A quick AI experiment

I grew up at just the right time, to have access to broadband when the internet was still wild. Ever since then, I’ve ignored the Spotify’s and relied on an every growing collection of mp3’s. Every new pc starts with the same ritual: install the venerable Foobar2000 as a music player, point it to the collection, and wait for it to load in the folder structures it needs to allow searching. I always wait for the first album to pop up, and for the 20 years, it’s been 50 Cent’s ‘Get Rich or Die Tryin’, the only time I ever listen to it.

It’s simple, it works, it’s perfect. Except for one problem: my phone. I use local files there too, but I only have space for a tenth of my collection. Once in a blue moon I’ll upload some new music, but it’s a hard to navigate the collection if you’re not looking for anything specific. I end up scrolling for a while, picking things at random, ending up not listening to most of it.

So I wondered if AI could help me with it. I told it to gather 100Gb of music. Maintain a genre split similar to my listening history (stored by last.fm). Bias it towards critically acclaimed albums, but allow merely good or mediocre ones. Add albums I listen to a lot, but also add things I never listen to. And convert lossless files to mp3’s (something I didn’t bother with before). I also told it that 30% should be full albums suited for running, with my existing running playlist as a guide.

It worked first-try ish (most problems were uploading so many files, ambiguity of greatest hits collections, and some bad folder structure on my part). It grabbed genres and rating online, faster than a copy of a file. The whole process took about 5 prompts and 90 minutes. I spotchecked the output, and it looks fine. Which is an excellent result.

So why did it work well?

• I specified. I didn’t ask it for just a ‘varied’ set of albums. I told it to look at critical reception, and have a spread. Same for ‘famousness’ of the artist (eg 33% famous, 33% known, 33% obscure)
• I pointed to existing data. I had an export from lastfm listing everything I ever listened to, of which it inferred genres, and the great shortcut of also giving me partially albums I already listened to a lot
• I had a huge margin of error. Even if 75% of the albums don’t appeal to me at all, I still have 200 that do. I think this, if anything, sets the bar low enough

Ultimately, I spend 20 minutes of my own time on it, so I would call it a success. I first put on my wireless headphones, paired with my phone, I immediately hear a coin drop. Usually it continues the previous played song, but I deleted the entire folder to make room for the new list. It must have just played the first song it could find, in alphabetical order. 50 Cent’s ‘Get Rich or Die Tryin’.

… it also gave me Drake