Rendered at 19:07:22 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
netruk44 1 days ago [-]
I’ve been playing around with mapping steam games + their reviews based on Qwen embeddings. I revisit this project every year around this time to see what improvements I can make to make a “better” map than last year (According to my own preferences). [0/1]
One thing I learned this year is that naive text mapping (embed a corpus of reviews with a Qwen embedding model, use umap to plot, then compare where things fall) primarily compares embedded text length more than anything else. What a review says affects the position less than how long the review is.
This was a little surprising to me. For whatever reason, I didn’t think the embedding would so strongly describe the length of the embedded passage. So most of this year’s time was spent fixing that artifact.
was one of very few useful results of the first decade of TREC research... And the main thing it addresses is balancing various factors which could cause a full text search engine to privilege either long or short documents.
Roxxik 24 hours ago [-]
Not sure how related the cause is, but the fix seems to apply too.
In single cell genomics you often have a large correlation between the data and the length of the sequencing reads.
You can calculate the correlation between your metadata and every PC dimension and exclude the highly correlating dimensions before doing the umap. Similar can be done for other metrics. I've seen it done for patient gender for example. I could imagine looking at the correlation for publish year or genre.
Edit: fixed a misconception
tomaskafka 1 days ago [-]
I have played with similar problem in different domain, and if you could have “which users play which games” (or in general “which articles are read by which people”), you can do really good similarity search and self organizing map based on no content at all, just the metadata.
Which is of course why Facebook can say they don’t read your conversation content, while still be able to target ads precisely based on conversion metadata.
Alulim 1 days ago [-]
This is very cool! The web pages didn't run well on my phone, but that was easily fixed using Opus 5.5 and a local fork. Thanks for creating this!
Jtarii 1 days ago [-]
Interesting, a way to filter by number of reviews and review score would be nice.
netruk44 1 days ago [-]
Yes, I agree. Those are coming. I only added the current filters and pins to the 2026 map yesterday.
(Edit: They're live now, you'll need to hard refresh/clear your cache to get the updated HTML and JS)
Joel_Mckay 1 days ago [-]
In general, ranking relevant keywords in the context of the Negative reviews will likely have more hits the longer people complain.
The random fluctuations of individual experience only cancel out bias if and only if the results are truly unbiased normal distributions.
In the case of positive:negative ratio feedback scoring, it only makes sense if you ignore the entire message body of what people enjoyed.
Negative-weighted ratio ranking more accurately describes the end users experience. Any game with paying unique customer greifer total scores under 15% without game-play reliability related complaint keywords is going to be a fairly good game if the genre matches your profile preferences.
Otherwise, people will just see shit scores from the response bias of pissed off users on dodgy OS/Networks. Would recommend checking your conclusions with the average current active players in the SteamDB plugin. If users aren't engaged with the paid gameplay experience 6 months after launch, than the release likely isn't fun or broken garbage. =3
A good game, but takes some skill to play (4270 daily players 10 years later):
The ratings displayed on the map are taken directly from Steam's review summary on each game's store page. I don't assign a rating to a review based on its text, and then combine that into a single game rating. That would be very flaky, as you explained. The text is only used to locate the game on the map.
I also believe (though I don't have proof) the rating that Valve generates only considers the binary "Recommend" and "Do not recommend" options that you can pick from before you write a review.
plagiarist 1 days ago [-]
This is an interesting idea, do you find it has more interesting properties than doing cosine similarly by reviewers? Have you found things adjacent to games you enjoyed that were also fun to play?
netruk44 1 days ago [-]
Yes, I’ve actually found a few interesting games I had never heard of before! So the project has been a success for me already.
But there’s still plenty of places where games are placed strangely on the map, and I like to investigate why that’s happening (it’s usually because Steam reviews are pretty noisy signal).
The first version of this idea was actually based on cosine similarity. The user would enter arbitrary text or an existing steam app id, and would get a list of games back.
