Message in a bottle

AI Should Make You Feel (like a conductor)

Hello AI engineer at OpenChatGPTMetaSoftAppleZon,

When I look around, I notice AI already making an impact in human labor: ChatGPT writing Facebook ads trying to get me hooked on stupid shit that I don't need, Github copilot writing lines of code that will most likely result in a copyright violations, and Stable Diffusion creating images that otherwise-hired artists would have made. A little over a month ago, Google released a paper describing MusicLM, the AI system which intends to make a similar impact felt in the field of music. As these developments unfolded, I thought to myself that multiple things:

These are the sorts of questions that also popped into the minds of those in the art world as these AI systems proliferated and made their advanced capabilities more apparent. As someone who plans on building more of these sorts of systems for music in the future, I feel that I have a responsibility for how my technology will be used— I certainly don't want my systems to deny current artists and musicians economic opportunities— and I absolutely hope that you feel that you have this responsibility too. But I also feel a sense of excitement because I think these sorts of AI systems are cool; when I see them in action, I feel a sense of childlike wonder that I used to feel when watching sci-fi as a kid.

And I refuse to believe that these two sentiments are mutually exclusive.

This is why I want to echo a design principle for these AI systems from this quarter which I hold dearly: instead of making AI models that replace artists and musicians, make ones that complement them and/or put them in control. I believe that AI which replaces musicians and artists is bound to make bad music and art, no matter how impressive it sounds. Why? Because music and art made solely by AI is music and art without meaning. AI can't distill meaning, and it doesn't have intent or purpose; it's just statistical pattern matching on a complex n-dimensional manifold derived from a pre-existing data distribution at a scale that is unfathomable to the human mind. It can sound technically good with high fidelity, but it doesn't carry the intention of its creator; it doesn't communicate a feeling. In fact, it is the epitome of the derivative. It regresses towards the mean of its own training data distribution. But music and art isn't about that! It's about communicating feelings which the human word can't possibly express and do justice, and it's about breaking conventions which deserve to be broken!

So if AI can't do these things, then ultimately, that responsibility falls unto us, the humans. Thus ideally, these systems should enable their users to be conductors, directors, musicians, setting the stage for what it is that *they* want, with the AI system responding and catering to their needs and desires. If something goes wrong, then the users should tell the AI system to stop, and to correct whatever it is doing. These users should be in their zone, masters in their little corners of the universe where what they say goes. That's what AI should do for the musician.