> Every single time a post about atproto hits Hacker News, somebody asks in the comments: “But where are all the Bluesky instances?”. The problem is, there are no instances in atproto! The question is a category error. Instances are a Mastodon-brained concept, and I wanted something I can link to that explains this clearly.
I feel like you've (perhaps purposefully?) misinterpreted "instances" just to plug ATProto specifically at the expense of ActivityPub (and RSS, a bit). I think you lower yourself by doing this:
1. it forces you to omit and contort the interesting technical truths about ATProto and Activitypub, like Relays and their pros/cons for ATProto and account migrations and pros/cons for ActivityPub
2. it creates unnecessary conflict and criticism and seems unnecessarily divisive for 2 platforms solving problems in such a similar space
It's also just seems a bit silly: why would you assume that when someone asks "where are the instances?" they're not using the common mainstream use of the word "instances", like, servers, or running software, or VMs, or containers?
Sorry if this is overly harsh or I've misunderstood, but it gives me a strong vibe that it was motivated by disdain and frustration towards ActivityPub and ActivityPub users rather than wanting to legitimately inform the world about ActivityPub.
I did enjoy the diagrams and the explainers though! I just felt like the subtle digs and pops at activitypub were an unnecessary distraction.
> why would you assume that when someone asks "where are the instances?" they're not using the common mainstream use of the word "instances", like, servers, or running software, or VMs, or containers?
Of course depends on the context, but in a lot of discussions about ATProto, ActivityPub, Mastodon and nearby areas, people talk about "instances" as in "ActivityPub instances that host my data and my profile uses its URL as a 'name'". The blog post is specifically for that context I think.
It's less about trying to hide around the issue, and more reframing how you see the concepts, as people start to associate words with concepts and structures. So when people talk about "decentralized social media", lots of people think about ActivityPub, which typically (always?) has a kind of federated architecture, and the instance is one of those nodes in the network. When these people see ATProto, instinctively (and perhaps rightly so) they literally ask "But why is there only one Bluesky instance that people join?" as those concepts map close to what they know.
Overall I think the post is a good and useful addition to the discourse, with perhaps not a completely novel perspective, but posted publicly for future reference when this inevitably gets asks again sometime in the future, specifically for the people who have these previous associations already formed in their head.
I'm being a bit cheeky in the article's tone but I am fairly confident from discussions in the past that "But where are Bluesky instances?" is a common question which usually demonstrates a misunderstanding of the architecture where "having instances of an app" is seen as a measure of decentralization.
My article was an attempt to dig at this specific misunderstanding by comparing it to "But where are Google Reader instances?" which I think illustrates its absurdity. I genuinely do think that the two pictures I provide close to the end clear this up in a way a lot of early atproto/ActivityPub discussions completely gloss over.
Re: Relays, I wrote here on why I didn't include them: https://news.ycombinator.com/item?id=48600963. They're kind of incidental perf optimizations rather than essential to the model. In the post, I wanted to focus on the model.
From my perspective, I care about the centralization/decentralization aspect a lot, and if I'm coming into the discussion with a much better understanding of the Mastodon side then _of course_ I'm going to ask about the instances--that's the vocabulary I'm going to use to try to probe for flaws and gaps. It's not necessarily that it's the instances specifically I care about, or that I'm somehow technically misguided.
What I hoped to read in the article is how we approach topics like centralization, censorship, moderation, data ownership--and with a technical lens. But I feel like all I got was "here's why instances are the wrong vocabulary" without substantively talking about the part I personally care about and want to marry the technical understanding with. Maybe I just read too shallowly and need to sit with it.
I see! That's a huge topic by itself since you're raising a lot of questions. Maybe this could be a different article. My aim with this one was just to clarify the network topology because it is a prerequisite to having the other discussion, and too often that prerequisite is not there.
If you ask a list of specific questions, that would help a lot. I might be able to write something or reply inline here.
I appreciate that a lot! The article has a deliberate and explicit scope, and covers it well.
I'm hoping that perhaps my personal perspective shades why "instances" comes up, or why the reaction on HN seems to include the wider scope than the article itself covers.
Thanks for the fair response, I agree you're being cheeky. Sorry, I'm being lazy not searching here, but have you written anything on if instances of something is a good measure of decentralisation? (FWIW, I feel independently owned/managed instances in the traditional non-mastodon-definition seems like an okay measure of decentralisation.)
I completely agree with the point in your link that relays are different to instances - I love architectures involving dumb-relay or zero-trust type nodes. But I think Relays should still be mentioned in your post, since they're probably the main architectural element which protect PDS instances from the scale issues heavily federated AP instances might face, right? (I only have a high level understanding of ATProto and very little experience with AP, happy to be told I just need to learn more for this to make sense.)
It's a comparison they are directly inviting, by constantly claiming it's decentralized. And then its defenders get upset when people rightfully point out that there is only a single instance, because that single instance going down takes the whole thing down. Like in Google Reader.
I found the distinction and comparison about Mastodon and ATProto are necessary. The fediverse model is easier to understand given existing social networks. ATProto is a novel concept that give users data sovereign and also the scalability of the centralized social networks.
I agree, a comparison and distinction is helpful, maybe even necessary. But I felt the author's bias came across a bit too strong in places and was a little distracting. Still interesting stuff though!
I think the analogy presented here is broken. RSS doesn't depend on Google Reader at all. Even at its prime, RSS depended less on Google Reader than email depends on Gmail now. In ATProto, AppViews heavily depends on Relays to be useful, and Relays are quite expensive to run. Also, the yellow circles which represent blogs in the RSS illustration are really not of the same nature as the same circles which represent posts on Facebook. Blogs are self-sufficient, for example.
I'm not saying ATProto is bad at all, but I feel like this blog post adds more confusion than it clarifies anything.
Relays are actually quite cheap now! They used to be a bit more expensive when they archived all the traffic, but in sync 1.1 that was dropped and they can be run on $20/mo VMs pretty trivially now
While relays are among the more intensive parts of AT Protocol infrastructure, their cost of operation is still something most people can afford: approximately $30/mo now. What is truly expensive and difficult is something that will be immutably so regardless of how centralized or decentralized you are: moderation.
As far as I can tell, Relays[1] are the glue that makes ATProto work performantly. I think they're supposed to be content-agnostic — they just shuttle data through, reducing the number of services each AppView needs to be aware of.
As the blog mentions, the big improvement vs Mastodon is that Relays, AppViews and PDSes are separate services with their own distinct scaling demands. It's a rather beautiful solution to a system design problem.
Yeah, Relays are one way to do that. I've mostly skipped them because they're an invisible optimization and there are other strategies. E.g. many smaller apps today rely on Constellation (https://constellation.microcosm.blue/) instead of building their own database index, so they don't use a Relay at all.
They do remove content directly from relays. They claim they only remove content that is illegal to host, but I don't know how true that is, and there is always the risk it could change in the future.
I want the bsky org to be able to choose what content they host (and I think the internet would be a better place without section 230 protections allowing hosts to ignore the content they distribute); the promise as I understood it was that relays could be hot pluggable. If someone stopped carrying content (maybe it was illegal in /their/ region and not yours) you could failover to another relay.
However there is very little incentive to mirror any of the firehouse if someone else is doing it for free.
An important distinction is that blogs have their own websites and they're not required to publish full articles in their RSS feed.
Bluesky doesn't normally work that way - everything in the PDS gets replicated. They are also encouraging people to put put full blog posts in the PDS for easy replication. So, anyone who wants to index it gets a copy and you have no control over what they do.
You don't have to do it that way, though. You can publish your blog on your own website and just publish links to it on Bluesky.
Honestly that’s just as much because atproto is a raw data protocol. Putting an http frontend on an atproto account is something we encourage and a lot of folks do. I do that on pfrazee.com for instance, and my leaflet blogposts (which are canonically on atproto) render on my blog.
Web pages aren’t digitally signed, aren’t necessarily indexed by search engines, and there are ways to block bots with things like captchas. You also have much more control over the UI. If your blog has comments, you can moderate them, for better or worse.
With a PDS, the replication happens first, before anyone reads it, and the UI is out of your control.
Maybe that’s okay, but people should understand the tradeoffs.
ATproto sacrifices true decentralization for consistency, Mastodon and AP does the opposite, sacrifices true consistency for more accessible decentralization.
At least that's how I understand it, because running an AP node is much more accessible to regular selfhosters than running one of those content relays in AT.
So all you'll ever "decentralize" in AT is your own data, it's more about owning your data rather than collectively owning a part of the network.
I think that’s a part of it but doesn’t state it fully.
AT doesn’t just give consistency, but a shared data model across apps. So apps can reference and render content from other apps. It’s really kind of like a web of typed JSON. Different apps are lenses through which you can see the same network. Anyone can build new experiences on top of old data. There’s nothing remotely equivalent in AP.
AP couples data to apps. In AT, it’s more like there’s one global database with entire world’s data that every app can query.
I don’t understand why the discussion always bumps into Relays. Running a Relay if you want to is cheap-ish ($30/mo) these days. There’s multiple existing ones (Bluesky or community) you can use for free. And many apps don’t use one at all and rely on community indexes like Constellation (https://constellation.microcosm.blue/). Some don’t even run their own server or database.
