You knew w what? I am actually into this where the algorithm used sre published
latest app build showing invalid on my android
i guess loops is browser based now
This is outstanding!
Not being based on “rengagement” or “monetization” means it’s purely interest-based, with a touch of serendipity.
One of BSKY’s distinctive features was to have “pluggable” algorithms. Fediverse would do well to support it so people who are not into the technical weeds could choose how their feed is curated.
Well. This is good news.
Isn’t loops open source?
Yes
So then this is just a shorter explanation, and not “publishing” the algorithm? Because before already anyone could read the code and understand how it works, it would have just taken longer.
I suppose that by using the word “publish”, I made it sound like the algorithm had been private until this point. I could probably have chosen a better phrasing like “published breakdown illustration”.
Technically, yes the info would be available but not everyone is able to read code so this is a welcome additional bit of info and transparency. I also see it as an example of how future open source social media platforms can promote and demonstrate the way their algorithms work as opposed to the black boxes of proprietary web apps like Instagram etc (which, while not algorithmically transparent or open, have been pretty well established to have algorithms that prioritize maximizing time on platform and engagement over all else with some serious negative repercussions for that).
anyone could read the code and understand how it works
“Anyone” is doing some heavy lifting here.
Sorry for being pedantic since your comment is true anyway, but “could” is the one doing the heavy lifting.
Personally, I would not be able to understand this, looking at the code (or any code). I appreciere that they shared a visualization of it. If you’re hung up on the word “publish” instead of “share”, “posts” or “shows”, then I think its redundant
The same way “anyone” could make an HTML5 compliant browser. The reality is you can’t do it unless Google lets you.
Can you personalize your own ranking algo that can be shared with friends? That’s the breakthrough we’re all waiting for…
Not that I know of, but you could make a feature request if one doesn’t already exist https://github.com/joinLoops
That would be extremely groovy. Even just having a few flavours of algorithm that you can choose from would be really cool.
that’s kinda how bluesky works with the feeds and all
I can’t even zoom into the picture on mobile on this website
this is from the post, that this post links to
Yeah after making an algorithm for peertube I’m gonna say this is too complicated to implement. That graphic might be the the plan and it’s something more simplistic.
Making a simple cosign vector for peertube was a pain
“I failed at making something, therefore someone else could certainly not make something more complex than the thing I failed at”
K
I didn’t fail, I actually succeeded. Plus judging by how hard it was to do a simple algorithm an algorithm like this would take way too much effort.
Given from the infographic they didn’t even think about an algorithm let alone worked on it
Seriously though, it’s literally already implemented. Might be too complex for you to implement, but apparently not for them.
How do you know it’s implemented as described in the picture?
How do you know it’s not?
I already explained how I know. So you can answer my question or just dodge?
Saying you tried something simpler and it was hard is not proof that this implementation is a fraud. If anything, it just makes you look less credible yourself. If you have specific aspects you can point to that are impractical, and the technical reasons they wouldn’t work, feel free, but it’s not a dodge to say that the diagram makes sense and this team had done a lot of impressive work already.
“Because I did something substantially simpler than this, and struggled to get it done, they must be lying.”
K
Lol you really don’t understand recommendation algorithms 🫵🤣
K
What’s too complicated exactly? They claim their program does precisely that. I believe you could review the code yourself.
oh lord your looking for a fight. i tell you this i kinda already won since your looking my input to criticize, but ill say this the personalized focus block alone. that’s an exact “thing” and if you want more well your gonna have to ask more specific questions and your gonna demand that i help you.
also I helped loops in the past. I can tell you that the GitHub code isn’t the code that’s used
Not trying to fight, just wanted to get some perspective of someone critical in a thread where everybody else is very positive. I didn’t read your username and didn’t notice I replied three times to the same guy.
Nice so you got my perspective then?
I can tell you that the GitHub code isn’t the code that’s used
really? given that the license is AGPL and they do have some external contributors, they shouldn’t be running an unpublished branch of the code!
Yeap, turns out you can post junk code on GitHub and say it runs whatever. I never seen anyone independently verify that the GitHub code runs similar to loops or compiles the code for the app
You’re*
Hoping for something like this for PeerTube and personal customisation of the algorithm.
This infographic reeks of AI slop.
What about it?
