The bull case for AI music

by Crispy Chicken

I am on record as telling people that they’ve been sniffing glue when it comes to claims that AGI is just around the corner, and I stand by that. I do believe AI is going to have an enormous impact on the economics and affordances of our world, though.

Here’s one important case-study: music.

The bear case on AI music is simple to make: it’ll be technically flawless, but it’ll revert to the mean in melody, lyrics, texture, and anything else important.

To be fair, this is what 99% of people who try to make AI music work will do, and it’ll suck and no one will care about those people.

But a few people are going to realize a simple recipe

  1. Improved versions of models like Jukebox are going to capture certain styles well enough and be good enough at making interesting” discoveries in the space of possibilities to be sifted through by paid or unpaid crowds to be useful at getting the music part right.
  2. Improved versions of models like GPT-3 are going to be able (already?) able to generate eerily captivating and lucid snippets of insight when conditioning on relevant information…which is really all you need for good lyrics.
  3. Combine and profit.

What I’m describing here is fastmusic: music generated to order to target niches, such as nascent subcultures, as soon as they arise. Being there first can matter a lot: you can become a schelling point for the Elders and sneak in your own ideas a bit if your clever. It’s a whole new kind of branding.

Companies may find these subcutltures through quantitative or qualitative methods. They may choose to target only large subcultures or it may be possible to serve-up custom ads with custom music to a few dozen folks. And many songs may only be listened to once.

It really doesn’t matter—because if producing these kinds of aesthetic snipe attempts is cheap enough i(n terms of hardware to run big neural networks on) there will be plenty of money in using it to create schelling points that you and your friends socialize around, even if only one song in a million hits.

Isn’t that what a lot of gpeople are doing now, in place of AIs?