This is an interesting take because AtProto feels both more accessible AND more decentralized to me (at least with my current mental model).
With ActivityPub, because running an instance requires hosting the data, the application, and dealing with all the subsequent scaling challenges, you kinda have to choose between being taking on active ops responsibilities or tying yourself to someone else's instance (which will probably be one of the bigger, more centralized ones).
If you decide you don't like an instance you picked and decide to move (unless things have changed) you're kinda stuck needing to start fresh.
With AtProto, it's trivial to jump ship to a different application platform and continue using your same identity. Exporting your data from a platform and self-hosting is a bit of a UX challenge, but at least it's possible.
As an example, I recently started using Tangled for the first time and was able to login using my existing bsky-backed domain (h14h.com). No need to create a new account or pick a new username -- it was as if I were already there. Then getting set up w/ self-hosting my git repos on a VPS was an afternoon of work at most, and it's just some backend service chugging away that I almost never have to think about.
The worst that will ever happen is I see a banner message in tangled.org saying something like "your repo is out of date and may be compatible with the latest version of Tangled", which I can solve by simply rebuilding & redeploying a docker image w/ the latest versions.
Granted, AtProto is definitely harder to wrap your head around architecturally. But actually interfacing it with a user is much simpler, IMO.
I'm not sure there is such a thing as "true" decentralization :) In my mind it's more of a buffet of tradeoffs rather than a single sliding scale.
FWIW, in the AP world there are several individuals and small teams running relays/mirrors/caches/AppViews and so on -- but you're right that this could get more expensive as things grow.
I appreciate how this explains the difference between the two.
But I also found it a little frustrating, because it answered one part of the question but failed to answer the question so what does ATProto do to solve the problems that instances solve?
For example, when this article dismisses defederation as merely a mysterious reason you might not see posts from your friends, it fails to answer "so how does atproto solve the problems that defederation solves?". Because the default reasonable answer to assume, given this framing, is "it doesn't".
> it fails to answer "so how does atproto solve the problems that defederation solves?".
The better way to ask this is, how does ActivityPub solve the problems that defederation causes? It's essentially the thing Microsoft does with email. Discard messages from all but the largest providers, defederate by default, forcing users to use Microsoft or another major incumbent if they want their messages to be delivered. Then new instances can't have their messages delivered, therefore can't get users. Which is obviously a perverse incentive for the major incumbents to not federate with new instances.
It's an architectural choice that has the long-term effect of cementing an oligopoly.
Meanwhile the claim is that it's necessary to prevent spam, but there are other providers that don't do this, e.g. in general you can deliver to Gmail as long as you have DKIM and reverse DNS etc. configured correctly, and those providers don't have any more of a spam problem than the ones who block innocent small servers by default.
Moreover, there is an obvious way to do this without giving the major instances a perverse incentive. You do the filtering on the client so that the filter list(s) you use are provided by something in the nature of uBlock rather than something in the nature of Microsoft, since the former doesn't operate any instances and therefore isn't trying to pressure everyone to use theirs.
If you’re asking about moderation, it works similarly as you’d expect it to work in a everything-RSS world.
At the hosting level, the hosting you use will likely ban you for clearly illegal stuff. Same as blogspot dot com or Cloudflare could ban you for certain things.
At the application level, application admins/mods would moderate as any app does. This is similar to running any web service today with user generated content. It’s up to app developers to choose. Apps can also provide primitives for userland moderation, like Reddit does, or even ability to plug your own extra moderation services (which Bluesky allows). But again, this is largely how it works on any app with user-generated content.
There’s no “defederation” because there’s no analog of “community instances” that may fight with each other. There’s hosting, there’s apps, and there’s app-level moderation that works according to each app’s developer’s choices.
> Apps can also provide primitives for userland moderation, like Reddit does, or even ability to plug your own extra moderation services (which Bluesky allows).
This is the part I would be looking for, in an article talking about "there are no instances". Is there a standard protocol for this, so that anyone can spin up a shared moderation service that people can subscribe to if they're aligned with it, and be able to plug that into any standard app built on the protocol (not just Bluesky-the-company's app)? Or is this something specific to Bluesky-the-company?
If this is a standardized part of the protocol, then that answers the question of "how does ATProto solve the same problems defederation solves".
There are several other things I can think of in the category of "how do you solve the problems that ActivityPub uses instances to solve", but they're things I've already asked in other parts of this thread, namely "how do you make the parts of the system not shown in the tidy hosts->apps M:N graph decentralized, too".
Of course, nothing stops an app from doing moderation differently and not using any of that. This is more for better composability and interoperability.
It doesn't solve those problems, except in an alternative universe where there are a very large number of appviews capable of consuming the entire firehose and you can freely choose between them and cheaply run your own. ATProto is like RSS in a universe where you can only read RSS through Google Reader (or a clone of Google Reader running on the same scale).
If it quacks like a duck... An account has a single Personal Data Server (PDS), right? The DID links to a PDS which is the canonical data feed for a user, and where a user's writes go. Data can be replicated but the PDS is treated as canonical. That's much closer to client/server architecture than distributed architecture. There's no P2P database. There's no writes into a DHT or peers. You write to your PDS, then those writes are optionally mirrored. Also discovery happens via DNS, so you don't even ask peers for data. You connect to a relay not as a peer but as a client asking a server for a copy of data that's canonically hosted by the PDS. I don't think it's a stretch to call the PDS an instance and the relay a mirror.
One thing I like about the ATProto is the stable identity, I could change the hosting as the author did recently and the other users see no difference.
The one down side of the system is the cost. It's cheap to host a PDS but expensive for other components. Users could not relies on "someone" for running those components for free forever.
The cost of ATProto is only high if your AppView has a goal of "every user should be able to see and access every post and interaction indefinitely". If one limits the scope of what users see on their AppView to, say, a similar scope as what popular ActivityPub applications do (only posts and interactions from users that use that instance, and people that they interact with / follow are available / retained), it becomes not that expensive at all.
The difference is that it's easy to scale ATProto AppViews up beyond the reach of their users. It can scale down though. It is not easy to scale ActivityPub instances up beyond the scope of the people who use it, and it would probably be way more expensive if you tried.
I wouldn't say other components are expensive. Common ones are:
- Relay as an optimization. That's cheap-ish ($30/mo) or free-ish if you pool with others.
- Your own app server. That's on par with normal web apps as long as you can keep a socket open. What's expensive is if the app you're making is a fully capable copy of Bluesky itself with gigabytes of existing posts — but is that the app you're making? The economics here are identical to normal web app stuff.
I wonder why were relays mentioned only mentioned in passing and there was no elaboration on how they interact with the rest of the network. Maybe because doing that would show that there are in fact "instances" in atproto, but who knows?
I mostly skipped over it because a Relay is an optimization and not essential to the shape of the network. It's not a fundamental element in the same way that PDS (hosting) and AppViews (app servers) are. It's more like a "next reasonable thing" an engineer would bolt on to make it easy to create apps.
A Relay is not an "instance" in any meaningful sense because it is a dumb retransmitter. It is cheap to run one, and it is easy to pool them between multiple apps. (Fun fact for nerds: the Relay's API for subscriptions is literally the same as a single server's. So a Relay is kind of a facade for "a bunch of servers" that lets you listen to their events combined.)
Early on (more than a year ago), running a Relay used to be more expensive because any Relay was expected to store the entire network archive. This is no longer a part of the contract, but a lot of discussions still reference or assume that. The current cost of running your own Relay (if you don't want to pool with anyone) is about $30/month. There are community-run Relays like https://firehose.network/ that you can use too.
It's not centralized in any way that matters. The Relay Bluesky uses is open source, you can run your own for $30/month if you really insist on doing that, it's trivial to pool between multiple apps if you want to lower costs, and there are already a few independent ones you can just use directly now, for example:
Google Reader feels like an ominous pick for an analogy. Sure, RSS survived the Google Reader shutdown, but not all the communities that used RSS (many that still don't know what RSS is) survived.
It feels almost "Freudian" to claim a thing is decentralized and then by analogy keep pointing to a massive (social) centralization of a decentralized ecosystem as a good thing. But especially one that we already know the ending for. Google Reader united a lot of RSS houses, value added a social graph and social commentary between them, and then at the whims of executives Google Reader fell and nearly killed RSS, but certainly destroyed an impressive social graph.
As an analogy that doesn't give me a lot of confidence in ATProto.
ATProto is an interesting protocol and there is lots of room to argue about its plus and minuses (and about Fedi/Masto's), but lots of people are already doing that so I won't.
My concern isn't technology or culture, it's money. At the moment, ATProto is existentially dependent on Bluesky PBC, a venture-funded startup ($100M from Bain Capital). There are people doing good work to make it more decentralized, more power to them, but at the moment it's still deeply centralized. And it's hard to see what the business model is that will support what Bsky PBC does at a global scale. Eventually Bain will want to see a revenue stream that justifies their investment; maybe there's a way to do that that doesn't involve enshittification but it's certainly not obvious.