The arbitrary text results were never very good because to get related results you needed to write a fake review of the kind of game you were looking for, which is a difficult concept to explain to users. Most users just entered text like "beautiful game", and that lead to bad results. "Not a beautiful game" winds up very close to that.
The results based on steam app id were much better, because it was comparing like-for-like. That lead to me considering different ways of presenting the games based on their review similarity. And so this map was made.
The umap projection does introduce a fair amount of distortion, so it's not a strictly better result than cosine similarity between specific Steam games. I'm still playing with the umap parameters to try and find something which causes the least problems.
plagiarist 23 hours ago [-]
That original idea also sounds interesting to me, if you modified it to work in a semantic space where "beautiful" and "not beautiful" differed by more. Write a hypothetical review and it finds the closest match. It's a very cool project, thank you for sharing this with us.
Scene_Cast2 1 days ago [-]
The methodology feels either heavily undercooked to the point of being misleading, or heavily under explained. Why are we in 3D? At which point did dimensionality reduction hit? And if we're doing dimensionality reduction, does the "plane" really mean anything?
vhold 1 days ago [-]
It's under explained. It's in 3D because of PCA. It mentions PCA all over the place, but never even says what it is, less alone briefly describe what it does.
There are a lot of more fancy ways to reduce dimensions and create visualizations like these that tend to preserve more of the higher dimensional structure.
I'm very familiar with PCA. There are so, so many ways you can apply it to this problem.
Assuming the simplest interpretation of what the analysis did, then it's saying "the variance of the biggest variance direction stays constant". Which in itself is an interesting thread to tug on, but the charts aren't how you'd present such a result.
thece0 1 days ago [-]
[flagged]
pohl 1 days ago [-]
> Why are we in 3D?
Isn’t there an obvious answer to this? It’s about visualization. There’s a natural upper limit to dimensions we can see, and 3 is where we hit it.
oldjohnnylane 1 days ago [-]
There’s no meat here. It’s just someone throwing data into a plotting tool with no meaningful results or conclusions. I was expecting a meaningful analysis. Nope.
supermatt 1 days ago [-]
Isn’t this last layer analysis basically saying that after a bunch of Chinese tokens you get a Chinese token, and after some Latin tokens you get a Latin token? That seems expected given the setup. Would be interested in seeing what translations look like in the middle layers.
About a year ago, I made a very simple language detection model from the token embedding tensor of gemma 3. I used UMAP to show the projections of how languages separated into a 2D space. It's very interesting to look at IMO.
I wish there had been a third broad language style such as arabic although perhaps Qwen doesnt use arabic.
It made me think what if totally different knowledge bases could exist in a model where as my mental model was similar to OPs translation from pure knowledge source.
It also explains why rewriring prompts to another language sometimes lets you bypass restrictions.
mcswell 1 days ago [-]
Also, Arabic script is used for a number of non-Arabic languages, including Urdu and Farsi (Persian). It would be interesting to see if the projections of those languages have similarities as a group to the group using Latin script.
Another interesting (to me) comparison would be between Hindi (using Devanagri) and Urdu (using Arabic script). The two languages are very closely related, being largely mutually intelligible in their spoken forms.
levocardia 1 days ago [-]
Even the blog post is 100% AI-generated
big-chungus4 22 hours ago [-]
Using Qwen 3 is a dead giveaway of AI generation with no oversight. Qwen3.5 2B hasn't reached knowledge cutoff yet
numpad0 1 days ago [-]
So what does that all say? Everything except Chinese are English or LLM personality is hard split between Chinese and everything else?
Definitive and undeniably digestible proof for linguistic relativity in a single picture?
Or something else a lot less controversial? Are there other visualizations that bunch of things cluster apart beautifully like the olympic games flag?
supermatt 3 hours ago [-]
Nothing so interesting as that. It is literally the last layer, so all it is saying is after chinese characters you probably get a chinese character, and after latin characters you probably get a latin character.
