I miss (at least) three things when I go to a new social network: 1\ ability to follow people that I followed in other social nets, 2\ algorithmic feed that knows about me 3\ functional "semantic" search. All three could be developed independently of actual social net used, and that's actually where the most value is.
Yeah, 100% agree! I've been reading the ActivityPub specs and it seems to me the Actor objects https://www.w3.org/TR/activitypub/#actor-objects actually hold the follower relationships, so that could be something that lives outside. It could also be extended to add "tags" or "topics" or something. But in most stuff that's built on top of ActivityPub, the accounts live within the new service. I've been looking through projects but it is really hard to tell what is mature and what's just a prototype, etc.
What do you mean with "semantic" search? You mean something that works well to find content not just by full text search but by topic?
Semantic search is for me search that returns not only exact term matches, but topically relevant results. I think that's difficult to do right without personalization, therefore knowing about the user is important.
What I'd like to have is a big graph with people (& communities) and posts (& topics) as nodes (represented by vectors), and links between these nodes. A lot can be built on top of such representation.
Yeah, I see. I agree, I've been talking to people (also as input for fedistats) and what comes up often is that people want an easy way to get relevant stuff, maybe also extending from your current timeline, etc.
Still trying to wrap my head around the ActivityPub specs. It is based on JSON-LD, which I think theoretically links up all those entities you mentioned. But it's hard to extract such very concrete implications from very general specs :)
It says that "[ActivityPub] provides a client to server API for creating, updating and deleting content, as well as a federated server to server API for delivering notifications and content", so I'd say that this provides API to fetch data to build service that provides search and recommendations functionality. There's no mention of some kind of registry of accounts at different social nets, as that's not their target, so APi could be extended to include that. But even "just" building search and recommendation requires crawling a lot of these servers. That seems a bit more technical involved that building such services for centralized social nets.
Yeah, you'll definitely need to crawl many different sites. There are also things that are missing but standard ways of doing things have emerged, for example webfinger for account search, and http signatures for some level of server-to-server authentication.
Eugen Rochko has a pretty good post on the minimal building blocks.
But yeah, I could imagine it works like this: You authenticate with the actor, pull their inbox to see the content they have been getting, compute embeddings and then do the search against other users you know. Or maybe you could extend the Actor object to include embeddings or some representation of topic markers, and so on.
I miss (at least) three things when I go to a new social network: 1\ ability to follow people that I followed in other social nets, 2\ algorithmic feed that knows about me 3\ functional "semantic" search. All three could be developed independently of actual social net used, and that's actually where the most value is.
Yeah, 100% agree! I've been reading the ActivityPub specs and it seems to me the Actor objects https://www.w3.org/TR/activitypub/#actor-objects actually hold the follower relationships, so that could be something that lives outside. It could also be extended to add "tags" or "topics" or something. But in most stuff that's built on top of ActivityPub, the accounts live within the new service. I've been looking through projects but it is really hard to tell what is mature and what's just a prototype, etc.
What do you mean with "semantic" search? You mean something that works well to find content not just by full text search but by topic?
Semantic search is for me search that returns not only exact term matches, but topically relevant results. I think that's difficult to do right without personalization, therefore knowing about the user is important.
What I'd like to have is a big graph with people (& communities) and posts (& topics) as nodes (represented by vectors), and links between these nodes. A lot can be built on top of such representation.
Yeah, I see. I agree, I've been talking to people (also as input for fedistats) and what comes up often is that people want an easy way to get relevant stuff, maybe also extending from your current timeline, etc.
Still trying to wrap my head around the ActivityPub specs. It is based on JSON-LD, which I think theoretically links up all those entities you mentioned. But it's hard to extract such very concrete implications from very general specs :)
It says that "[ActivityPub] provides a client to server API for creating, updating and deleting content, as well as a federated server to server API for delivering notifications and content", so I'd say that this provides API to fetch data to build service that provides search and recommendations functionality. There's no mention of some kind of registry of accounts at different social nets, as that's not their target, so APi could be extended to include that. But even "just" building search and recommendation requires crawling a lot of these servers. That seems a bit more technical involved that building such services for centralized social nets.
These three articles also look interesting, also contrasting how Mastodon (like Twitter) vs Lemmy (like reddit) are using AP.
https://seb.jambor.dev/posts/understanding-activitypub/
Yeah, you'll definitely need to crawl many different sites. There are also things that are missing but standard ways of doing things have emerged, for example webfinger for account search, and http signatures for some level of server-to-server authentication.
Eugen Rochko has a pretty good post on the minimal building blocks.
But yeah, I could imagine it works like this: You authenticate with the actor, pull their inbox to see the content they have been getting, compute embeddings and then do the search against other users you know. Or maybe you could extend the Actor object to include embeddings or some representation of topic markers, and so on.