You can dislike the instance-centeredness of Fedi/Masto (seems to have worked OK for email over the decades) but it's an actual thing that's actually working. And offers account migration without losing followers if you don't like the instance you're on. And has multiple really excellent client software packages. And seems to be covering its costs through a mixture of Patreon, co-operative & nonprofits, some Euro-gov help, all without any VC input. It can't be bought or owned by anybody.
Put another way, this is a really interesting space. But the technology is less interesting than the culture, and the culture is less interesting than the m money.
Yea that's fair. My stubborn point is that Fedi is trying to do the right thing using the wrong technology, and the resulting tradeoffs in user experience are why it will always remain niche. I hope that more resources will eventually flow from Fedi-centered solutions to more independent stuff in the atproto ecosystem.
I think email is a pretty poor example of decentralization in the modern internet. I pay a company to not have to worry about hosting my own and I've still had emails blackholed by the massive providers for unknown reasons I can only assume are some combination of the custom domain* and the provider. I went back to using Gmail on my resume and for applications after nearly missing out on an interview because of it.
I feel like the charts could be clearer if they showed the primary user experience difference between RSS and AT/AP, which is how do the arrows flow for Bob's response to Cat's post. I understand it fairly well for AP, I don't think I actually get it for AT.
TLDR: at the hosting layer, it's like <a href> links, but in JSON. Links can refer to other person's repo (even if it's hosted somewhere else). And then apps index everything, so they can present a coherent aggregated view (like a post thread).
Hah, well that's my bad for not reading articles I've saved, after reading those I remembered Puzzmo mentioned this recently and they, of course, got inspired by that second article: https://blog.puzzmo.com/posts/2026/03/02/atproto/
- There's no "instances" so I don't know what you mean by this.
- Re: PLC directory, indeed, there is only one of those. I think this is a legit point but it's worth considering the whole point of PLC directory is to be the single minimal stateless open source part that lifts identities out of apps and hostings. PLC governance and maintenance is being spun out into a Swiss organization (https://atproto.com/blog/plc-directory-org). Longer term the idea is for it to have a similar role to ICANN. Here's more on that: https://youtu.be/9z0z-Qu66yM?si=_8Dcw1M3VSKFGZhm&t=493
- Re: full Relays, they're easy/cheap to run, and you can run one yourself if you think the other ones are coordinating with Bluesky and don't trust their decisions. You don't need to depend on something else to do that.
> PLC governance and maintenance is being spun out into a Swiss organization (https://atproto.com/blog/plc-directory-org). Longer term the idea is for it to have a similar role to ICANN.
And since that sounds like a massive centralization problem, how do we have a dozen more of them with independent governance that aren't all controlled by either the same legal entity or by whoever has legal leverage to compel that entity?
Well, I think you also need to consider what PLC is. It’s an open source implementation of an open source spec. The implementation holds zero private state and exposes a verifiable log of operations for audit. There’s ongoing work on mirrors and replicas. Also, its output itself is cryptographically self-verifying.
I get that it’s not ideal but I think it’s worth keeping in mind that there’s not much you can mess up with it other than refusing to update requests. The threat model is very limited and it would immediately be obvious that this is happening, killing the credibility.
I’d also call out that activitypub has the same threat model in the form of ICANN, as it’s also heavily dependent on DNS for identity. I believe these are reasonable trades to make; realistically the alternative is to use a blockchain, which few people are keen to do.
Bluesky's moderation actions are generally implemented as moderation labels which take effect at the AppView level. Sometimes they'll take down someone's PDS or censor from their relays, but I don't believe third-party relays are aware of those relay takedowns at all.
Most decentralized systems I can think of tend to follow a power law distribution (roughly) where one or a few platforms dominate. In the fediverse that's mastodon.social (or maybe Threads), in email that's Gmail, in AP that's Bluesky. I'm not sure this is unique to the AT protocol?
I think the centralization on mastodon.social is a bit over estimated. According to fedidb.com there are over 1.1M active users in the fediverse. Of these mastodon.social has about 273k. That means mastodon.social has about 1/4 of the active users. The rest are scattered over nearly 43k servers. So I'd say the fediverse is pretty decentralized. That would make it pretty resilient.
On fedi there are several distinct pockets apart from the mastodon crowd which although don't dominate in raw headcount have their own flavour and unique view of the network (in 2nd place is "dark fedi", which has more shitposters, hackers and free speech adjacents; there's also a contingent of Japanese users mostly on Misskey instances). What I get out of AT is that this wouldn't really happen, since the content distribution would dominate your view with Bluesky posts regardless of your relationship with the rest of the network.
In AT you also get communities, in the same way that twitter had weird twitter, black twitter, and so on. But because you, by design, don't really see what PDS people use, it's more fluid.
I personally think that this actually leads to healthier communities but that may be a matter of taste!
Yeah, I mean a big reason why Fedi is segmented the way it is, is because the Mastodon crowd has a lot of merited and unmerited prejudices against the other parts of the network, and put a lot of effort into splitting it with centralized blocklists. The W3C working group that works with the AP spec (which refuses to cooperate with credible people representative of projects like Pleroma) is now looking to adopt a centralized moderation model similar to Bluesky's. It's one of the issues with the instance model, even though users have the tools to shape their own view of the network, to not have to see things they don't like, instance admins still get to dictate who they're allowed to talk to and what they see. I don't think that's right, even if the reasons are justifiable.
The guy who runs FSE (one of dark fedi's more notable instances) has written a lot of good blog posts on his side of the technical and social details, his one about running FediList is a good picture of the type of one-sided politics involved (and, if you're into the technical stuff, is rich with that too): https://blog.freespeechextremist.com/blog/about-fedilist.htm...
re: censorship & moderation, the folks @ BlackSky have been active in building out an alternative moderation structure (alongside running PDSes & AppViews).
All the nitpicking in this thread is missing the forest for the trees. You know what’s really bad for decentralization? The rest of the world living on fully centralized apps mostly belonging to one of 3 billionaires with no public protocol in sight.
ATProto making some idealistic compromises to improve the protocol as a product is a more effective half-court shot at the winning users from the oligarchy of apps than AP will ever be with its current design.
There’s a lot of talent in this community that could be spent building an ecosystem around the protocol far more likely to make a dent in social media centralization, but we’re stuck letting perfect be the enemy of the good.
Whoever operates hosting can decide to ban someone from hosting. This isn't different from how Cloudflare would ban you for hosting something illegal.
Whoever operates an app can decided to ban someone from that app. This isn't different from how a forum moderator can ban you for something they don't like.
Whoever operates services in the middle may decide to ban someone for network spam/abuse, same as cloud services may do if you abuse their limits.
You can always try to host your stuff elsewhere, and you can always access the same network from another app whose decisions you prefer.
So it's basically same as usual on the internet, but each role is separate, and you can mix and match what works for you.
Exactly this, people seem to confuse decentralization with federation. The Blockchain is decentralized, the fediverse is federated, big difference. I think it's because what we actually want and need right now (with western regimes dialing up censorship and repression to eleven) is decentralized social media, but that doesn't exist yet unfortunately.
Let's say I make a post on Bluesky, which is decentralized. My post is very contentious. It is blocked by the moderators, and the moderation service can't be disabled on bsky.app. I am now invisible on bsky.app.
So when this happens where do we go? Forget about "instance brain", your problem is Bluesky is vastly more centralized in practice than the theoretical marketing. Because if it was truly practically decentralized you could actually point to numerous instances of the service, but last time I raised this point there were... 3. Except one of them was actually not running the full appview and we weren't 100% sure the other one was either.
I'm sorry man, but this isn't going to cut it. A lot of people are absolutely right to not be sold on ATProto as it stands: there is no obvious reason to believe it will become more meaningfully decentralized over time rather than less. As it grows larger, the feasibility of having more "instances" that can run completely independently of Bluesky PBC becomes even less plausible.
If over 99% of the users are using Bluesky PBC infrastructure and placeholder DIDs, almost all of the keys to the kingdom lie in one place, and at that point you have invented Twitter with a ridiculous number of extra steps.
Can you explain to me why I would ever run my own PDS? Why would I pay to selfhost stuff while allowing someone to control almost everything I can see and do?
Unfortunately, this will never get answered. It's very easy to write a long blog post explaining how ATProto is technically decentralized. It's much harder to unpack how it actually isn't really.
I mean, this literally already happened (a person was banned, and Blacksky reversed that ban on their app server). So their account only works when seen thorough the Blacksky app. What is a better solution you’d like to see? I think it’s reasonable that there’s a market between these and if there’s enough demand, another app server can become popular. I don’t think it’s reasonable to expect that the apps shouldn’t be in control of their own moderation.
> I mean, this literally already happened (a person was banned, and Blacksky reversed that ban on their app server).
Blacksky is literally the only such example of alternative infrastructure that I know of, and obviously, it will not be applicable to the vast majority of people. Given the rising cost of hosting combined with the fact that the compute needs of running appviews and relays should theoretically only go up, I have a strong feeling that there will not be a lot more of them, either. It's already bigger than ActivityPub I believe and we're in the very low single digits at best.