I’m not too happy to spend time pointing out flaws in AI slop. That kind of bullshit asymmetry feels a bit too much like work. But, since you’re polite about it, and seem to ask in good faith…
First of all this is presented as a technical infographic on an “algorithm” for how a recommendation engine will work. As someone whose job it is to design similar things, it explains pretty much nothing of substance. It does, however, include many concepts that would be part of something like this, with fuzzy boxes and arrow that make very little sense. With some minor trivial parts you can assume from the problem description itself. It’s all just weird and confusing. And, “confusing” not in the “skill issue” sense.
So let’s see what this suggested algorithm is.
-
It starts out with “user requests the feed”, and depending on whether or not you have “preference” data (prior interests or choices, etc), you give either a selection based on something generic, or something that you can base recommendations on. Well… sure. So far, silly, and trivial.
-
“Scoring and ranking engine”. And below this, a pie diagram with four categories. Why are there lines between only the two top categories, and the engine box? Seems weird, but, OK. I suppose all four are equally connected, which would be clearer without the lines. Also, what are the ratios here? Weights for importance, of some sort? “Time-Decayed”? I hope that’s not the term that stuck for measuring retention/attention time.
-
On the three horizontal “Source Streams” arrows coming in from the left, its all just weird. The source streams are going to be… generated content, no? But let’s give it the befit of the doubt and assume it’s suggesting that, given generated content, some of it might can be considered relevant for “personal preference” and has a “filter: hidden creators”, but, none of that makes any sense. The scoring and ranking engine is already suggested to do this part… The next one is “Popular (high scores) filter: bloom filter (already seen)”. Which mixes concepts. A bloom filter is the perfect thing to confuse an LLM, because it has nothing to do with filters in the exact same context “filters” was used for the above source stream. Something intelligent wouldn’t make this mistake. But, it does statistically parrot it’s way to suggest that a bloom filter might have something to do with a cost effective predicate function that could make sense for a “has seen before”. However, why is this here?
I’ll just leave it at that. This infographic would make a lot of sense if it was created by some high schoolers who were tasked to do something like this. Came up with some relevant sounding concepts. Didn’t fully understand any of them. Which is also exactly the kind of stuff LLMs do
I don’t think loops hired a bunch of kids, so LLM it is.
And the like “Our new For You algorithm is pretty complex, so we created this infographic to make it easier to understand!”, doesn’t help the case against LLM either. There a many complex parts of a recommendation engine, but none of the things in this infographic explain or illuminate those complex parts…
But, I might be wrong, and this is their earnest attempt at explaining how their algorithm works. In which case, they are just bad at either explaining it, or at designing it, most likely both. Then again, if I’m right, and this is generated by an LLM still gives the same impression, but leaves some room for “someone who isn’t technical, asked an LLM, and phoned this in because it looked cool, and people who don’t know any better will think so too!”
Ty for the effort post. It’s all french to me so I was looking for arrows to nowhere, crooked lines, and messed up text.
Happy to hear. Cheers
Just because you overanalyzed something to the point of confusing yourself does not mean that it is AI slop, or equally confusing for others.
To address the specific points you raised as “evidence” of AI:
- The two top categories have lines going to them because those are the things that a user controls with their activity on the platform. Prior to that, the “for you” recommendation engine is not active, since it has nothing to base it’s recommendations on. Seems pretty clear to me.
- Time decayed, in the context of that category means when you last interacted with a post. If you haven’t interacted with a post for a while, it will no longer show up in your for you feed. Again, really quite straight forward.
- What about filtering hidden creators makes no sense? You hide a creator, they don’t show up in your feed. That’s one aspect of personalization, from the start, the rest of it is the two categories that, once they make it past the “hidden creator” filter, determine how likely it is to show up.
- Bloom filter is literally explained right there, it’s if you have seen a post yet or not. Lemmy clearly does not have this sort of filter, because you keep seeing the same shit over and over until it drops off from whatever category of the feed you’re viewing. Really not sure what is hard to understand there.
You’re using a lot of fancy words in your analysis here, but the actual analysis is nonsensical. Almost makes me wonder if you yourself are actually a bot.
I think you might have missed my point. I wasn’t listing stuff I had trouble understanding. I was listing stuff that didn’t make much sense. The distinction is relevant. The end result, even if you manage to find some excuse that extends the already generous benefit of doubt, it still doesn’t result in anything useful or informative.
