Comments URL: https://news.ycombinator.com/item?id=48207281

Points: 2

# Comments: 3

https://toposonico.com/#lon=14.4313&lat=-1.0200&z=9.10&entit...

That's it, not the funniest project for sure but hope that's useful to someone.

repo: https://github.com/peppedilillo/gcn-notice-parser docs: https://peppedilillo.github.io/gcn-notice-parser/

Comments URL: https://news.ycombinator.com/item?id=48192163

Points: 1

# Comments: 1

This is toposonico, a music recommender and navigable map. At core it's a skipgram word2vec model trained over ~6M playlists. Tracks are embedded in a 128d space. Embeddings for albums, artists and labels are computed marginalizing over tracks. The 2D map was built with UMAP.

Both the model and UMAP were trained in the cloud over a NVIDIA A100. All things considered it cost me around ~50EUR, over two main training sessions and a few experiments. For the slippy map I experimented with a few libraries. Ended up with Maplibre GL JS. Loved working with it, kudos to their developers. For the recommender indexes I used FAISS, another great piece of software. Pretty happy with the thing running on a small and cheap box. It was the first project where I used LLMs extensively. Still I wouldn’t say this was vibecoded: getting the thing to work as I wanted was the usual pain in the arse and I ended up rewriting by hand multiple times. Mostly Claude on Pro plan when the plan was still usable, then I switched to codex (which eventuay became unusable too, but that’s a another story).

It still has some problem. For example very big artists often ends up very disconnected from the rest of the embeddings. I guess this has to do with the high number of mono-artist playlists these ends up on. But I’m not very sure still, and would like to know if anyone else faced a similar issue and how it was fixed. I’m interested in any kind of feedback on UI/UX. Love working on these but, coming from a scientific computation background, have little experience.

Repo link: https://github.com/peppedilillo/toposonico

Comments URL: https://news.ycombinator.com/item?id=48035569

Points: 2

# Comments: 0

This is toposonico, a music recommender and navigable map. At core it's a skipgram word2vec model trained over ~6M playlists. Tracks are embedded in a 128d space. Embeddings for albums, artists and labels are computed marginalizing over tracks. The 2D map was built with UMAP.

Both the model and UMAP were trained in the cloud over a NVIDIA A100. All things considered it cost me around ~50EUR, over two main training sessions and a few experiments. For the slippy map I experimented with a few libraries. Ended up with Maplibre GL JS. Loved working with it, kudos to their developers. For the recommender indexes I used FAISS, another fantastic piece of software. Pretty happy with the thing running on a small and cheap box.

Two things influenced me in making this. The first is decade-old idea: human navigation and exploration skills work in information spaces too. Many ML concepts fit this idea especially well. It would be nice to see more experiments in this direction. The second goes more like a story. Before moving out and selling my turntable, I used to visit record fairs and I always ended up finding something novel there. To find a new record you didn't sit down and listen to ten records in a row, selected based on a supposed model of your personality. You wandered around, speaking with dealers and looking through the crates they brought with them. The crates often leaned on some genre more than another, reflecting the dealer's history and taste. There were huge stalls and there were small ones. Some were crap. Finding oddities was very easy. I wanted to write something that felt like that.

Let me know what you think, if you find any bugs or if you have any idea for improving it.

Repo link: https://github.com/peppedilillo/toposonico

Comments URL: https://news.ycombinator.com/item?id=47985369

Points: 4

# Comments: 2

For the moment the page is reachable at:

http://ec2-16-171-47-81.eu-north-1.compute.amazonaws.com/

Note the http there, not https, I have yet to make a certificate. Hopefully hn will not kill the link. The content is made-up faker bs.

Repo is here:

https://github.com/peppedilillo/ist

Any feedback from people more experienced than me with webdev is appreciated. bests

Comments URL: https://news.ycombinator.com/item?id=42264891

Points: 5

# Comments: 1

Comments URL: https://news.ycombinator.com/item?id=41568050

Points: 5

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=39384998

Points: 1

# Comments: 1