Meanwhile, if we really did get a lot of these instances, then it really begs the question what the actual benefit of Bluesky's ATProto architecture is: if someone is banned on Bluesky and not Blacksky... won't users see a totally different view of the world? Isn't that the same problem ActivityPub sees? How does this really differ from defederation in practice?
> What is a better solution you’d like to see? I think it’s reasonable that there’s a market between these and if there’s enough demand, another app server can become popular.
If I knew how to fix this, I would probably be trying to help rather than criticizing ATProto. I don't think it can be fixed, so I don't have any suggestions.
> I don’t think it’s reasonable to expect that the apps shouldn’t be in control of their own moderation.
It kind of sounds like you're admitting that there is no real difference from a user standpoint with browsing to twitter.com vs bsky.app that have anything to do with decentralization.
I know I'm not going to win a popularity contest here, you don't even have to bother responding, honestly. But just being honest, I know you're a pretty intelligent person and the work you have done has benefited my life as a developer. I have a feeling deep down you also realize there is an inherent contradiction with Bluesky and ATProto's marketing pitch. I wish you would be honest about it.
The Fediverse has value specifically because of its downsides. A version of decentralized social media without those downsides inherently picks up almost all of the disadvantages of centralized social media. To me it seems apparent that all you can do is move the sliders around a bit, and Bluesky appears to net a very tiny percent of benefit from decentralization while bearing immense cost for it.
> It kind of sounds like you're admitting that there is no real difference from a user standpoint with browsing to twitter.com vs bsky.app that have anything to do with decentralization.
Bluesky users can interact with Blacksky users and vice versa unless Bluesky has applied moderation to the Blacksky user, because they are decentralized via ATproto. ~Twitter~ X users cannot interact with users on any other application, because X is not decentralized.
Yep. Effectively, there are a low single digit number of ATProto "instances" - they just pull posts from a more decentralized data layer than Fedi instances.
https://JFKSocial.com/ is a full Twitter (X.com) clone that works on BlueSky. It is new and launched.
Open source. BlueSky + Mastodon + NOSTR. Advanced feed algorithm. 100% of the code base is open source (in 2 weeks).
https://m.jfksocial.com/
It makes the feed rank algorithm metrics clearly visible on the right bar. People can confirm they aren't being censored or deboosted by seeing their own scores made transparent.
I'm pretty sure this comment is not written by AI, but those last few sentences really really made me think it is. It's a little sad that this type of antithesis now always has this AI-written connotation to it.
"That's not you cheating. It's you looking for emotional fulfillment. And frankly? That's courage."
I'd say atproto gives you a clear sense of what's tied to each of your identities — you can go and explore your repo in a browser. There's nothing to say your identity has to be "tied to a person" anymore than your Mastodon account on some server is "tied to a person". It's true atproto has a "scraping is the default, so expect it" vibe, whereas maybe you're arguing Mastodon allows security by obscurity?
To be fair, from 2022. I would argue that moderation in Fedi is holding together reasonably well. There are a few popular instances that lots of people think are overly aggressive/purist in their moderation policies, but the people using them seem to like what they're getting.
I haven't used Mastodon, but my experience with Lemmy was a lot of petty fighting about who should be defederated. The OP's reference to warring fiefdoms was very spot-on for that scene. I imagine it's a huge turn off to casual, well-adjusted people exploring the space. But I wouldn't know.
There's, of course, another sort of grossness to corporate moderation from above, but at least with ATProto you can take your identity and content to another AppView, if it wasn't shown there already. AFAIK, any fediverse migration tooling requires a cooperative host server you haven't already been banned from.
AT does have instances. They are just grouped differently.
In BlueSky, there is only one single "AppView" instance in the entire network. There is one instantiated "Firehose". Each user can instance his own "PDS".
In ActivityPub/Mastadon, the instances are "sender's server" vs "receiver's server."
The difference isn't that there aren't "instances" in AT proto. It's just that the instances are segmented differently.
Sure, there are servers, but the different grouping is the whole point because they're not coupling hosting to apps. When people say "where are Bluesky instances", they're asserting that it's useful to run many copies of the Bluesky database server. My article is an attempt to show that this way of thinking is very Mastodon-brained because these "instances" are the only unit of decentralization that's available in Mastodon. But you don't have to think this way.
In atproto, you can swap hosting (without changing apps), and you can create and use different apps (without changing hosting). That's the thing you can't do in Mastodon because it hard-couples hosting + apps into monolithic "instances".
In Mastodon, "receiver" and "sender" talk to each other, as you say. In atproto, hosting servers never talk to one another. The data from them flows into apps.
You're right that there's often a firehose in the middle, but that's also misleading. There doesn't have to be one firehose — there's a bunch of community-ran ones. It's relatively cheap to run one yourself these days (about $30/mo). It's easy to pool them between apps. And many apps don't use Firehose at all, and instead query community indexes like Constellation (https://constellation.microcosm.blue/). So "one firehose" is misleading.
Running your own firehose is not expensive, fwiw, it's $30/mo. If I were making a "serious" app I'd probably do that. Otherwise, relying on community-maintained ones seems fine.
Running an AppView for your own app is not expensive at all. It can be as cheap as you want. It's only expensive if you want to store gigabytes of Bluesky posts and serve them to millions of users — i.e. if you want to build the full Bluesky AppView. But why would you want to build a Bluesky AppView? That's part of what I'm alluding to in my article — atproto isn't "for Bluesky". You can build any social app.
I don't understand how running your own Relay is related to competing with Bluesky. A Relay is just a dumb websocket broadcaster. Yes, you can absolutely run one on your own if you don't want to rely on any of the existing ones. I don't think this has to do with competition.
I'm saying that if there is any required component of a full ATProto setup whose lowest-friction implementation is "use the One True Central Implementation, which every tool defaults to and which will be very painful to change", then it's not a decentralized protocol. Are there any components of ATProto that are found not through a service discovery mechanism that would seamlessly migrate to a new service, but by every individual app going "here's the hardcoded URL we never expect you to want to change"?
I like the concept of working like RSS. I don't like the idea of having a massive ecosystem coordination problem with game-theoretic network effects, for any component of the system.
When the first paragraph starts out by insulting people for not using the exact jargon you want them to, it really doesn’t make me want to read the rest.
This article is deeply, deeply misleading, in that it leaves out relays and appviews in all of its diagrams of the ATProto network. Like, are ATProto identities namespaced by instance/home-server like Mastodon or Matrix identities? No. Does who you are able to follow dependent on the appview you connect to? Yes. Are you able to run your own appview? Probably not (this answer has been upgraded from "No" in only the last few months).
> Every single time a post about atproto hits Hacker News, somebody asks in the comments: “But where are all the Bluesky instances?”. The problem is, there are no instances in atproto! The question is a category error. Instances are a Mastodon-brained concept, and I wanted something I can link to that explains this clearly.
I feel like you've (perhaps purposefully?) misinterpreted "instances" just to plug ATProto specifically at the expense of ActivityPub (and RSS, a bit). I think you lower yourself by doing this:
1. it forces you to omit and contort the interesting technical truths about ATProto and Activitypub, like Relays and their pros/cons for ATProto and account migrations and pros/cons for ActivityPub
2. it creates unnecessary conflict and criticism and seems unnecessarily divisive for 2 platforms solving problems in such a similar space
It's also just seems a bit silly: why would you assume that when someone asks "where are the instances?" they're not using the common mainstream use of the word "instances", like, servers, or running software, or VMs, or containers?
Sorry if this is overly harsh or I've misunderstood, but it gives me a strong vibe that it was motivated by disdain and frustration towards ActivityPub and ActivityPub users rather than wanting to legitimately inform the world about ActivityPub.
I did enjoy the diagrams and the explainers though! I just felt like the subtle digs and pops at activitypub were an unnecessary distraction.
> why would you assume that when someone asks "where are the instances?" they're not using the common mainstream use of the word "instances", like, servers, or running software, or VMs, or containers?
Of course depends on the context, but in a lot of discussions about ATProto, ActivityPub, Mastodon and nearby areas, people talk about "instances" as in "ActivityPub instances that host my data and my profile uses its URL as a 'name'". The blog post is specifically for that context I think.
It's less about trying to hide around the issue, and more reframing how you see the concepts, as people start to associate words with concepts and structures. So when people talk about "decentralized social media", lots of people think about ActivityPub, which typically (always?) has a kind of federated architecture, and the instance is one of those nodes in the network. When these people see ATProto, instinctively (and perhaps rightly so) they literally ask "But why is there only one Bluesky instance that people join?" as those concepts map close to what they know.
Overall I think the post is a good and useful addition to the discourse, with perhaps not a completely novel perspective, but posted publicly for future reference when this inevitably gets asks again sometime in the future, specifically for the people who have these previous associations already formed in their head.
I'm being a bit cheeky in the article's tone but I am fairly confident from discussions in the past that "But where are Bluesky instances?" is a common question which usually demonstrates a misunderstanding of the architecture where "having instances of an app" is seen as a measure of decentralization.