I’m also not using fancy words (or…?). The only fancy thing that stands out is the the “Bloom filter”, which isn’t a fancy word. It’s just a thing, in particular a data structure. I referenced it because its an indication of an LLM, in behaving like the stochastic parrot that it is. LLMs don’t know anything, and no transformer based approach will ever know anything. The “filter” part of “bloom filter” will have associations to other “filters”, even tho it actually isn’t a “filter” in any normal use of that word. That’s why you see “creator filter” in the same context as “bloom filter”, even though “bloom filter” is something no human expert would put there.
The most amusing and annoying thing about AI slop, is that it’s loved by people who don’t understand the subject. They confuse an observation of slop (by people who… know the subject), with “ah, you just don’t get it”, by people who don’t.
I design and implement systems and “algorithms” like this, as part of my job. Communicating them efficiently is also part of that job. If anyone came to me with this diagram, pre 2022, I’d be genuinely concerned if they were OK, or had some kind of stroke. After 2022, my LLM-slop radar is pretty spot on.
But hey, you do you. I needed to take a shit earlier and made the mistake of answering. Now I’m being an idiot who should know better. Look up Brandolini’s law, if you need an explanation for what I mean.
I just explained how the things you claim don’t make sense, do in fact make sense. Saying “this does not make sense” implies you don’t understand it. I have seen plenty of AI slop, and this is not it.
You didn’t use the term “bloom filter”, the diagram did. I know what it is, and it makes perfect sense in the context, so it’s really weird that you would claim it doesn’t. The fancy words I was referring to was “predicate function” and “asymmetrical”. Both are jargon words/phrases that don’t add anything to your statement as far as illuminating your point, but make you sound smart.
The thing to me that is not really amusing at all, but very annoying, is when someone has experience in a technical field, but then think that experience makes them an expert in every other field that might be tangentially related, and uses that assumption to pedantically (and often erroneously) dissect and dismiss the work of others.
Let me ask you this tho. When you say “do in fact make sense”. Are you basing it that in the context of what you think this diagram is saying? Or do you mean “do in fact make sense” in the context of knowing how such an algorithm would be constructed?
You still keep missing my points. And they aren’t difficult points either. The fancy jargon words were a basic ass description of what a bloom filter does. So you’re kinda making my argument, which is funny for reasons I’m sure won’t be appreciated.
I’m also not tangentially an expert, for fucks sake. I’m the kind who’s day job is to design simpler things than what this diagram is trying to “explain”, and telling you, that it comes across as if made by with a toddler’s understanding. I also didn’t say this was 100% guaranteed to be LLM, I said it smelled like it. I have suggested other possible explanations: stupidity, incompetence, and even a mental stroke.
Your take on being tangentially an expert might be a woosh moment
I’m also out of shits to give at this point. Literally.
Do me a favor here, as a self proclaimed expert. Define a bloom filter, and then explain to me, a stupid pleb, why it would not work in this context. Cause from everything I have read on them, the description in this diagram is literally what it is used for.
-
probably because the block under the cpu looking thing doesn’t indicate how it interacts with the cpu looking block and the block that ranking engine feeds into the ranked “for you” feed also there’s two user controls
It seems that the pie chart under the cpu describes the weights of video characteristics that push to the top of your algorithm. But that’s a guess, and it should be clearer than that if the platform wants to be transparent.
“Everything I don’t like is AI”
That’s way too reductive.
Cooool
Loops… Algorithm???
… yes?
That is surprising for those who have a long familiarity with Loops. The chronological timeline was a major selling point.
A calm chronological Following feed for the people you trust — and a For You feed that surfaces new creators from across Loops and compatible ActivityPub servers.
Loops kind of sucks.
Compared to…?
I honestly ask because out of all the short-form video apps I know of, it sucks the least.
No u
I mean, tell me I’m wrong. Tell me that there is good content. Tell me it functions well.
It’s still in early stages. Give it time :)
You can do something about the content 😊
I tried. The UI sucks.
I think the stuff i post is pretty neat… Ben Glish is on there now too, that’s cool!
It’s literally in its infant stages. You think any of the major social media sites just started out with a bajillion accounts and a neverending feed? I remember signing up for Facebook and Instagram in the very early days. It looked like the fediverse.
It’s kind of in beta.
