Critical response #1: Stanford CS 470 / MUS 356, Music and AI
Prompt: What do you (really) want from AI music generation?
As someone who quit their job to work on generative audio about a year and a half ago, the last few weeks have been filled with embarrassment (that I didn’t make the tools everyone is talking about) and nervousness (that the train has left the station on making impactful contributions). Putting personal feelings aside doesn’t offer much relief because the questions that are forced into the conversation around music feel more centered on corporations taking a slice of the pie from artists than they do on making creative tools for musicians.
In Music and AI it feels like there are two separate conversations
- What are the implications of technologists having a growing impact on the music creation process?
- What are the fun/playful/creative ways we can use technology to make music?
To me, talking about these questions in the same breath is challenging because the emotions they evoke are so different. I guess my goal in this post will be to move the conversation forward to the point where I can talk about the future of technology in music in a more hopeful way.
Phase 1: Optimism
Okay! Technology can do amazing things and is so fascinating! I’m a forward thinking musician and there’s all these SWEs trying to make tools for me, how cool! I live in the 21st century so I’m really only good at one aspect of making music. WLOG let’s say I can sing. That’s great because now I can use my creative inputs — a 15 second melody that I sang into my usb microphone — to condition MusicLM2 as it makes me a “guitar and drum backing track in the style of green day and 100 gecs.” Only in 24k and filled with the noise of a machine learning algorithm generating audio at the sample level? No problem! I will simply use Landr to make this appealing to the masses. A few clicks on distrokid later and I’ve become a TikTok star with a record deal that has Universal thinking “man, I hope this kid can play a live show.” Of course there’s no hope of me playing a live show… I make all my music on my iPhone 16 and haven’t collaborated with another musician since high school choir.
Phase 2: Trying to be Logical
If I had to list 3 major ways technology has influenced music in the last 150 years I would probably pick
- Ability to record: artists have new revenue streams and can collaborate. people listen to music while they make dinner.
- Electronic synthesis: new sounds drive musical styles which evolve faster than before. Music production becomes data driven as we quantify what makes a song exciting
- Streaming / subscription model: more artists reach an audience, but their revenue must come from live, people listen to more and a wider variety of music
At every step, new tools stick when the technologists that build them are able to share revenue. This typically happens when more people listen to and spend money on music. What implications of AI generated music fit into this framework?
- More people have the capability of producing a compelling song
- Separation of music for art and music for sync
In order to become mainstream, the business of AI music will to transition from demos that excite programmers to products that interact with content that everyday people enjoy.
Phase 3: MusicLM
The biggest questions in AI music right now are
- Who will own the rights to AI generated content?
- Will people listen to AI generated music?
One surprise about MusicLM is that, as far as I can tell, it’s not trained on establishment (owned by major label) music. I think right now it still exists in this space where people are excited by developments but nothing’s good enough to stand on its own. The question of “how much better would these models be if they were trained on top 100 hits” is interesting because it sways the answer to “who will own the rights to AI generated content.” From WaveNet to Jukebox, Riffusion, and now MusicLM, AI generated music has evolved dramatically in the last 7 years. To me, however, it hasn’t started its ascent up the other side of Masahiro Mori’s uncanny valley.
At the same time, I think very few people are listening to the results of MusicLM more than a few times. Equivalently, nobody turns on these outputs when asked to play music in the car. With the question “would you watch Atlas Dance,” I’m asking if, after you get over the fact that it’s an amazing feat of engineering to be able to make a bipedal robot that can do flips off scaffolding, you would say “that is beautiful” as it hops around the stage with Tchaikovsky playing in the background.
Phase 4: What do you really want from AI generated music?
It’s too easy to say that I just want music to be left tf alone because truthfully I like the fact that I can listen to recordings, hear genres change, and do it without thinking about whether I want to spend $10 to check out a new artist. I want AI to enable new creative outlets and increase the capacity of people to make music. I want it to change the business of music to the degree that OpenAI gets a grammy and Robert Glasper has an h-index that a professor would gawk at. I want it, and I think it has the opportunity to revitalize our appreciation for musicians as a class of people who express themselves creatively through sound and communicate specific experiences in general terms in a way that moves listeners.
Most of all, however, as a consumer, I want it to leave music the f*** alone. I want it to be something that sits in the background as latent as the space it interpolates from. I’m sure that most recorded music will have components that are AI generated but I don’t want to know it’s there. I don’t want to think about it, I don’t want it to try to steal the show, I just want to listen to music.
