admin said in #4944 5d ago:
https://cprimozic.net/blog/building-embedding-visualizations-from-user-profiles/
I've been thinking for a while I'd love to represent topical clusters on a 2d or 3d map that users can actively explore. Instead of just having The Algorithm locate you on some manifold and then feed you specially calibrated slop, focus the algorithms on actually illustrating the structure of the manifold visually and by other means, so that you can explore it. This gives the agency to the user instead of the machine. I envision a sofiechan "Atlas" that works like this. Not needed until we have a lot more, but a neat idea that I like to think about.
I randomly came across this article from a guy chasing the same hunch about the value of embedding visualizations. I wanted to share for discussion. Besides the pretty pictures and technical details, I liked this little bit of the philosophy he's taking, which I share:
>Behind the scenes, it's very likely that the recommendation models are using embeddings or something very similar to them to achieve this. Embedding visualizations allow for a self-led and exploratory method for finding new things compared to the one-dimensional and opaque view given by "for you" lists or recommendation engines.
>I have a lot of interest in breaking through the black box of these kinds of opaque systems and piecing together the hidden relationships between things — and I've found embedding visualizations to be ideal tools for this.
Embedding visualization is the high tech superior alternative to the for you page.
I've been thinking for a while I'd love to represent topical clusters on a 2d or 3d map that users can actively explore. Instead of just having The Algorithm locate you on some manifold and then feed you specially calibrated slop, focus the algorithms on actually illustrating the structure of the manifold visually and by other means, so that you can explore it. This gives the agency to the user instead of the machine. I envision a sofiechan "Atlas" that works like this. Not needed until we have a lot more, but a neat idea that I like to think about.
I randomly came across this article from a guy chasing the same hunch about the value of embedding visualizations. I wanted to share for discussion. Besides the pretty pictures and technical details, I liked this little bit of the philosophy he's taking, which I share:
>Behind the scenes, it's very likely that the recommendation models are using embeddings or something very similar to them to achieve this. Embedding visualizations allow for a self-led and exploratory method for finding new things compared to the one-dimensional and opaque view given by "for you" lists or recommendation engines.
>I have a lot of interest in breaking through the black box of these kinds of opaque systems and piecing together the hidden relationships between things — and I've found embedding visualizations to be ideal tools for this.
Embedding visualization is the high tech superior alternative to the for you page.
referenced by: >>4947
I've been thinking f