3 ms·

I also don't want mood playlists, the majority of the time. I'm saying I use AI generations for exactly the opposite — so that I can explore and listen to thing

I also don't want mood playlists, the majority of the time. I'm saying I use AI generations for exactly the opposite — so that I can explore and listen to things that are more intellectually interesting in the ways I find intellectually interesting, because the human music it's easy to find is common denominator music. The way you describe AI ('continuous stream of auditory stimuli') is the way I'd describe Spotify. Sure, you could use AI to make a faux Spotify, but, like, why would you? The popular stuff already has saturating supply, and it will sound much better than an AI generation.

moritzwarhier 1y ago
I was also having Spotify in mind when writing about algorithmic curation and GenAI in music ;) Regarding this: > I'm saying I use AI generations for exactly the opposite — so that I can explore and listen to things that are more intellectually interesting in the ways I find intellectually interesting I just have not found any AI music that would satisfy this description. But I am very interested in failure modes of GenAI. Especially in Suno, it was cracking me up at times. I'm also sure there will be a space for interesting and/or challenging music generated with neural networks involved. But I don't see any revolution here so far. Care to share examples of AI-assisted music you find interesting? To elaborate, I don't find jarring or curious combinations of cliches interesting. AI could not invent a new style, it seems to me. To repeat this point. And I've never had any problem finding interesting music. Key to me is diving into labels, artists and their philosophy, after I got interested into particular ones (the other way around doesn't work for me). I adore discogs.com for that. Regarding interviews and stuff, there's sadly a huge decline in quality written material about music, I feel. "Lowest-common-denominator music" is exactly what Suno produces, at least in my ears. I could go on and list music I like, but generally avoid that. Wait, I'll do it anyway for a bit... at the moment, I like Punctum - Remote Sensing EP (Caterina Barbieri) and AtomTM vs Pete Namlook - Jet chamber LP just for example I also love so much other music. To me, such music is miles apart from the slop I heard from AI. I heard there's research into generating music in the style of JS Bach as well. How's that going? I'd guess: probably bot too well, because the genius of Bach is not only in complexity, or counterpoint rules. His music is very emotional to me (at least the portions I like). And, like any good music, it has moments of surprise. It's not just a formula, or a "vibe", or a "genre". Could AI create a new Techno, a new Blues, a new Bossa Nova? I doubt it.
- Veedrac 1y ago
  I've been avoiding trying to class what counts as 'interesting' because it feels like the wrong point to make, for a medium so entwined with personal preferences that are so often at odds with each other. To single out my tastes feels like it would be saying 'paleontology is the true interesting subject, unlike architecture'. The tracks you listed here aren't really my thing, but I can see why AI audio doesn't help with it. That's fine. I will also repeat that I'm well aware that the best stuff is definitely all human. It's not my genre either, but traditional composers like Bach certainly made extremely interesting, clever, even deeply-studiable pieces and AI 'in the style of' those composers surely won't capture much of that. There's a lot of stuff AI can't do wholesale; one particularly strong example is if you're Jacob Collier, AI is not going to make the complex harmonizations and song structures there. AI is pretty bad at these textural or instrument exploration things like from Collier above or Mike Dawes or Yosi Horikawa or Yoko Kanno or Keiichi Okabe. There's a bunch of music I listen to because it's generically a genre or mood I like and it's well produced, which I won't list here, and AI audio can often do stuff like that at baseline but not especially well. There's also nostalgia; I'm also certain a huge part of the reason I like the Celeste soundtrack so much is in part that I liked the game so much. But then there's a whole category of music I listen to where the texture is supplemental to the part that defines it, like most of Acapella Science or Bug Hunter or Tom Lehrer. Eg. Prisencolinensinainciusol isn't interesting to me because it's musically complex; the part I care about is that it's a listenable execution of an idea, not precisely how it was executed on. I don't keep coming back to I Will Derive by some random schoolkids recorded on a potato 17 years ago annually because it's sung well or they were particularly clever with how they took another song and changed the words; I come back to it because it's fun and reflects for me onto a part of my past that I remember fondly, and these things make me happy. All these words and I've still only addressed half the comment. Ok, let's consider the idea that it's not enough for AI audio to facilitate the creation of interesting musical pieces, and it instead has to create whole interesting musical styles. I take issue with this in a bunch of places. I don't reject artists who I judge not likely able to create a new Bossa Nova. I judge artists based on whether the output they produce is something I want. I do the same for AI. I also think the question about whether AI could 'create' a new style is somewhat misplaced. A style is a cultural centroid, not just a piece of audio. AI can definitely create new musical textures or motifs, but it's always being pulled towards the form of what it's being asked to produce. As long as we're talking about systems that work like today's systems, the question still needs to involve the people that are selecting for the outputs they want. Could that connected system create something as distinct and audibly novel as a new genre? Yeah, probably, given time and a chance for things to settle. That's a different question from whether it'll do so to an inspecific prompt thrown at it.

