The AT protocol is the most obtuse crock of shit

Mastodon is definitely bad. I had to do an emergency migration to a new instance and lost all of my posts.

Okay well. I work on Bluesky and helped build the AT Protocol. I'm sorry Sam differs with us on this, and I'm glad that Activity Pub is already there for him. However, Sam doesn't understand the ATProto very well and I want to clear it up a bit.

Before I do, let me just say: Bluesky and the AT Proto are in beta. The stuff that seems incomplete or poorly documented is incomplete and poorly documented. Everything has moved enormously faster than we expected it to. We have around 65k users on the beta server right now. We _thought_ that this would be a quiet, stealthy beta for us while we finished the technology and the client. We've instead gotten a ton of attention, and while that's wonderful it means that we're getting kind of bowled over. So I apologize for the things that aren't there yet. I haven't really rested in over a month.

ATProto doesn't use crypto in the coin sense. It uses cryptography. The underlying premise is actually pretty similar to git. Every user runs a data repository where commits to the repository are signed. The data repositories are synced between nodes to exchange data, and interactions are committed as records to the repositories.

The purpose of the data repository is to create a clear assertion of the user's records that can be gossiped and cached across the network. We sign the records so that authenticity can be determined without polling the home server, and we use a repository structure rather than signing individual records so that we can establish whether a record has been deleted (signature revocation).

Repositories are pulled through replication streams. We chose not to push events to home servers because you can easily overwhelm a home server with a lot of burst loads when some content goes viral, which in turn makes self hosting too expensive. If a home server wants to crawl & pull records or repositories it can, and there's a very sensible model for doing so based on its users' social graph. However the general goal is to create a global network that aggregates activity (such as likes) across the entire network, and so we use large scale aggregation services to provide that aggregated firehose. Unless somebody solves federated queries with the sufficient performance then any network that's trying to give a global view is going to need similar large indexes. If you don't want a global view that's fine, then you want a different product experience and you can do that with ATProto. You can also use a different global indexer than the one we provide, same as search engines.

The schema is a well-defined machine language which translates to static types and runtime validation through code generation. It helps us maintain correctness when coordinating across multiple servers that span orgs, and any protocol that doesn't have one is informally speccing its logic across multiple codebases and non-machine-readable specs. The schema helps the system with extensibility and correctness, and if there was something off the shelf that met all our needs we would've used it.

The DID system uses the recovery key to move from one server to another without coordinating with the server (ie because it suddenly disappeared). It supports key rotations and it enables very low friction moves between servers without any loss of past activity or data. That design is why we felt comfortable just defaulting to our hosting service; because we made it easy to switch off after the fact if/when you learn there's a better option. Given that the number one gripe about activitypub's onboarding is server selection, I think we made the right call.

We'll keep writing about what we're doing and I hope we change some minds over time. The team has put a lot of thought into the work, and we really don't want to fight with other projects that have a similar mission.

It seems to me that a lot of the problems discussed in this thread (it's using a new protocol that doesn't work with existing tools, it uses crypto, handles auth in the protocol, server can goes down) are just frustrations that don't have to do with the core innovation that Bluesky promises to deliver, and are instead confusing the AT protocol to be another ActivityPub-related protocol, rather than something completely different. In fact, it misses that having pull-based indexes is part of the idea of separating data publication vs. data curation. Yes, invite codes are kind of a weird thing to shove so deep into the protocol but sometimes these kinds of things are necessary to bootstrap a social network, and become relatively unimportant in the long term.

But on the other hand, the AT protocol is not all what it seems either. It's fundamentally not possible to do what they promise unless there is some kind of synchronization point where your most up-to-date identity can be searched. The documentation goes on extensively about the format of DIDs and the fact that there _is_ a resolution scheme, but it fails to mention that this resolution is being done by Bluesky themselves, and is not planned to expand into something that others can control. As a decentralized protocol, you would expect this to be something DNS-based, or if you really wanted something more peer-to-peer, DHT lookup, or at the worst case, a distributed blockchain. Seeing as this is a fundamental protocol-level change, the fact that they're rolling with the current approach makes me believe they will not be changing this in the near future. Bluesky is just converting one problem into another.

Having done a decent bit of hacking around ActivityPub, when I read the (thin, which 'pfraze and others have copped to) documentation I immediately went "oh, this is going to be way more scalable than ActivityPub once it's done."

It's not all roses. I'm not sold on lexicons and xrpc, but that's probably because I am up to my eyeballs in JSON Schema and OpenAPI on a daily basis and my experience and my existing toolkit probably biases me. I think starting with generally-accepted tooling probably would've been a better idea--but, in a vacuum, they're reasonably thought-out, they do address real problems, and I can't shake the feeling that the fine article is spitting mad for the sake of being spitting mad.

While federation isn't there yet, granted, the idea that you can't write code against this is hogwash. There's a crazily thriving ecosystem already from the word jump, ~1K folks in the development Discord and a bunch of tooling being added on top of the platform by independent developers right now.

Calm down. Nobody's taking ActivityPub away from people who like elephants.

The DID method has to satisfy a lot of requirements. I did a ton of research on DHT and blockchain approaches and none of them give the right performance, reliability, and cost outcomes while also supporting key rotation. DNS isn't far off but it's a little obtuse to cram it into, so we're starting with did:web and did:plc. We'll add others if they have the right characteristics.

The only options that seem reasonable for PLC is for it either to be operated by some trusted multi-stakeholder institution like ICANN or to be a closed-group blockchain. I'm open to either approach but they require the creation of a consortium, which requires buyin from a set of stakeholders that don't exist yet. So until that happens, we run it, and that's why we called it Placeholder. We'll backronym it into something else once we get there (the C most likely being Consortium).