Computers Still Learning

I was reading a piece about GPT and humor the other day (which I won't bother to look up again to link because I don't remember where I saw it and all it said is what I'm about to tell you) that said that these large language models are good at explaining jokes, but not good at making new ones. It's interesting how those are different projects. Analytical understanding of why something is funny really is different from the ability to craft an emotional experience that will provoke laughter. It's easier after the fact to process why it worked, and harder to do the thing.

Whatever Spotify is using to categorize songs is imperfect in a similar way. It likes to do 'mood' playlists, and it often produces notable failures. For example, on today's "Romantic" playlist the first song was this one:


That's a beautifully sung song, but it is not a romantic song. It's about the heartbreaking failure of a romance.

Likewise, they recently had a "Happy" playlist that headlined with this piece:


This song kind of sounds happy, because of the band's signature strategy (which I believe they adopted from the Blues Brothers) of marching through blues pieces in major chords. It's definitely not a happy song, though: it's about a man who is dying of alcoholism, and finds himself powerless to stop it.

I guess it's good that it's still easy to fool the Terminator.

No comments:

Post a Comment