My article was an attempt to dig at this specific misunderstanding by comparing it to "But where are Google Reader instances?" which I think illustrates its absurdity. I genuinely do think that the two pictures I provide close to the end clear this up in a way a lot of early atproto/ActivityPub discussions completely gloss over.
Re: Relays, I wrote here on why I didn't include them: https://news.ycombinator.com/item?id=48600963. They're kind of incidental perf optimizations rather than essential to the model. In the post, I wanted to focus on the model.
From my perspective, I care about the centralization/decentralization aspect a lot, and if I'm coming into the discussion with a much better understanding of the Mastodon side then _of course_ I'm going to ask about the instances--that's the vocabulary I'm going to use to try to probe for flaws and gaps. It's not necessarily that it's the instances specifically I care about, or that I'm somehow technically misguided.
What I hoped to read in the article is how we approach topics like centralization, censorship, moderation, data ownership--and with a technical lens. But I feel like all I got was "here's why instances are the wrong vocabulary" without substantively talking about the part I personally care about and want to marry the technical understanding with. Maybe I just read too shallowly and need to sit with it.
I see! That's a huge topic by itself since you're raising a lot of questions. Maybe this could be a different article. My aim with this one was just to clarify the network topology because it is a prerequisite to having the other discussion, and too often that prerequisite is not there.
If you ask a list of specific questions, that would help a lot. I might be able to write something or reply inline here.
I appreciate that a lot! The article has a deliberate and explicit scope, and covers it well.
I'm hoping that perhaps my personal perspective shades why "instances" comes up, or why the reaction on HN seems to include the wider scope than the article itself covers.
Thanks for the fair response, I agree you're being cheeky. Sorry, I'm being lazy not searching here, but have you written anything on if instances of something is a good measure of decentralisation? (FWIW, I feel independently owned/managed instances in the traditional non-mastodon-definition seems like an okay measure of decentralisation.)
I completely agree with the point in your link that relays are different to instances - I love architectures involving dumb-relay or zero-trust type nodes. But I think Relays should still be mentioned in your post, since they're probably the main architectural element which protect PDS instances from the scale issues heavily federated AP instances might face, right? (I only have a high level understanding of ATProto and very little experience with AP, happy to be told I just need to learn more for this to make sense.)
It's a comparison they are directly inviting, by constantly claiming it's decentralized. And then its defenders get upset when people rightfully point out that there is only a single instance, because that single instance going down takes the whole thing down. Like in Google Reader.
If atproto app goes down and it’s open source, anyone can put it back up with all public data intact.
Even if it’s not open source, anyone who wants to write the code can still get it back up with all public data intact.
I think it’s a substantial difference with “takes the whole thing down”. Can we acknowledge that?
"but where are the instances?" is asking "if Bluesky the company disappeared or turned evil, would there be a Bluesky network that kept going?".
Perhaps ATProto vs. ActivityPub will been seen as the Fediverse's East-West Schism.
Instead of decrees over the "filioque" we get blog posts about the definition of "federation" where both parties talk past each other.
I found the distinction and comparison about Mastodon and ATProto are necessary. The fediverse model is easier to understand given existing social networks. ATProto is a novel concept that give users data sovereign and also the scalability of the centralized social networks.
I agree, a comparison and distinction is helpful, maybe even necessary. But I felt the author's bias came across a bit too strong in places and was a little distracting. Still interesting stuff though!
I think the analogy presented here is broken. RSS doesn't depend on Google Reader at all. Even at its prime, RSS depended less on Google Reader than email depends on Gmail now. In ATProto, AppViews heavily depends on Relays to be useful, and Relays are quite expensive to run. Also, the yellow circles which represent blogs in the RSS illustration are really not of the same nature as the same circles which represent posts on Facebook. Blogs are self-sufficient, for example.
I'm not saying ATProto is bad at all, but I feel like this blog post adds more confusion than it clarifies anything.
Relays are actually quite cheap now! They used to be a bit more expensive when they archived all the traffic, but in sync 1.1 that was dropped and they can be run on $20/mo VMs pretty trivially now
> Relays are quite expensive to run
While relays are among the more intensive parts of AT Protocol infrastructure, their cost of operation is still something most people can afford: approximately $30/mo now. What is truly expensive and difficult is something that will be immutably so regardless of how centralized or decentralized you are: moderation.
The author of this piece wrote about this common misconception about relays 9 months ago: https://news.ycombinator.com/item?id=45077291#45078223
As far as I can tell, Relays[1] are the glue that makes ATProto work performantly. I think they're supposed to be content-agnostic — they just shuttle data through, reducing the number of services each AppView needs to be aware of.
As the blog mentions, the big improvement vs Mastodon is that Relays, AppViews and PDSes are separate services with their own distinct scaling demands. It's a rather beautiful solution to a system design problem.
[1] https://atproto.com/guides/glossary
Yeah, Relays are one way to do that. I've mostly skipped them because they're an invisible optimization and there are other strategies. E.g. many smaller apps today rely on Constellation (https://constellation.microcosm.blue/) instead of building their own database index, so they don't use a Relay at all.
They do remove content directly from relays. They claim they only remove content that is illegal to host, but I don't know how true that is, and there is always the risk it could change in the future.
https://docs.bsky.app/blog/blueskys-moderation-architecture#...
I want the bsky org to be able to choose what content they host (and I think the internet would be a better place without section 230 protections allowing hosts to ignore the content they distribute); the promise as I understood it was that relays could be hot pluggable. If someone stopped carrying content (maybe it was illegal in /their/ region and not yours) you could failover to another relay.
However there is very little incentive to mirror any of the firehouse if someone else is doing it for free.
If that becomes a massive problem - host a relay with different moderation policy.
The scale-down floor for running an ATProto relay or appview is much, much higher than the floor for running a fediverse instance.
An important distinction is that blogs have their own websites and they're not required to publish full articles in their RSS feed.
Bluesky doesn't normally work that way - everything in the PDS gets replicated. They are also encouraging people to put put full blog posts in the PDS for easy replication. So, anyone who wants to index it gets a copy and you have no control over what they do.
You don't have to do it that way, though. You can publish your blog on your own website and just publish links to it on Bluesky.
Honestly that’s just as much because atproto is a raw data protocol. Putting an http frontend on an atproto account is something we encourage and a lot of folks do. I do that on pfrazee.com for instance, and my leaflet blogposts (which are canonically on atproto) render on my blog.
> So, anyone who wants to index it gets a copy and you have no control over what they do.
How does this differ from scrapers hitting the blog directly?
Web pages aren’t digitally signed, aren’t necessarily indexed by search engines, and there are ways to block bots with things like captchas. You also have much more control over the UI. If your blog has comments, you can moderate them, for better or worse.
With a PDS, the replication happens first, before anyone reads it, and the UI is out of your control.
Maybe that’s okay, but people should understand the tradeoffs.
How can people control the comments about their blog post on Hacker News? I think my example is closer to what happen in PDS and App View.
ATproto sacrifices true decentralization for consistency, Mastodon and AP does the opposite, sacrifices true consistency for more accessible decentralization.
At least that's how I understand it, because running an AP node is much more accessible to regular selfhosters than running one of those content relays in AT.
So all you'll ever "decentralize" in AT is your own data, it's more about owning your data rather than collectively owning a part of the network.
And we've been over this many times before on HN.
I think that’s a part of it but doesn’t state it fully.
AT doesn’t just give consistency, but a shared data model across apps. So apps can reference and render content from other apps. It’s really kind of like a web of typed JSON. Different apps are lenses through which you can see the same network. Anyone can build new experiences on top of old data. There’s nothing remotely equivalent in AP.
AP couples data to apps. In AT, it’s more like there’s one global database with entire world’s data that every app can query.
I don’t understand why the discussion always bumps into Relays. Running a Relay if you want to is cheap-ish ($30/mo) these days. There’s multiple existing ones (Bluesky or community) you can use for free. And many apps don’t use one at all and rely on community indexes like Constellation (https://constellation.microcosm.blue/). Some don’t even run their own server or database.
This is an interesting take because AtProto feels both more accessible AND more decentralized to me (at least with my current mental model).
With ActivityPub, because running an instance requires hosting the data, the application, and dealing with all the subsequent scaling challenges, you kinda have to choose between being taking on active ops responsibilities or tying yourself to someone else's instance (which will probably be one of the bigger, more centralized ones).
If you decide you don't like an instance you picked and decide to move (unless things have changed) you're kinda stuck needing to start fresh.
With AtProto, it's trivial to jump ship to a different application platform and continue using your same identity. Exporting your data from a platform and self-hosting is a bit of a UX challenge, but at least it's possible.
As an example, I recently started using Tangled for the first time and was able to login using my existing bsky-backed domain (h14h.com). No need to create a new account or pick a new username -- it was as if I were already there. Then getting set up w/ self-hosting my git repos on a VPS was an afternoon of work at most, and it's just some backend service chugging away that I almost never have to think about.
The worst that will ever happen is I see a banner message in tangled.org saying something like "your repo is out of date and may be compatible with the latest version of Tangled", which I can solve by simply rebuilding & redeploying a docker image w/ the latest versions.
