based, if you have one massive server come on, like the Threads server, things really centralize around that. And so, it might functionally not be very open because people are tied to the server and it’s hard to move, and it’s going to be really up to Threads whether they ever become fully protocol-compliant and let people move off Threads easily. Then, is there going to be that composability there? Are users really going to have that much control? Those are open questions, and we’ll see where Threads takes things. 

But I think that we’ve built that in from the start, this openness and composability. We’re also open-source, and so that’s the other guarantee. Our code is open-source, all these interfaces we’ve built. That’s not the case at Threads. It’s going to be something that, if it becomes fully protocol-compliant, I think there will still be risks to users.

One of the reasons you can see why Threads might prefer ActivityPub, just based on this conversation, is it’s very server-based. I’m guessing Meta loves the idea of running the big server. There’s some benefits to that, too. Meta has a giant compliance department and a huge content moderation team. There are parts of the Bluesky approach that seem like it’ll be very challenging to run at scale. Just really dumb — the data is everywhere. It seems undeletable because it’s so public. If there’s a copyright claim, can you actually pull something down? Can you get rid of it?

Ultimately, data has a place where it’s being hosted, so whatever server that you’re running, the server runs in the jurisdiction where you posted it, and whoever’s running that server is going to have to be responsive to their legal jurisdiction. So, there ultimately is a host for data. It just moves around more.

Are you thinking about that stuff, like, “Boy, we’re going to run in a bunch of jurisdictions, and the Indian government wants Bluesky to have an actual person in India that it can arrest if they don’t get the moderation rules they want”? Those seem like very thorny issues for a company that’s very early, especially one that has to explain, “Hey, maybe we don’t have all the controls that you’re used to from a TikTok or a YouTube.”

This is one of those things where I’m really glad we just hired Aaron because he has experience–

… with platforms at scale and government requests, and our goal right now is to grow and get to that kinds of scale where these will be a good set of problems to have, but bringing in the expertise early so that we can be prepared to deal with it and someone who really understands our whole architecture and can help us navigate those questions is where we are right now, in terms of preparing for that.

One of the things you talked about rolling out soon is some of the composable moderation. I read your blog post about it. Again, just to be reductive — and tell me if I’m wrong — it seems like the core idea here is more people should be able to label more things in more ways. We’re going to allow the users to say, “Here are the labels, and we’ll get the labels out there.” It’s an election year. I feel like the election-year experiments with labels or the covid-year experiments with labels have taught us all a lot. What are you pulling from all of those previous experiences with big platforms doing labels to say, “Okay, our big idea is labels, but decentralized”?

I think Birdwatch was this project that we thought was really cool on Twitter, and seeing Community Notes come out of it was a success in terms of them pioneering an algorithm that could usefully augment and annotate information out there. And, basically, the interface we built is for a thousand algorithms like this to be experimented with, and for some of them to be complex algorithms that were very specialized algorithms like the Birdwatch algorithm, Community Notes algorithm. 

Or some of them to be very manual, like people are like, “I’m an expert in this thing. I’m just going to label for this thing.” And so then that allows all these things to come together to an experience that, again, it’s about harnessing collective intelligence, letting experts, people with local context, move faster on these things than the company whose interests might not be entirely aligned in solving this niche problem or this jurisdiction or that’s just too slow because it’s a big company or it’s a small team or whatever.

So, our goal is to let that whole ecosystem just iterate and experiment, and then we try to have some amount of leadership in terms of what we’re encouraging people to build, how we’re creating and surfacing the best stuff that gets built and bringing it to user’s attention and helping them install it. I think that it is going to produce a better system, and there’s some ideas right now in the academic world of this concept of moderation middleware that they’ve been designing and thinking around doing this within the centralized platform paradigm of offering, “Hey, instead of the only way to make change at these companies is to have nonprofits, interest groups, governments lobbying the company for change, maybe you could create a middleware interface where people could directly add some of their input into things.” Composable labeling is like middleware but as a first-class thing built in where anyone can build something like this, and because it’s fundamentally open, it’s taking that from the fundamental premise rather than trying to add that as an intervention into a centralized system not designed for it.

Can I ask you a question that I always ask the AR companies but, I think, plays in the world where we have different kinds of labeling systems? I can imagine a world where someone posts a picture of the US Capitol, and for an AR company, this is very difficult, right? You’re looking at the United States Capitol, and your Apple Vision Pro has to tell you what happened there, and you can either put January 6th or you can put the passage of Obamacare — two things that would enrage different people in very different ways.

If you have composable moderation, you have different kinds of algorithms, appending different kinds of labels. You still might break that shared reality. Someone might post a picture of the US Capitol, and one person might tell you, “This is where Donald Trump did a coup,” and someone else might say, “This is where the election was stolen from Donald Trump.” Are you worried about that sort of breaking of reality, or do you think the market will actually converge on the truth?

