Every one of these "How AT proto works" explainers focuses on data ownership—which is where ATProto shines—and glosses over data processing, where ATProto is decidedly weaker than ActivityPub.
ATProto is built on a global, public view of the world, where all events are visible to a trusted global "AppServer" that can make all of the decisions for you—how to create your feed, who can see who's posts, etc—all of those decisions have to be made by a trusted intermediary. ActivityPub is more like RSS or email—your local server only has to manage the feeds you subscribe to, and your inbox is directly built from all of the posts you have access to. People you subscribe to send you your posts, and you don't have to process them at all.
This is why Bluesky could never have "private likes" in the same way Twitter or ActivityPub does—every AppView needs to track the like counts of every post in the network manually. It's a huge hassle! I just don't see this architecture winning out in the long term, when compared to the AP feed-subscription architecture.
primarily because multiple programs can access the same identity
Actually, this was how AP was originally designed as well—it was just that the most popular early implementations took shortcuts to remove that functionality to fit them into their existing architecture. This is a direct consequence of the fact that the biggest AP implementations when it was initially adopted were descendants of older OStatus social networks, and not built to be "ActivityPub-native" from the ground up.
> This is why Bluesky could never have "private likes" in the same way Twitter or ActivityPub does
I didn't know "private likes" even existed, but if atproto includes public key encryption, I could publish a record containing a "like" that I have encrypted with the "liked" user's public key. Only that user would know what the record contained. Though, the fact that the encrypted message exists and had a cleartext "@name" is itself informative to adversaries. Concealing that level of info would require other measures.
Correct. All information in the system is part of a public, append-only ledger. That's the thing I'm pointing out is a fundamental weakness of the system.
I think the lack of owner's custody of the private key in bluesky is more the result of a convenience design than of a technical limitation. If the user is expected to manage the private key, you might end up with something similar in complexity to Nostr. But it might still be possible if you're running your own PDS.
I haven't worked with Fedify before, but what I'm talking about is the difference between a service that has a "translation layer" between their own internal data model and ActivityPub, and a server that actually stores full ActivityPub object's in the user's inbox and outbox, and displays them unchanged to clients. 90% of deployed ActivityPub applications—like yours!—are the former, not the latter.
A true "ActivityPub server" is almost as simple as a Bluesky PDS—all it's responsible for is 1) storing blobs of data, 2) fanning out subscriptions and 3) collecting incoming data for you to view. In the original way ActivityPub was designed, all of the actual data presentation layers—Mastodon, PeerTube, Pixelfed—were designed to be specialized clients that could communicate with the user's generic server. However, the first popular implementations cut out the client-to-server part of the protocol, so now we're stuck in a place where everyone thinks ActivityPub means you need a separate identity for every client application.
So, what I would recommend for your own username/password site is implementing an ActivityPub client, and recommending that users use it to connect to a third-party ActivityPub server. That way, the user owns the data, and they simply use your service to get access to (filtered version of) it.
Unfortunately, since this is a less implemented part of the protocol, the client APIs necessary to make this a reality haven't seen much development. And you're facing an uphill battle for user adoption. In practice, users don't seem to mind having separate accounts and identities for different clients. It reminds me a bit of the "key management problem" in e2e cryptography. Having a stable cryptographic identity is doable if you're technically minded, but most people just muddle along and don't really care about it that much—they create new keys every time they get a new laptop instead of trying to figure out cross-signing, and everything works itself out more or less fine.
I mess around with fedify a lot, it's great. It's fun to integrate with existing websites.
I've thought a lot about ATProto and integrating it in similar ways. I'd love to have a look at what you're doing and how. The struggle I have is that I think the ATProto repos have a fairly strong cryptographic structure compared to AP
If someone requests an object over AP, that object contents can come from anywhere easily, and can be signed easy. So for me, when someone requests an activitypub object of one of my notes via fedify, it just reads the truth from my markdown note files and returns it. If I edit my markdown files, it's no real issue, the next request gets the latest version of that markdown (there's some signing nuances in places, but it's generally straightforward)
With ATProto PDS and repositories use things like Merkle Search Trees and other things which I assume means the backend data needs to be a lot more... consistent. Like the data has to live in the PDS, and that has to become the source of truth to maintain the merkle structures, including updates.
But with AP via fedify, it feels super easy and nice for my source of truth to be whatever backend store I like (markdown notes).
I've done enough with crypto to see the benefits provided by the transparent verifiable history of merkle like structures, but honestly, this is social media not cash: I don't care if someone wants to subtly change something to manage how they come across with their own social media. In that respect, I feel the ATProto repositories overcomplicate things a bit.
This is why Bluesky could never have "private likes" in the same way Twitter or ActivityPub does—every AppView needs to track the like counts of every post in the network manually. It's a huge hassle! I just don't see this architecture winning out in the long term, when compared to the AP feed-subscription architecture.
Actually, this was how AP was originally designed as well—it was just that the most popular early implementations took shortcuts to remove that functionality to fit them into their existing architecture. This is a direct consequence of the fact that the biggest AP implementations when it was initially adopted were descendants of older OStatus social networks, and not built to be "ActivityPub-native" from the ground up.