Granted, AtProto is definitely harder to wrap your head around architecturally. But actually interfacing it with a user is much simpler, IMO.
I'm not sure there is such a thing as "true" decentralization :) In my mind it's more of a buffet of tradeoffs rather than a single sliding scale.
FWIW, in the AP world there are several individuals and small teams running relays/mirrors/caches/AppViews and so on -- but you're right that this could get more expensive as things grow.
I appreciate how this explains the difference between the two.
But I also found it a little frustrating, because it answered one part of the question but failed to answer the question so what does ATProto do to solve the problems that instances solve?
For example, when this article dismisses defederation as merely a mysterious reason you might not see posts from your friends, it fails to answer "so how does atproto solve the problems that defederation solves?". Because the default reasonable answer to assume, given this framing, is "it doesn't".
> it fails to answer "so how does atproto solve the problems that defederation solves?".
The better way to ask this is, how does ActivityPub solve the problems that defederation causes? It's essentially the thing Microsoft does with email. Discard messages from all but the largest providers, defederate by default, forcing users to use Microsoft or another major incumbent if they want their messages to be delivered. Then new instances can't have their messages delivered, therefore can't get users. Which is obviously a perverse incentive for the major incumbents to not federate with new instances.
It's an architectural choice that has the long-term effect of cementing an oligopoly.
Meanwhile the claim is that it's necessary to prevent spam, but there are other providers that don't do this, e.g. in general you can deliver to Gmail as long as you have DKIM and reverse DNS etc. configured correctly, and those providers don't have any more of a spam problem than the ones who block innocent small servers by default.
Moreover, there is an obvious way to do this without giving the major instances a perverse incentive. You do the filtering on the client so that the filter list(s) you use are provided by something in the nature of uBlock rather than something in the nature of Microsoft, since the former doesn't operate any instances and therefore isn't trying to pressure everyone to use theirs.
If you’re asking about moderation, it works similarly as you’d expect it to work in a everything-RSS world.
At the hosting level, the hosting you use will likely ban you for clearly illegal stuff. Same as blogspot dot com or Cloudflare could ban you for certain things.
At the application level, application admins/mods would moderate as any app does. This is similar to running any web service today with user generated content. It’s up to app developers to choose. Apps can also provide primitives for userland moderation, like Reddit does, or even ability to plug your own extra moderation services (which Bluesky allows). But again, this is largely how it works on any app with user-generated content.
There’s no “defederation” because there’s no analog of “community instances” that may fight with each other. There’s hosting, there’s apps, and there’s app-level moderation that works according to each app’s developer’s choices.
Does this help clarify it?
> Apps can also provide primitives for userland moderation, like Reddit does, or even ability to plug your own extra moderation services (which Bluesky allows).
This is the part I would be looking for, in an article talking about "there are no instances". Is there a standard protocol for this, so that anyone can spin up a shared moderation service that people can subscribe to if they're aligned with it, and be able to plug that into any standard app built on the protocol (not just Bluesky-the-company's app)? Or is this something specific to Bluesky-the-company?
If this is a standardized part of the protocol, then that answers the question of "how does ATProto solve the same problems defederation solves".
There are several other things I can think of in the category of "how do you solve the problems that ActivityPub uses instances to solve", but they're things I've already asked in other parts of this thread, namely "how do you make the parts of the system not shown in the tidy hosts->apps M:N graph decentralized, too".
Great questions! The protocol does have suggested app-agnostic moderation primitives (see https://atproto.com/guides/labels, https://atproto.com/guides/creating-a-labeler). That's the basis of how Bluesky's own moderation works under the hood. There's also Bluesky-specific tools and APIs like Ozone (https://atproto.com/guides/using-ozone), but they build on top of those primitives.
Of course, nothing stops an app from doing moderation differently and not using any of that. This is more for better composability and interoperability.
It doesn't solve those problems, except in an alternative universe where there are a very large number of appviews capable of consuming the entire firehose and you can freely choose between them and cheaply run your own. ATProto is like RSS in a universe where you can only read RSS through Google Reader (or a clone of Google Reader running on the same scale).
If it quacks like a duck... An account has a single Personal Data Server (PDS), right? The DID links to a PDS which is the canonical data feed for a user, and where a user's writes go. Data can be replicated but the PDS is treated as canonical. That's much closer to client/server architecture than distributed architecture. There's no P2P database. There's no writes into a DHT or peers. You write to your PDS, then those writes are optionally mirrored. Also discovery happens via DNS, so you don't even ask peers for data. You connect to a relay not as a peer but as a client asking a server for a copy of data that's canonically hosted by the PDS. I don't think it's a stretch to call the PDS an instance and the relay a mirror.
One thing I like about the ATProto is the stable identity, I could change the hosting as the author did recently and the other users see no difference.
The one down side of the system is the cost. It's cheap to host a PDS but expensive for other components. Users could not relies on "someone" for running those components for free forever.
The cost of ATProto is only high if your AppView has a goal of "every user should be able to see and access every post and interaction indefinitely". If one limits the scope of what users see on their AppView to, say, a similar scope as what popular ActivityPub applications do (only posts and interactions from users that use that instance, and people that they interact with / follow are available / retained), it becomes not that expensive at all.
The difference is that it's easy to scale ATProto AppViews up beyond the reach of their users. It can scale down though. It is not easy to scale ActivityPub instances up beyond the scope of the people who use it, and it would probably be way more expensive if you tried.
I wouldn't say other components are expensive. Common ones are:
- Relay as an optimization. That's cheap-ish ($30/mo) or free-ish if you pool with others.
- Your own app server. That's on par with normal web apps as long as you can keep a socket open. What's expensive is if the app you're making is a fully capable copy of Bluesky itself with gigabytes of existing posts — but is that the app you're making? The economics here are identical to normal web app stuff.
I wonder why were relays mentioned only mentioned in passing and there was no elaboration on how they interact with the rest of the network. Maybe because doing that would show that there are in fact "instances" in atproto, but who knows?
Author here!
I mostly skipped over it because a Relay is an optimization and not essential to the shape of the network. It's not a fundamental element in the same way that PDS (hosting) and AppViews (app servers) are. It's more like a "next reasonable thing" an engineer would bolt on to make it easy to create apps.
An app can work without a Relay (like https://reddwarf.app/ does). There are caches like Constellation (https://constellation.microcosm.blue/) that you can just query directly.
A Relay is not an "instance" in any meaningful sense because it is a dumb retransmitter. It is cheap to run one, and it is easy to pool them between multiple apps. (Fun fact for nerds: the Relay's API for subscriptions is literally the same as a single server's. So a Relay is kind of a facade for "a bunch of servers" that lets you listen to their events combined.)
Early on (more than a year ago), running a Relay used to be more expensive because any Relay was expected to store the entire network archive. This is no longer a part of the contract, but a lot of discussions still reference or assume that. The current cost of running your own Relay (if you don't want to pool with anyone) is about $30/month. There are community-run Relays like https://firehose.network/ that you can use too.
> Maybe because doing that would show that there are in fact "instances" in atproto, but who knows?
I wonder why you are vagueposting here instead of stating your position firmly. Maybe because you are afraid to be shown wrong, but who knows ?
I just searched up Relay and I guess Relay is like Feeder in the analogy? I don't need it? And if I want it I choose one (many?) or run my own.
You need a relay unless you want to have all the same NxM scaling problems as Mastodon.
Mostly because having a big centralized firehose relay is less decentralized than people want to admit.
That's not true.
It's not centralized in any way that matters. The Relay Bluesky uses is open source, you can run your own for $30/month if you really insist on doing that, it's trivial to pool between multiple apps if you want to lower costs, and there are already a few independent ones you can just use directly now, for example:
- https://pdsls.dev/firehose?instance=wss%3A%2F%2Fatproto.afri...
- https://pdsls.dev/firehose?instance=wss%3A%2F%2Feurope.fireh...
Why is that a problem if you can host relay for $50 per month as TFA mentions?
Google Reader feels like an ominous pick for an analogy. Sure, RSS survived the Google Reader shutdown, but not all the communities that used RSS (many that still don't know what RSS is) survived.
It feels almost "Freudian" to claim a thing is decentralized and then by analogy keep pointing to a massive (social) centralization of a decentralized ecosystem as a good thing. But especially one that we already know the ending for. Google Reader united a lot of RSS houses, value added a social graph and social commentary between them, and then at the whims of executives Google Reader fell and nearly killed RSS, but certainly destroyed an impressive social graph.
As an analogy that doesn't give me a lot of confidence in ATProto.
Ominous, but accurate. Imagine RSS but you couldn't use a desktop or mobile RSS reader, but only Google Reader or a comparably scaled clone.
In exchange we have many excellent RSS Reader. Just as recently someone talked about NetNewsWire.
ATProto is an interesting protocol and there is lots of room to argue about its plus and minuses (and about Fedi/Masto's), but lots of people are already doing that so I won't.