There is no interesting internal model state being shown here at all, just what the model is about to output. Nothing to do with "personality", just the token space itself, which is pretty obvious (latin script languages use latin characters, chinese use chinese) and not very noteworthy.
Sophira 21 hours ago [-]
I'm just a layperson in this field, but I'm wondering if part of the reason for the separation could be the cultural expectations set by the language.
After all, to be able to understand a question being asked, you have to understand the culture that it's presented within, and in a lot of cases, being asked a question in a specific language tells you what cultural expectations that person has... and a lot of people tend to think of "Western culture" as a homogenous blob.
Of course, it could also be simple that all of the tested languages, apart from Chinese, use the Latin alphabet. I have no idea, I'm just throwing out ideas here.
It would be interesting to see how, say, Swahili (as a language that uses the Latin alphabet but whose cultural heritage is distinct from "Western culture") compares here.
big-chungus4 22 hours ago [-]
I swear, next time someone uses Qwen 3, I will start committing atrocities
jrmg 1 days ago [-]
Naive question:
This is visualizing end states, right? That’s why there are no ‘paths’ between the orange initial cloud and the ‘result’ clouds?
Wouldn’t we expect the languages to be ‘separate’ by that point? The output is necessarily different after all - and ‘more different’ for Chinese and English than for other Latin script languages and English.
It feels like it would be more interesting to see the intermediate states. Do questions about the same topics in Chinese and English ‘traverse’ similar areas in the vector space?
Am I completely misunderstanding?…
1 days ago [-]
slickytail 1 days ago [-]
I think the raw tokens embeddings shouldn't have been included in the PCA (past the first or second graph). Since they're disjoint from the conversation embeddings, you're essentially losing a dimension and compressing all the conversations into a plane. There may well have been much more obvious structure in the language embeddings that you'd be able to see.
crib-award-cont 1 days ago [-]
The language comparison angle is interesting; a broader set of languages would make this even more informative.
One thing I learned this year is that naive text mapping (embed a corpus of reviews with a Qwen embedding model, use umap to plot, then compare where things fall) primarily compares embedded text length more than anything else. What a review says affects the position less than how long the review is.
This was a little surprising to me. For whatever reason, I didn’t think the embedding would so strongly describe the length of the embedded passage. So most of this year’s time was spent fixing that artifact.
[0]: 2025’s map: https://storage.danieltperry.me/share/steam-map/
[1]: 2026’s map (still a work in progress, both the ui and the review processing are incomplete): https://storage.danieltperry.me/share/steam-map-refresh/
https://en.wikipedia.org/wiki/Okapi_BM25
was one of very few useful results of the first decade of TREC research... And the main thing it addresses is balancing various factors which could cause a full text search engine to privilege either long or short documents.
In single cell genomics you often have a large correlation between the data and the length of the sequencing reads.
You can calculate the correlation between your metadata and every PC dimension and exclude the highly correlating dimensions before doing the umap. Similar can be done for other metrics. I've seen it done for patient gender for example. I could imagine looking at the correlation for publish year or genre.
Edit: fixed a misconception
Which is of course why Facebook can say they don’t read your conversation content, while still be able to target ads precisely based on conversion metadata.
(Edit: They're live now, you'll need to hard refresh/clear your cache to get the updated HTML and JS)
The random fluctuations of individual experience only cancel out bias if and only if the results are truly unbiased normal distributions.
In the case of positive:negative ratio feedback scoring, it only makes sense if you ignore the entire message body of what people enjoyed.
Negative-weighted ratio ranking more accurately describes the end users experience. Any game with paying unique customer greifer total scores under 15% without game-play reliability related complaint keywords is going to be a fairly good game if the genre matches your profile preferences.
