This started yesterday, a series of insulting songs about me, but he’s been doing it today, on Father’s Day, too, the adorable scamp
And the thing is, these robo-songs are actually very listenable. Apparently you can choose from a wide variety of styles. First the bots took our art, then they took our programming abilities, and now it’s song writing and creation. What is left, creatively, for us humans?
Here’s an example in what’s described as ‘guitar-driven edgy post-grunge’. I would describe it more as early 90s power pop, but whatevs. It even name-checks the Straightdope.
No, he put in some keyword prompts and the AI wrote the bulk of the lyrics, though he is vowing to start writing all the lyrics by himself soon.
The tech is advancing scary-fast. When you say the lyrics are a lot better than what LLMs are able to produce, I think you were thinking of the state of LLMs two or three weeks ago
Did you see Rick Beato’s recent video about it?
I did a chatgpt thing a year or so ago asking “make a song in AABACA in 4/4 time” and it wasn’t too bad.
Here’s a very illuminating video - it’s worth a watch all the way through if you are interested in this kind of thing. I would not be at all surprised if AI does the heavy lifting melody-wise in future music - and you won’t know if it is, as this video illustrates. Because in my opinion it’s a pretty good song. Better than a lot of the stuff out there.
My buddy and I, when we were in college in the late 90s, created this myth that we were in a boy band. It was half-heartedly to “impress chicks” (don’t laugh, it was the peak of Backstreet and NSYNC), but mostly if was just a goof amongst our friends.
Anyway, we’ve been messing around for the past couple weeks with Suno, and holy shit, we’ve put together a halfway decent 90s-era boy band album. It’s freaking bonkers how good this ai music creation is.
I messed with it a good bit after this thread went up.
It can’t make lyrics for shit, but that’s an LLM problem and not a Suno problem. If you’re willing to put a little time into lyrics, it’s very easy to get music that imo is as good as a lot of what’s out there today.
Its biggest weakness is the same weakness that image generators - particularly earlier versions - have. Generate twenty songs in the same genre and you’ll see that it reuses the same structures over and over again, and so everything it makes starts to sound the same. But this will get better as the training sets get larger and the models get more advanced.
It’s also much harder to share a 3-minute song with your friends than a picture, in the sense that you’re asking people to commit time to listening, so it’s not as much fun to mess around with for me.