My concern isn't technology or culture, it's money. At the moment, ATProto is existentially dependent on Bluesky PBC, a venture-funded startup ($100M from Bain Capital). There are people doing good work to make it more decentralized, more power to them, but at the moment it's still deeply centralized. And it's hard to see what the business model is that will support what Bsky PBC does at a global scale. Eventually Bain will want to see a revenue stream that justifies their investment; maybe there's a way to do that that doesn't involve enshittification but it's certainly not obvious.
You can dislike the instance-centeredness of Fedi/Masto (seems to have worked OK for email over the decades) but it's an actual thing that's actually working. And offers account migration without losing followers if you don't like the instance you're on. And has multiple really excellent client software packages. And seems to be covering its costs through a mixture of Patreon, co-operative & nonprofits, some Euro-gov help, all without any VC input. It can't be bought or owned by anybody.
Put another way, this is a really interesting space. But the technology is less interesting than the culture, and the culture is less interesting than the m money.
Yea that's fair. My stubborn point is that Fedi is trying to do the right thing using the wrong technology, and the resulting tradeoffs in user experience are why it will always remain niche. I hope that more resources will eventually flow from Fedi-centered solutions to more independent stuff in the atproto ecosystem.
It doesn’t fully resolve your concerns, but I do want to highlight that atproto now has an IETF working group to home its spec development
Never rely a service just because they _hope_ to one day be less decentralised.
"Seems to have worked OK for email"
THANK YOU. It seems like far too few in this space really understand the benefits of actual decentralization.
ATProto feels like "centralized, except also we get other people to do the hard work with few of the benefits."
I think email is a pretty poor example of decentralization in the modern internet. I pay a company to not have to worry about hosting my own and I've still had emails blackholed by the massive providers for unknown reasons I can only assume are some combination of the custom domain* and the provider. I went back to using Gmail on my resume and for applications after nearly missing out on an interview because of it.
*: Not an .xyz, has proper SPF and DKIM records
I feel like the charts could be clearer if they showed the primary user experience difference between RSS and AT/AP, which is how do the arrows flow for Bob's response to Cat's post. I understand it fairly well for AP, I don't think I actually get it for AT.
I have a different article that goes into this here: https://overreacted.io/open-social/#open-social
Also here: https://overreacted.io/a-social-filesystem/#links
TLDR: at the hosting layer, it's like <a href> links, but in JSON. Links can refer to other person's repo (even if it's hosted somewhere else). And then apps index everything, so they can present a coherent aggregated view (like a post thread).
Hah, well that's my bad for not reading articles I've saved, after reading those I remembered Puzzmo mentioned this recently and they, of course, got inspired by that second article: https://blog.puzzmo.com/posts/2026/03/02/atproto/
There's basically only one instance.
There's only one PLC directory.
There's very few full relays (edit: appviews), none that I'm aware of that don't mirror bluesky censorship/moderation decisions.
- There's no "instances" so I don't know what you mean by this.
- Re: PLC directory, indeed, there is only one of those. I think this is a legit point but it's worth considering the whole point of PLC directory is to be the single minimal stateless open source part that lifts identities out of apps and hostings. PLC governance and maintenance is being spun out into a Swiss organization (https://atproto.com/blog/plc-directory-org). Longer term the idea is for it to have a similar role to ICANN. Here's more on that: https://youtu.be/9z0z-Qu66yM?si=_8Dcw1M3VSKFGZhm&t=493
- Re: full Relays, they're easy/cheap to run, and you can run one yourself if you think the other ones are coordinating with Bluesky and don't trust their decisions. You don't need to depend on something else to do that.
> PLC governance and maintenance is being spun out into a Swiss organization (https://atproto.com/blog/plc-directory-org). Longer term the idea is for it to have a similar role to ICANN.
And since that sounds like a massive centralization problem, how do we have a dozen more of them with independent governance that aren't all controlled by either the same legal entity or by whoever has legal leverage to compel that entity?
Well, I think you also need to consider what PLC is. It’s an open source implementation of an open source spec. The implementation holds zero private state and exposes a verifiable log of operations for audit. There’s ongoing work on mirrors and replicas. Also, its output itself is cryptographically self-verifying.
I get that it’s not ideal but I think it’s worth keeping in mind that there’s not much you can mess up with it other than refusing to update requests. The threat model is very limited and it would immediately be obvious that this is happening, killing the credibility.
I’d also call out that activitypub has the same threat model in the form of ICANN, as it’s also heavily dependent on DNS for identity. I believe these are reasonable trades to make; realistically the alternative is to use a blockchain, which few people are keen to do.
There are many relays, here's a list of 13 but it's not comprehensive: https://tangled.org/firehose.club/community-firehose-list/bl...
Bluesky's moderation actions are generally implemented as moderation labels which take effect at the AppView level. Sometimes they'll take down someone's PDS or censor from their relays, but I don't believe third-party relays are aware of those relay takedowns at all.
You're right, I was probably confusing appviews with relays.
Most decentralized systems I can think of tend to follow a power law distribution (roughly) where one or a few platforms dominate. In the fediverse that's mastodon.social (or maybe Threads), in email that's Gmail, in AP that's Bluesky. I'm not sure this is unique to the AT protocol?
I think the centralization on mastodon.social is a bit over estimated. According to fedidb.com there are over 1.1M active users in the fediverse. Of these mastodon.social has about 273k. That means mastodon.social has about 1/4 of the active users. The rest are scattered over nearly 43k servers. So I'd say the fediverse is pretty decentralized. That would make it pretty resilient.
On fedi there are several distinct pockets apart from the mastodon crowd which although don't dominate in raw headcount have their own flavour and unique view of the network (in 2nd place is "dark fedi", which has more shitposters, hackers and free speech adjacents; there's also a contingent of Japanese users mostly on Misskey instances). What I get out of AT is that this wouldn't really happen, since the content distribution would dominate your view with Bluesky posts regardless of your relationship with the rest of the network.
In AT you also get communities, in the same way that twitter had weird twitter, black twitter, and so on. But because you, by design, don't really see what PDS people use, it's more fluid.
I personally think that this actually leads to healthier communities but that may be a matter of taste!
Yeah, I mean a big reason why Fedi is segmented the way it is, is because the Mastodon crowd has a lot of merited and unmerited prejudices against the other parts of the network, and put a lot of effort into splitting it with centralized blocklists. The W3C working group that works with the AP spec (which refuses to cooperate with credible people representative of projects like Pleroma) is now looking to adopt a centralized moderation model similar to Bluesky's. It's one of the issues with the instance model, even though users have the tools to shape their own view of the network, to not have to see things they don't like, instance admins still get to dictate who they're allowed to talk to and what they see. I don't think that's right, even if the reasons are justifiable.
The guy who runs FSE (one of dark fedi's more notable instances) has written a lot of good blog posts on his side of the technical and social details, his one about running FediList is a good picture of the type of one-sided politics involved (and, if you're into the technical stuff, is rich with that too): https://blog.freespeechextremist.com/blog/about-fedilist.htm...
re: censorship & moderation, the folks @ BlackSky have been active in building out an alternative moderation structure (alongside running PDSes & AppViews).
All the nitpicking in this thread is missing the forest for the trees. You know what’s really bad for decentralization? The rest of the world living on fully centralized apps mostly belonging to one of 3 billionaires with no public protocol in sight.
ATProto making some idealistic compromises to improve the protocol as a product is a more effective half-court shot at the winning users from the oligarchy of apps than AP will ever be with its current design.
There’s a lot of talent in this community that could be spent building an ecosystem around the protocol far more likely to make a dent in social media centralization, but we’re stuck letting perfect be the enemy of the good.
Nobody asks _how_ are all the Bluesky instances :(
Thanks! Your name makes me think of a funner pre-LLM time when Elm and Redux was new and cool. Great explainer!
So all the censor needs to do is cut off one host? And then you upload everything to different host and connect that?
What kind of censor?
Whoever operates hosting can decide to ban someone from hosting. This isn't different from how Cloudflare would ban you for hosting something illegal.
Whoever operates an app can decided to ban someone from that app. This isn't different from how a forum moderator can ban you for something they don't like.
Whoever operates services in the middle may decide to ban someone for network spam/abuse, same as cloud services may do if you abuse their limits.
You can always try to host your stuff elsewhere, and you can always access the same network from another app whose decisions you prefer.
So it's basically same as usual on the internet, but each role is separate, and you can mix and match what works for you.
Exactly this, people seem to confuse decentralization with federation. The Blockchain is decentralized, the fediverse is federated, big difference. I think it's because what we actually want and need right now (with western regimes dialing up censorship and repression to eleven) is decentralized social media, but that doesn't exist yet unfortunately.
> So all the censor needs to do is cut off one host? And then you upload everything to different host and connect that?
Which host?
If you're curious about atproto and feel like a n00b about it, this is a great place to start.
Let's say I make a post on Bluesky, which is decentralized. My post is very contentious. It is blocked by the moderators, and the moderation service can't be disabled on bsky.app. I am now invisible on bsky.app.
So when this happens where do we go? Forget about "instance brain", your problem is Bluesky is vastly more centralized in practice than the theoretical marketing. Because if it was truly practically decentralized you could actually point to numerous instances of the service, but last time I raised this point there were... 3. Except one of them was actually not running the full appview and we weren't 100% sure the other one was either.