moritzwarhier 1y ago

I was also having Spotify in mind when writing about algorithmic curation and GenAI in music ;) Regarding this: > I'm saying I use AI generations for exactly the opposite — so that I can explore and listen to things that are more intellectually interesting in the ways I find intellectually interesting I just have not found any AI music that would satisfy this description. But I am very interested in failure modes of GenAI. Especially in Suno, it was cracking me up at times. I'm also sure there will be a space for interesting and/or challenging music generated with neural networks involved. But I don't see any revolution here so far. Care to share examples of AI-assisted music you find interesting? To elaborate, I don't find jarring or curious combinations of cliches interesting. AI could not invent a new style, it seems to me. To repeat this point. And I've never had any problem finding interesting music. Key to me is diving into labels, artists and their philosophy, after I got interested into particular ones (the other way around doesn't work for me). I adore discogs.com for that. Regarding interviews and stuff, there's sadly a huge decline in quality written material about music, I feel. "Lowest-common-denominator music" is exactly what Suno produces, at least in my ears. I could go on and list music I like, but generally avoid that. Wait, I'll do it anyway for a bit... at the moment, I like Punctum - Remote Sensing EP (Caterina Barbieri) and AtomTM vs Pete Namlook - Jet chamber LP just for example I also love so much other music. To me, such music is miles apart from the slop I heard from AI. I heard there's research into generating music in the style of JS Bach as well. How's that going? I'd guess: probably bot too well, because the genius of Bach is not only in complexity, or counterpoint rules. His music is very emotional to me (at least the portions I like). And, like any good music, it has moments of surprise. It's not just a formula, or a "vibe", or a "genre". Could AI create a new Techno, a new Blues, a new Bossa Nova? I doubt it.

Veedrac 1y ago
I've been avoiding trying to class what counts as 'interesting' because it feels like the wrong point to make, for a medium so entwined with personal preferences that are so often at odds with each other. To single out my tastes feels like it would be saying 'paleontology is the true interesting subject, unlike architecture'. The tracks you listed here aren't really my thing, but I can see why AI audio doesn't help with it. That's fine. I will also repeat that I'm well aware that the best stuff is definitely all human. It's not my genre either, but traditional composers like Bach certainly made extremely interesting, clever, even deeply-studiable pieces and AI 'in the style of' those composers surely won't capture much of that. There's a lot of stuff AI can't do wholesale; one particularly strong example is if you're Jacob Collier, AI is not going to make the complex harmonizations and song structures there. AI is pretty bad at these textural or instrument exploration things like from Collier above or Mike Dawes or Yosi Horikawa or Yoko Kanno or Keiichi Okabe. There's a bunch of music I listen to because it's generically a genre or mood I like and it's well produced, which I won't list here, and AI audio can often do stuff like that at baseline but not especially well. There's also nostalgia; I'm also certain a huge part of the reason I like the Celeste soundtrack so much is in part that I liked the game so much. But then there's a whole category of music I listen to where the texture is supplemental to the part that defines it, like most of Acapella Science or Bug Hunter or Tom Lehrer. Eg. Prisencolinensinainciusol isn't interesting to me because it's musically complex; the part I care about is that it's a listenable execution of an idea, not precisely how it was executed on. I don't keep coming back to I Will Derive by some random schoolkids recorded on a potato 17 years ago annually because it's sung well or they were particularly clever with how they took another song and changed the words; I come back to it because it's fun and reflects for me onto a part of my past that I remember fondly, and these things make me happy. All these words and I've still only addressed half the comment. Ok, let's consider the idea that it's not enough for AI audio to facilitate the creation of interesting musical pieces, and it instead has to create whole interesting musical styles. I take issue with this in a bunch of places. I don't reject artists who I judge not likely able to create a new Bossa Nova. I judge artists based on whether the output they produce is something I want. I do the same for AI. I also think the question about whether AI could 'create' a new style is somewhat misplaced. A style is a cultural centroid, not just a piece of audio. AI can definitely create new musical textures or motifs, but it's always being pulled towards the form of what it's being asked to produce. As long as we're talking about systems that work like today's systems, the question still needs to involve the people that are selecting for the outputs they want. Could that connected system create something as distinct and audibly novel as a new genre? Yeah, probably, given time and a chance for things to settle. That's a different question from whether it'll do so to an inspecific prompt thrown at it.