Both. I think that basically — to go very meta with this for one moment — I think that every new form of technology that comes along causes a lot of disruption, particularly information technology. So, when the printing press happened, people said, “This is going [out] in the world. We’re going to have heresies like revolutions.” All those things did happen, and there was a lot of chaos for a while because the shared truth of the old order broke down. But then eventually, we developed new institutions that built around the printing press, and we developed new ways of understanding the written word and arriving at truth. We developed encyclopedias. We developed the academic citation system. We developed a legal system that’s very text-based and builds off previous texts, and so we developed all these systems that helped us create a society that works.

I think the information era has once again totally fragmented shared reality that the printing press had established and the written word and then broadcast media, and so now everyone can broadcast on social media. And how do we develop new kinds of systems where we can come to shared truth again? We’ll have to develop systems that [have] new algorithms like community notes, that kind of approximate shared truth from all these people broadcasting into the web.

And how do we do that? Is it going to be a centralized effort where one company just funds and develops that and gets the right answer? Or are we going to open it up to experimentation and then maybe lots of people going at the problem start to arrive at the right answer? And our approach is the latter, which is opening up the system, having lots of people experiment on building new ways of curating, like arriving at approximating truth, pulling lenses apart to look at this shared information and then combining them again will approximate at some point. People will start to converge around new ways of doing things. 

So, we also, as a company in the ecosystem, will be trying to approximate: what is a good way to do this? What are the best labels out there? How do we bring those in? Maybe if we’ve talked to a reputable fact-checking org and they’re running a really good fact-checking service, we’ll promote this in the app, but then someone else could build a different app, and then how is the ecosystem going to evolve? Well, it’s an open system, and eventually, at some point, we’ll probably reach a point where we know what the good sorts of ways of arriving at truth are and the good institutions, like providing objective stuff or stuff that gets things right most of the time, and we’ll converge around that.

That’s a very high-minded idea. You have to fund it. It sounds like your plan to fund it and be sustainable is to operate those marketplaces to take a cut. So, I buy an algorithm from you, and you’ll take a piece of that transaction.

Yeah. As a company, but then also because this is an open ecosystem, it’s not just limited to our funding model. We are in the early stages of building out this vision, and so we need to keep going, prove out the vision, show this, make money as a company, create a marketplace, but then also somebody running one of these services on its own, already people building clients like other services in the ecosystem are… some of them are charging, some of them are taking donations. I think some things will end up nonprofit-funded. If you look at things like email, you have big companies that run email services. You also have nonprofits that run their own email servers, and you have institutions like your university might give you an email, or your company might provide you an email.

And so, this is something where if we show value, once again, money follows value; we show enough value in this ecosystem. Some of these might become institutions funded by nonprofits who are interested in a better information ecosystem. They might fund moderation services, might fund feeds, things like this. Institutions might run them. Communities might get together and self-fund. Some of them might be paid services that payments happen outside of us. Others might be paid services where the payments happen through us.

Do you think any of them will be businesses unto themselves? I’m reminded of this Bill Gates quote about Windows as a platform, where he is like, “A platform is when the economic value of everybody that uses it exceeds the value of the company that made the platform.” So, all the people who built companies on Windows, you add them all up, they exceed the value of Microsoft. You’re kind of talking like that, right? You’re going to build a platform. You’re going to let people buy and sell on it and build products on top of it. Do you see someone saying, “Okay, there’s an opportunity to build a great content moderation company that runs on top of AT Protocol that Bluesky users can buy”?

We’ve already seen some people start to move in that direction. I think that is something that’s going to happen.

How long until that happens?

I wish I could see the future, but I don’t actually have that ability.

You’re probably better at seeing it than I am. I’m very curious. This feels like there’s some inkling of “this is how it should work,” but you’ve got to train a whole bunch of people to think differently, and they’re right up next to Elon Musk being like, “I’ll just do it for you. Just show up and tweet.” That seems like the hard part, right?

Right now, a lot of what we’re doing is trying to communicate the vision, like talking to you. Earlier on, before we built stuff out, yeah, like last year, I was turning down a lot of media because we had to build this out, and we really want to show, not tell. Now we have examples that we can point to, and it makes it a lot more real rather than me just sitting here expounding on the printing press. It’s very high-minded and not concrete. But now we actually have examples of many different feeds and many different services that have been built in this ecosystem, and so one of the things we’re starting to do is to fund the developer ecosystem.

