algorithms exist for a reason

%at=2025-12-06T12:34:56.789Z

this article is made for people that already know what vrchat is (and preferably already played it). if you haven't, take a look

#author_luna #recommendation_algorithms #vrchat

I will be using the word "algorithm" to represent recommendation systems. as such, here's a quote from wikipedia:

A recommender system (RecSys), or a recommendation system (sometimes replacing system with terms such as platform, engine, or algorithm) and sometimes only called "the algorithm" or "algorithm", [1] is a subclass of information filtering system that provides suggestions for items that are most pertinent to a particular user.

how large is the internet #

while I wasn't there for the early days of the internet, I think we (me and you, the reader) can agree that the internet has reached an absurdly meteoric rise in its use, capabilities, and just general size. the precursor systems of the internet were just a couple of universities, but it's now insane. from the Facts and Figures 2025 from the UN International Telecommunication Union [^1] :

Today, in 2025, 74 per cent of the world’s population are online, compared with 71 per cent a year earlier. In absolute numbers, this corresponds to 6 billion people, up from 5.8 billion in 2024. Year-on-year growth increased slightly to 3.3 per cent, up from 2.9 per cent previously.

6 billion people is a lot of people, almost every country is connected to the internet in some way shape or form, and the fact it even works (somewhat) is a modern miracle.

from ix.br (original pt, translated en) in April 2025:

The IX.br (Brazil Internet Exchange) has just reached another record, registering 40 Tbit/s of aggregate traffic across the 38 locations where it operates.

the main point I want to bring here is that a lot of people connected to the internet invariably leads to more data, either being pushed or pulled just because there's a larger amount of people.

while the threshold for "too large" varies from person to person, I think we can agree that it's effectively impossible for any random person on the internet to ingest and process every single piece of information on the internet (in either image, video, audio, text, etc), this has been already the case for multiple decades and may sound obvious to you but this builds up to further points in this article.

as such, if someone wants to process information on the internet in general they MUST have some way to deny data. in this case a "filter" also works conceptually. I think we (me and you) can agree with this because of how large the internet is (as shown above), the scale of global data is increasing everywhere and given we (as humanoids) can't read faster than the collective power of the world, we must design systems that let us drop data without looking at the data ourselves.

and that's where I think the first search engines were born, out of this deep asymmetry between a singular person's ability to look at data versus how much there is. maybe it was possible for you to read "literally everything on the internet" back in the day, but after a small while it probably (I wasn't there) would be impossible for a random person to just acknowledge everything.

and yet, the internet grew, with the search engine paradigm growing to be one of the primary ways people access information on the web (and in this case, the rise of Google as a large player including all the ramifications of that, good and bad). but from a philosophical perspective (not from the past or current execution of Google themselves), I think we can agree search engines are a good thing for the world. they provide value in information filtering without which it would be impossible for anyone to look anything up.

from filtering to recommendation #

I'm not a historian, so I can't build much of a timeline. but I can say that my first interactions with a recommendation algorithm were the early social media websites, namely Facebook. in those websites (or Platforms) there's the same kind of problem which can happen at a much smaller scale than the internet itself: too much data, can't look at everything, must filter.

the separation from a "filter system" like early Google and a "recommendation algorithm" like Facebook is (in my opinion!) the explicit personalization of the type of content it may provide based on the type of user that is looking at stuff. it's hard to draw a proper line because nowadays Google can provide a lot of personalization itself so we can't just put Google into either box and close it off.

you generally want this for systems (like social media) where you don't have a full global view of everything. nobody follows literally Everyone on Facebook, but when you're searching Google you definitely are thinking from the perspective of "I am searching the entire internet". optimizations into better results look extremely different between Google's PageRank or Facebook's "algorithm" (I haven't gotten a proper paper from Facebook about how it works, so a nebulous document from them may help show the difference. the former is ranking pages globally and their relationships, with the latter is ranking based on your graph, your relationships, your preferred subjects)

someone on Facebook (or any other social media) would not enjoy having an unfiltered Global view of everything because of the same reasons Google exists to filter the internet based on a user's query: too much data, impossible to keep up, and not everything on the internet is something they would like.

effective. Power #

information filter systems, by definition, hold a lot of power over whoever is using them, and that is by intentional design! because of that a large amount of scrutiny is usually placed on those types of systems to ensure they're behaving correctly according to the metrics set by the designers of those systems. keep in mind I am not saying "ethically" here. that is not required. a recommendation algorithm's goal is whatever the designers said it should be, if Amazon's algorithm is driving you to give more money to Amazon, it wins. if YouTube's algorithm is making you give (or they extract from you, depending on your personal relationship with YouTube as a platform) more hours of your life watching them, it's working as intended.

a lot of contention happens because of that. different people have different views of the world and when they see a big bright shiny lever (metaphor! nothing is that simple. modern recommendation systems are absurdly complex) they see the ability to impose their views onto thousands or even millions of people, and that is caused by the fact usually these systems are deployed on a global scale (towards all of that system's users).

I don't think that's an inherently good thing. while in most cases you may want a singular system that filters everything for everyone, I think most platforms should start to ponder about systems that acknowledge the hyperplural world beyond just personalizing at the global system.

as practical examples, bluesky adopts something along those lines at the product level, and while it isn't perfect the fact it exists at all leads to possibilities that weren't possible before. as another example of alternative social media though on a more negative light, Mastodon (and Mastodon-following Fediverse implementations) went with a singular system: your Following feed, which relies on you getting some good social relationships which lets you crawl out to other people (there's the Global view, which depending on your instance is an absolute waterfall of information going at rapid speeds, exactly what you want to prevent -- some people can keep it under control with herculean effort.)

and this brings me to the main point of this article: VRChat.

