We should be able to select different fully open source algorithms from a drop down menu, and load custom ones from fediversealgorithmmenuwithdescriptions dot org, including “no algorithm”.
I assume that’s like a billion hours of work, but, goals.
When you sort your feed by hot vs top vs new, that’s already what you’re doing kinda.
But the platform has to have the data to support the algorithm, so you can’t just “load in” whatever algorithm you want. Besides, that sounds like a security nightmare for the platform lol
I mean, with the fediscovery project, people can make centralised applications from fediverse data (people who opt in) this makes indexing and other stuff that works better centralised possible.
If I understand correctly, that only works with data publicly available (or at least available to 3rd party instances). But there are going to be metrics that fediverse platforms simply don’t make public or even track.
for example: i dont imagine that peertube (or even loops) makes public who viewed which videos, when, for how long. and it’d be a huge privacy issue if they did. Even tracking things like who-liked-what are the kinds of things that a 3rd party probably shouldn’t be able to just check.
without these kinds of insights, it’d be hard to make a good recommendation algorithm, because you can’t really tell how an individual is interacting with content.
The client doesn’t have access to all the videos, it’ll have to query the (distributed???) platform to find matching videos.
And the platform is gonna need to serve up some kind of metadata for the client to track (are hashtags enough???)
And the platform is gonna need to be keeping stats on the videos for the client to match against
And what would that query even look like? A gargantuan weighted map of hashtags? I can’t imagine hashtags alone would be enough.
A centralised app, like an ATproto appview is what I meant by client. It could use a fediscovery api to index videos and then it could save a users like/watch history.
Remember when Musk took over Twitter and “open sourced” the algorithm, although it was impossible to reconstruct anything from what was given, and contained clear signs of being edited and incriminating details suggesting content categorization and prioritization?
What I really want to see is Facebook’s algorithm, because it seems to just produce a neverending stream of alt-right bullshit.
I was mistaken, it was not opensourced - there was a whitepaper on a recommendation system from tiktok’s parent company bytedance, and everyone just assumed it was the tiktok recommendation algorithm when it was published.
Well, its definetely possible on activitypub. Every “app” built on the atprotocol takes data from a relay’s firehose and then indexes it and does all the algorithm stuff. There is a project (https://www.fediscovery.org/) that will let people build centralised apps with fediverse data. Although, I could just make an algorithm that looks for keywords a user may be interested in, in the posts database and show it to them, it just wouldn’t have every post to its disposal.
Personally, I think there’s room in the Fediverse for an app with a “strong algorithm” provided it’s completely open ofc.
My biggest issue with algorithms isn’t the fact they exist, but that they’re proprietary black boxes so no one truly knows how it’s being manipulated
We should be able to select different fully open source algorithms from a drop down menu, and load custom ones from fediversealgorithmmenuwithdescriptions dot org, including “no algorithm”.
I assume that’s like a billion hours of work, but, goals.
“No algorithm” would load nothing at all. Everything is an “algorithm,” including listing all posts in chronological order.
Wanting “no algorithm” is like wanting food with “no chemicals” in it and not realizing that carbs, fats, proteins, etc. are “chemicals.”
Only eating noble gases.
Noble gases are chemicals too, damn it!
The only thing that isn’t “chemicals” is literally just vacuum.
Only a dehydrogen monoxide addict will say shit like that!
I would say freedom, fallacies, and magic are not chemicals.
They have at least little tendency to participate in chemical reactions.
When you sort your feed by hot vs top vs new, that’s already what you’re doing kinda.
But the platform has to have the data to support the algorithm, so you can’t just “load in” whatever algorithm you want. Besides, that sounds like a security nightmare for the platform lol
I mean, with the fediscovery project, people can make centralised applications from fediverse data (people who opt in) this makes indexing and other stuff that works better centralised possible.
If I understand correctly, that only works with data publicly available (or at least available to 3rd party instances). But there are going to be metrics that fediverse platforms simply don’t make public or even track.
for example: i dont imagine that peertube (or even loops) makes public who viewed which videos, when, for how long. and it’d be a huge privacy issue if they did. Even tracking things like who-liked-what are the kinds of things that a 3rd party probably shouldn’t be able to just check.
without these kinds of insights, it’d be hard to make a good recommendation algorithm, because you can’t really tell how an individual is interacting with content.
A client could be made on one, that tracks those things, and then suggests videos from that.
I don’t think it could.
The client doesn’t have access to all the videos, it’ll have to query the (distributed???) platform to find matching videos.
And the platform is gonna need to serve up some kind of metadata for the client to track (are hashtags enough???)
And the platform is gonna need to be keeping stats on the videos for the client to match against
And what would that query even look like? A gargantuan weighted map of hashtags? I can’t imagine hashtags alone would be enough.
A centralised app, like an ATproto appview is what I meant by client. It could use a fediscovery api to index videos and then it could save a users like/watch history.
Remember when Musk took over Twitter and “open sourced” the algorithm, although it was impossible to reconstruct anything from what was given, and contained clear signs of being edited and incriminating details suggesting content categorization and prioritization?
What I really want to see is Facebook’s algorithm, because it seems to just produce a neverending stream of alt-right bullshit.
Fwiw Tiktok apparently just open sourced their algorithm a week or so ago.
I wonder if loops will provide it as an option
Do you have a link? I can’t find it.
I was mistaken, it was not opensourced - there was a whitepaper on a recommendation system from tiktok’s parent company bytedance, and everyone just assumed it was the tiktok recommendation algorithm when it was published.
deleted by creator
Well, its definetely possible on activitypub. Every “app” built on the atprotocol takes data from a relay’s firehose and then indexes it and does all the algorithm stuff. There is a project (https://www.fediscovery.org/) that will let people build centralised apps with fediverse data. Although, I could just make an algorithm that looks for keywords a user may be interested in, in the posts database and show it to them, it just wouldn’t have every post to its disposal.
Hey thanks for the correction - my mistake!
It’s gonna happen in the AT protocol I’m sure.