I'm sorry man, but this isn't going to cut it. A lot of people are absolutely right to not be sold on ATProto as it stands: there is no obvious reason to believe it will become more meaningfully decentralized over time rather than less. As it grows larger, the feasibility of having more "instances" that can run completely independently of Bluesky PBC becomes even less plausible.
If over 99% of the users are using Bluesky PBC infrastructure and placeholder DIDs, almost all of the keys to the kingdom lie in one place, and at that point you have invented Twitter with a ridiculous number of extra steps.
Can you explain to me why I would ever run my own PDS? Why would I pay to selfhost stuff while allowing someone to control almost everything I can see and do?
Unfortunately, this will never get answered. It's very easy to write a long blog post explaining how ATProto is technically decentralized. It's much harder to unpack how it actually isn't really.
I mean, this literally already happened (a person was banned, and Blacksky reversed that ban on their app server). So their account only works when seen thorough the Blacksky app. What is a better solution you’d like to see? I think it’s reasonable that there’s a market between these and if there’s enough demand, another app server can become popular. I don’t think it’s reasonable to expect that the apps shouldn’t be in control of their own moderation.
> I mean, this literally already happened (a person was banned, and Blacksky reversed that ban on their app server).
Blacksky is literally the only such example of alternative infrastructure that I know of, and obviously, it will not be applicable to the vast majority of people. Given the rising cost of hosting combined with the fact that the compute needs of running appviews and relays should theoretically only go up, I have a strong feeling that there will not be a lot more of them, either. It's already bigger than ActivityPub I believe and we're in the very low single digits at best.
Meanwhile, if we really did get a lot of these instances, then it really begs the question what the actual benefit of Bluesky's ATProto architecture is: if someone is banned on Bluesky and not Blacksky... won't users see a totally different view of the world? Isn't that the same problem ActivityPub sees? How does this really differ from defederation in practice?
> What is a better solution you’d like to see? I think it’s reasonable that there’s a market between these and if there’s enough demand, another app server can become popular.
If I knew how to fix this, I would probably be trying to help rather than criticizing ATProto. I don't think it can be fixed, so I don't have any suggestions.
> I don’t think it’s reasonable to expect that the apps shouldn’t be in control of their own moderation.
It kind of sounds like you're admitting that there is no real difference from a user standpoint with browsing to twitter.com vs bsky.app that have anything to do with decentralization.
I know I'm not going to win a popularity contest here, you don't even have to bother responding, honestly. But just being honest, I know you're a pretty intelligent person and the work you have done has benefited my life as a developer. I have a feeling deep down you also realize there is an inherent contradiction with Bluesky and ATProto's marketing pitch. I wish you would be honest about it.
The Fediverse has value specifically because of its downsides. A version of decentralized social media without those downsides inherently picks up almost all of the disadvantages of centralized social media. To me it seems apparent that all you can do is move the sliders around a bit, and Bluesky appears to net a very tiny percent of benefit from decentralization while bearing immense cost for it.
> It kind of sounds like you're admitting that there is no real difference from a user standpoint with browsing to twitter.com vs bsky.app that have anything to do with decentralization.
Bluesky users can interact with Blacksky users and vice versa unless Bluesky has applied moderation to the Blacksky user, because they are decentralized via ATproto. ~Twitter~ X users cannot interact with users on any other application, because X is not decentralized.
Yep. Effectively, there are a low single digit number of ATProto "instances" - they just pull posts from a more decentralized data layer than Fedi instances.
https://JFKSocial.com/ is a full Twitter (X.com) clone that works on BlueSky. It is new and launched.
Open source. BlueSky + Mastodon + NOSTR. Advanced feed algorithm. 100% of the code base is open source (in 2 weeks). https://m.jfksocial.com/
It makes the feed rank algorithm metrics clearly visible on the right bar. People can confirm they aren't being censored or deboosted by seeing their own scores made transparent.
This sounds like BlueSky without the user-directed self and grouped moderation tools.
That’s not censorship. It’s showing you the door if you’re a jerk.
And it sounds awful.
I'm pretty sure this comment is not written by AI, but those last few sentences really really made me think it is. It's a little sad that this type of antithesis now always has this AI-written connotation to it.
"That's not you cheating. It's you looking for emotional fulfillment. And frankly? That's courage."
Yes, and this is exactly why ATProto is worse and more dangerous. Instances are safer. precisely because they are more genuinely decentralized.
The ability to forever tie your stuff to a person, strongly, is exactly what the surveillance state would want.
Mastodon's model gives you plausible deniability. It's safer.
I'd say atproto gives you a clear sense of what's tied to each of your identities — you can go and explore your repo in a browser. There's nothing to say your identity has to be "tied to a person" anymore than your Mastodon account on some server is "tied to a person". It's true atproto has a "scraping is the default, so expect it" vibe, whereas maybe you're arguing Mastodon allows security by obscurity?
Semi related post on why the moderation of federated Mastodon instances is a problem: https://blog.raed.dev/posts/mastodon_moderation
To be fair, from 2022. I would argue that moderation in Fedi is holding together reasonably well. There are a few popular instances that lots of people think are overly aggressive/purist in their moderation policies, but the people using them seem to like what they're getting.
I haven't used Mastodon, but my experience with Lemmy was a lot of petty fighting about who should be defederated. The OP's reference to warring fiefdoms was very spot-on for that scene. I imagine it's a huge turn off to casual, well-adjusted people exploring the space. But I wouldn't know.
There's, of course, another sort of grossness to corporate moderation from above, but at least with ATProto you can take your identity and content to another AppView, if it wasn't shown there already. AFAIK, any fediverse migration tooling requires a cooperative host server you haven't already been banned from.
AT does have instances. They are just grouped differently.
In BlueSky, there is only one single "AppView" instance in the entire network. There is one instantiated "Firehose". Each user can instance his own "PDS".
In ActivityPub/Mastadon, the instances are "sender's server" vs "receiver's server."
The difference isn't that there aren't "instances" in AT proto. It's just that the instances are segmented differently.
Sure, there are servers, but the different grouping is the whole point because they're not coupling hosting to apps. When people say "where are Bluesky instances", they're asserting that it's useful to run many copies of the Bluesky database server. My article is an attempt to show that this way of thinking is very Mastodon-brained because these "instances" are the only unit of decentralization that's available in Mastodon. But you don't have to think this way.
In atproto, you can swap hosting (without changing apps), and you can create and use different apps (without changing hosting). That's the thing you can't do in Mastodon because it hard-couples hosting + apps into monolithic "instances".
In Mastodon, "receiver" and "sender" talk to each other, as you say. In atproto, hosting servers never talk to one another. The data from them flows into apps.
You're right that there's often a firehose in the middle, but that's also misleading. There doesn't have to be one firehose — there's a bunch of community-ran ones. It's relatively cheap to run one yourself these days (about $30/mo). It's easy to pool them between apps. And many apps don't use Firehose at all, and instead query community indexes like Constellation (https://constellation.microcosm.blue/). So "one firehose" is misleading.
There are multiple Bluesky AppViews, Blacksky has a totally independent one. And there are multiple relays, each capable of serving a firehose.
You can have your own AppView and Firehose. They're just relatively expensive to run versus a PDS.
Running your own firehose is not expensive, fwiw, it's $30/mo. If I were making a "serious" app I'd probably do that. Otherwise, relying on community-maintained ones seems fine.
Running an AppView for your own app is not expensive at all. It can be as cheap as you want. It's only expensive if you want to store gigabytes of Bluesky posts and serve them to millions of users — i.e. if you want to build the full Bluesky AppView. But why would you want to build a Bluesky AppView? That's part of what I'm alluding to in my article — atproto isn't "for Bluesky". You can build any social app.
> But why would you want to build
The problem is that that sounds like "you shouldn't want to compete with Bluesky". Which makes it dangerously centralized.
I don't understand how running your own Relay is related to competing with Bluesky. A Relay is just a dumb websocket broadcaster. Yes, you can absolutely run one on your own if you don't want to rely on any of the existing ones. I don't think this has to do with competition.
I'm saying that if there is any required component of a full ATProto setup whose lowest-friction implementation is "use the One True Central Implementation, which every tool defaults to and which will be very painful to change", then it's not a decentralized protocol. Are there any components of ATProto that are found not through a service discovery mechanism that would seamlessly migrate to a new service, but by every individual app going "here's the hardcoded URL we never expect you to want to change"?
I like the concept of working like RSS. I don't like the idea of having a massive ecosystem coordination problem with game-theoretic network effects, for any component of the system.
When the first paragraph starts out by insulting people for not using the exact jargon you want them to, it really doesn’t make me want to read the rest.
This article is deeply, deeply misleading, in that it leaves out relays and appviews in all of its diagrams of the ATProto network. Like, are ATProto identities namespaced by instance/home-server like Mastodon or Matrix identities? No. Does who you are able to follow dependent on the appview you connect to? Yes. Are you able to run your own appview? Probably not (this answer has been upgraded from "No" in only the last few months).