VRChat Worlds #

this becomes the second half of this article. and from here onwards I assume the reader knows what VRChat is, has used VRChat, and knows the distinction between Worlds and Instances[^2].

VRChat is weird.

in VRC, the primary piece of content is a World. when you first start it up you enter the tutorial World, and when you login afterwards you default to the "VRChat Home" World (which you can change to something else, etc). this makes a World very different than a Website indexed by Google or a Post that someone makes on Facebook.

Worlds can be constantly updating, like Websites. or they can be static and immutable like a Post. Worlds can be ephemeral single-time things you go to, like Posts, but at the same time they're not required to be short like a Post. you can experience a World for multiple hours, literally non-stop, because the world is telling a story, or there are many trinkets in the world, or you (with your friends) started talking a lot without even interacting with the world itself.

because of that and even more factors I'm not thinking about, worlds are VERY hard to categorize. while you can create a classification for specific types of worlds you see out there, nothing can be 100% exact. yet, like Websites, there are a LOT of worlds and not enough time to actually look at all of them. this happened before, and will happen again to platforms that contain user-generated content (either on the Web with Websites, social media platforms with Posts, YouTube with Videos, and VRChat with Worlds).

on World Discovery #

how can you recommend them, then?

at present VRChat has approximately four methods by which Worlds get shared to users:

  1. in-product algorithms ("Popular", "New & Noteworthy", "Recently Updated", "Community Labs", etc)
  2. worlds containing portals to more worlds (also known as repositories)
  3. organically (in-platform person-to-person)
  4. organically (out of platform. example: social media to person)

all of these intertwine with each other: someone can find a world on VRChat's default algorithms (1) and share to someone (3) who posts on social media (4) which gets indexed by someone on their repository world (2), which then leads to another person finding a World, etc...

due to the nature of Platforms and how algorithms have power by looking like the "singular levers" that they are, a lot of discourse has spawned out of it. it's a pattern that I personally see again and again: different people have different opinions that can't be reconciled or compromised with. some people are okay with the current state of things, others believe that specific Worlds are gaming the system (and they can), some wish it was more varied or personalized. all of these concerns are valid but the avenue that they attempt to push forward is the lever metaphor: "VRChat must fix this, VRChat must make the algorithm more varied, VRChat must prevent this from happening", etc.

the argument I'm making here is that it's impossible to make everyone happy. algorithms by definition are filters, and because of the nature of filters someone will be left out of the calculation based on some metric (visits, aesthetic quality, technical merits, you can create dozens of objective or subjective (!!!!!!!!) metrics Worlds can optimize for), and because Worlds are much more complex than a Website I would say it's very hard, technically, to automatically process Worlds standalone into a recommendation system[^3].

I would say that due to the complexity of Worlds and the complexity of trying to find Worlds based on their standalone merits (outside of descriptions and tags set by the World creator, which are also problematic[^4][^5]), any kind of algorithm for Worlds should exploit the natural graph of people coming and going throughout Worlds. instead of trying to create a "one-size-fits-all" solution which may look perfect in the eyes of the person trying to design it (and then after a couple of months may just see it as incorrect due to some other subjective metric and then we're back at square one: everyone wants to pull the lever), solutions should consider the fact people are already curators of information (see world discovery methods 2/3/4).

my plea is for VRChat to integrate such methods which are based on the relationships between people and what people like beyond strict objective metrics (like aggregate playtime, or visit count, etc), because creating a subjective metric that works for everyone is impossible.

examples of this MAY involve:

I don't think any of these are silver bullets to the problem, but I think they're good ideas to build from. if someone likes to go through the popular stuff then they can, but if someone wants a specific "vibe" of World and knows people that align with such vibes, they can follow them and find more Worlds for that vibe (which may also change with time! not everyone in VRChat wants the same exact thing every time they log in[^6]). this leaves the problem of finding a specific metric in the hands of the user (they choose which levers to pull on a per-user basis), while also attempting (that relies on execution of those ideas, which could fail) to not overwhelm the user with too much information. if they want to interact with those systems then more power to them, if they don't then it's fine (sorta, lol).

thanks for reading! email vr at l4 dot pm or bluesky @l4.pm if you have any feedback or suggestions

sidenotes #

[^1]: while the UN ITU may or may not be a trustworthy source on the matter, anecdotal evidence from everyone that has been around for the past decade and a half can attest to how large the internet has become

[^2]: on the VRC half, I intentionally didn't talk about Avatars and Avatar Discovery as a concept because it's a differently shaped problem. Avatars deal with a constant identity and the design goals for an "Avatar Discovery" are much more different than World Discovery, imo

[^3]: at the very least you want a bot that walks through a World to figure out what it is about, which is a currently unsolved problem in the field of computer science. there may be specialized bots/agents that do something on some game, but VRChat Worlds are way too varied for this to work consistently (and most importantly at low cost)

[^4]: there's good worlds with descriptions in japanese. I'm an english-speaking user, if we rely purely on textual content of Worlds we're just splitting things off

[^5]: side-side-note: japanese worlds can still be played by english speakers: a broom is always going to look and (hopefully) act like a broom, and there is shared UI design across VRChat Worlds you can figure out what is a button (and then explore to find out what buttons do, etc)

[^6]: in the vrchat world that fixed me, strasz describes VRChat in a way I resonate with (timestamp 17:50 but I recommend the whole video): "The truth is more akin to what is described throughout MOBIUS: Making friends, sharing stuff with them, maybe making more than friends, having things not work or work and finding yourself growing or not growing in the process. It's very normal. It's very life. That is really something most media both built within VRChat and also externally about VRChat fundamentally miss about this place. VRChat is really just, I don't know, hyperreality, but still reality"