What I don't like about Mastodon is the way it federate servers. For example, if someone on server B "retweets" a post from server A, it does not carry over likes and comments from server A. That means personal instances are cumbersome and big servers always have a bigger leverage. That also implies you have no single page to just see all the comments about that post.
Bsky, on the other hand, does not have such hard splitting between servers (comments belong to the post on ATProto network, not to the instance) and allows you to run your own personal instances (PDS) natively. You can even transfer your profile and data (including followers) to other instances. IMO, it's a better tackle on building a federated social network.
But yeah, the native recommendation algorithm is lackluster.
True. Federation is a disaster. Like you say, the simple use case of having consistent likes and replies doesn't even work.
This breaks conversations. An OP hosting a popular thread will see the same/similar reply many times because people can't even see each other's replies.
Federation makes discovery and search impossible or at the very least extremely hard to implement, whilst this is such a core feature of social networks.
Federation is extremely wasteful. People hosting instances with just a few thousands of users are reporting gigantic bandwidth consumption and hefty bills. Not to mention storage bills for all that media.
Federation does not allow for the independence it suggests. A bitter activist can advocate for other admins to defederate you and then your instance is cut off. You're not independent, you answer to the rule of the purists and the Mastodon way.
Bsky, on the other hand, does not have such hard splitting between servers (comments belong to the post on ATProto network, not to the instance) and allows you to run your own personal instances (PDS) natively. You can even transfer your profile and data (including followers) to other instances. IMO, it's a better tackle on building a federated social network.
But yeah, the native recommendation algorithm is lackluster.