We’ve started a grant system, and we’re going to start giving money to some projects just to help boost this ecosystem, but I think over time, these funding models are going to shift and change, and we’re going to experiment with a lot of them because a whole ecosystem that’s like the web means a whole variety of experiments like the web itself kicked off, and I find that really exciting because social has been in a period of stagnation because it’s consolidated in a few companies, and early on, people built all sorts of Twitter clients and experiments on social, and even Facebook was more open back in the day, and then that whole ecosystem got closed down when these companies shut down their APIs.

But if we open that up again, I think social will be [in] another period of experimentation. There will be lots of companies and hobbyists, just people trying things and playing around again. It’ll be like a playground era of social but in a mixed iteration where we’re aware of a lot of the problems that happen through platforms and mass broadcast at scale. And we have a lot of new tooling, like machine learning has moved on a lot since the last time we had a lot of growth in social, so I think a lot of interesting things are going to come out of this era of building.

Let me end by bringing us all the way down on the ground. You have a lot of big ideas about protocols and platforms. (I’m obviously obsessed with them.) But you run a social network that’s growing in an election year. You’ve had to deal with some very real content moderation controversies on Bluesky already. There’s been some real problems with racism. There’s been problems with death threats, and you’ve just had to do some direct moderation, just take some direct moderation decisions. 

It’s an election year. It’s going to be weird. There’s going to be a bunch of AI deepfakes of both candidates coming. Do you think your new tools are going to let you address that stuff in a new way, or are you going to have to go back to the old ways and just take some direct moderation decisions as you’ve had to in the past?

I think the new tools will help a lot, and we’re also not relying entirely on them. As with everything we’ve done, we’re doing a hybrid approach where we are running our own moderation team with experienced leadership and we are going to do our best, and then we also have all these open interfaces where other people who are more experts or people who think we’re not doing it well come in and do it themselves, and then we’ll stitch a hole together out of both all the open experimentation and the decisions that we’re directly making and what we decide to carve out as, like, this is stuff we’re doing really well and that we think the foundation we need to set for this to be a good experience.

Are you worried about the flood of AI-generated garbage on every platform? Is that coming for Bluesky right now?

Not quite yet. I mean, we’ve been dealing with spam and all the other problems that networks run into, but we’ve also been building our own automated tooling, and it’s a constant cat-and-mouse game, but we’re trying to stay cutting edge in terms of how we address these things using a combination of human moderation and automated tooling to catch stuff. And then also, part of the open interface is we’ve already seen some cool experimentation going on. We had a hackathon a few weeks ago at [Y Combinator], and there were some groups that made deepfake detectors using state-of-the-art models that do deepfake detection, and so it’s this cat-and-mouse game of AI is being built to create deepfakes. Also, there’s deepfake detectors now, and so those things can come together into things that, if we don’t pull them in directly, they can run as independent services in this composable moderation thing, and if they’re running really well and super critical to the ecosystem, then we’ll find ways to integrate them more into the app experience.

Are you seeing that already taking place on Bluesky, or are you just watching this sort of Taylor Swift deepfake problem on X and hoping it doesn’t hit you before that stuff is ready?

We haven’t seen a lot of it yet, but we’re preparing to deal with it through both our policies and training, making sure everyone is aware and then also making sure that we’ve gotten this moderation tooling built out, which you’ll be able to see in action soon, that’s going to let people start to make these interventions.

Alright, last question. It feels like the people who run the more centralized platforms have lately become utterly beholden to their own posters. Like, Elon Musk seems very responsive to the posters on Twitter. Adam Mosseri, in the middle of the night, is like, “I don’t know. I’ll look into it.” You see that reaction, like, “I’ll be the face of the thing.” To some extent, you’re the face of Bluesky. You’re here doing this interview, but you have an out, right? It’s come up several times in this conversation, like, “I don’t know. We’ll see what the marketplace does. We’ll see what the community does.” Do you feel that pressure to be the face and to be responsive, or do you think that that out will actually take hold?

Yeah, I do, and I’ve accepted it to some extent, and I’ve also told people I am sort of the steward of this ecosystem right now and the creator of… let’s think of it as a city-state in an ecosystem we’re trying to build. My bio on Bluesky is, “Let’s build a federated republic starting with this server.” And so, there’ll be lots of other federations out there. I encourage you to start one, and I will do my best on this one. 

And then, I also am not… something we joke about on the app is, like, “posters madness” is a real thing that people can get on social media. We try to do some internal awareness on becoming audience captured by our own audience and stay aware of that. And then also, I try to do my best to — I’m a poster — engage with the community but not be the main character all the time. There was a joke at Bluesky I found very funny, which is, “The one thing I love most about this app is I don’t know who the CEO is.” Just like, great. If it’s a thriving enough community where it doesn’t even need me as a main character, that’s great.

Well, Jay, that is an incredible place to end it. You’re a great main character, by the way. Thank you so much for being here.