Otherwise, people will just see shit scores from the response bias of pissed off users on dodgy OS/Networks. Would recommend checking your conclusions with the average current active players in the SteamDB plugin. If users aren't engaged with the paid gameplay experience 6 months after launch, than the release likely isn't fun or broken garbage. =3
A good game, but takes some skill to play (4270 daily players 10 years later):
https://store.steampowered.com/app/219990/Grim_Dawn/
A similar genre game, but people score it higher (3650 daily players 1 year later):
https://store.steampowered.com/app/1145350/Hades_II/
A top ranked genre game, with a nostalgic series scoring bias (9651 daily players released this year):
https://store.steampowered.com/app/2536520/Diablo_II_Resurre...
I also believe (though I don't have proof) the rating that Valve generates only considers the binary "Recommend" and "Do not recommend" options that you can pick from before you write a review.
But there’s still plenty of places where games are placed strangely on the map, and I like to investigate why that’s happening (it’s usually because Steam reviews are pretty noisy signal).
The first version of this idea was actually based on cosine similarity. The user would enter arbitrary text or an existing steam app id, and would get a list of games back.
The arbitrary text results were never very good because to get related results you needed to write a fake review of the kind of game you were looking for, which is a difficult concept to explain to users. Most users just entered text like "beautiful game", and that lead to bad results. "Not a beautiful game" winds up very close to that.
The results based on steam app id were much better, because it was comparing like-for-like. That lead to me considering different ways of presenting the games based on their review similarity. And so this map was made.
The umap projection does introduce a fair amount of distortion, so it's not a strictly better result than cosine similarity between specific Steam games. I'm still playing with the umap parameters to try and find something which causes the least problems.
https://en.wikipedia.org/wiki/Principal_component_analysis
There are a lot of more fancy ways to reduce dimensions and create visualizations like these that tend to preserve more of the higher dimensional structure.
Here's a semi recent paper discussing and comparing some of them, I didn't dig too deep to try to find a more recent or better one : https://pmc.ncbi.nlm.nih.gov/articles/PMC12453773/
Assuming the simplest interpretation of what the analysis did, then it's saying "the variance of the biggest variance direction stays constant". Which in itself is an interesting thread to tug on, but the charts aren't how you'd present such a result.
Isn’t there an obvious answer to this? It’s about visualization. There’s a natural upper limit to dimensions we can see, and 3 is where we hit it.
https://huggingface.co/blog/dleemiller/wordllama-detect
If you just want to see the plot directly: https://cdn-uploads.huggingface.co/production/uploads/65ff92...
It made me think what if totally different knowledge bases could exist in a model where as my mental model was similar to OPs translation from pure knowledge source.
It also explains why rewriring prompts to another language sometimes lets you bypass restrictions.
Another interesting (to me) comparison would be between Hindi (using Devanagri) and Urdu (using Arabic script). The two languages are very closely related, being largely mutually intelligible in their spoken forms.
Definitive and undeniably digestible proof for linguistic relativity in a single picture?
Or something else a lot less controversial? Are there other visualizations that bunch of things cluster apart beautifully like the olympic games flag?
There is no interesting internal model state being shown here at all, just what the model is about to output. Nothing to do with "personality", just the token space itself, which is pretty obvious (latin script languages use latin characters, chinese use chinese) and not very noteworthy.
After all, to be able to understand a question being asked, you have to understand the culture that it's presented within, and in a lot of cases, being asked a question in a specific language tells you what cultural expectations that person has... and a lot of people tend to think of "Western culture" as a homogenous blob.
Of course, it could also be simple that all of the tested languages, apart from Chinese, use the Latin alphabet. I have no idea, I'm just throwing out ideas here.
It would be interesting to see how, say, Swahili (as a language that uses the Latin alphabet but whose cultural heritage is distinct from "Western culture") compares here.
This is visualizing end states, right? That’s why there are no ‘paths’ between the orange initial cloud and the ‘result’ clouds?
Wouldn’t we expect the languages to be ‘separate’ by that point? The output is necessarily different after all - and ‘more different’ for Chinese and English than for other Latin script languages and English.
It feels like it would be more interesting to see the intermediate states. Do questions about the same topics in Chinese and English ‘traverse’ similar areas in the vector space?
Am I completely misunderstanding?…