This is definitely both music and art.
Schoenberg mentions “one-finger composers”:
Harmonizing given melodies is in contradiction to the process of composition; a composer invents melody and harmony simultaneously. Subsequent correction may sometimes be necessary; improvements anticipating later developments and adaptations for changed purposes may challenge the composer’s technique. One might also be obliged to harmonize a melody—a folksong, or one by a “one-finger composer”. Again, this can be done only by one who has been born with or has acquired the sense for the evaluation of harmony.
and there was some discussion earlier about famous musicians who [gasp!] cannot read music (supposedly, Paul McCartney admitted he cannot read or write music and even used some AI to mix in John Lennon’s voice). So it does seem possible to be not classically trained and still be a “music creator”.
Just for fun, we went to the Facebook MusicGen demo and used it to generate some samples. These, looped and distorted beyond recognition, served as timbres for part of a crazy musical soundscape done up in Bespokesynth. The result was as non-musical as anything Eno did, but it is kind of fun to sit there and play with it. After about five minutes the exercise got a little boring… but it does seem to be a quick way of generating semi-novel weird sounds.
I wrote a paper last year for a class on AI-generated sounds, and that was a big takeaway from my review of the literature. Granted in the last year huge gains have been made in AI-generated composition, but at the time, a lot of the thinking was that these snippets of sound would be inspirational.
I do have to say, that I find programs such as AIVA a step too far. It does feel like those programs remove the vast majority of the artistry to the degree that it would be difficult to call the operator a musician/composer. Yes, I’m sure there are some people that are using these programs such as AIVA to kickstart their compositions, but there has been a glut (15 million IIRC) of songs just shoveled out to streaming services that required minimalistic effort on the part of the operator. I’m not opposed to algorithmic music, because the artist needs to design the algorithm to generate the music, and I recognize this is shades of grey, but for me, I don’t like it.
As I said in the other thread, my friends have asked me why I don’t use AI tools (especially as an AI researcher), and my answer has always been “because it wouldn’t be my music anymore.” But I recognize this is a subjective point of view. And it is a broader discussion, because just like with visual arts, there’s the question of where is the training data coming from, copyright, and are we basically going to make human art financially impossible.
My personal theory is that there will be a rejection of AI art on the grounds that I think people want something that has more human involvement, but time will tell.
It is - very few of the musicians I’ve worked with are classically trained, myself included. In fact, I’m barely literate on staff since that’s not how I learned to play and listen to music.
I haven’t played around with Music Gen, but I have toyed with MusicLM, and I agree it’s fun but not there yet. It might have potential use as a quick “prototype” system to see what a wacky combination might sound like without having to fire up a DAW - but that’s about it, at the moment.
I share the same feeling. There’s a HUGE difference, to me, between writing an algorithm/designing an interactive system and writing a prompt. I have the same concerns about the training data, but not copyright. The reason for that is, I’ve communicated with the developer of an AI drum synth (MACE, by Tensorpunk) and asked that specific question. Their reply was that I would own the copyrights on any music (or sample libraries[!]) generated with it. I’ve used it on some experiments, and that’s it.
A related question, is attribution - how do you properly credit an AI, how does publishing work? In general, it seems like a big mess I don’t feel like dealing with more than
I already have.
The main reason I don’t use AI tools, but will play with them, isn’t the ownership question, though. For me, it simply doesn’t meet my goals. As I understand it (and please correct me, if I’m wrong) AI is essentially creating derivative works. As a sound artist, for lack of better term, I have exactly 0% interest in new twists on old things - I prefer exploring the interactions of control and chaos.
I agree.
I stuck my big fat nose in this thread, so I may as well contribute.
I have no experience whatsoever with any sort of AI music generator thing. I also have no experience whatsoever with ChatGPT. I do, however, have experience AI art programs. Here is my reasoning:
I like to create AI Art, because, on my own, I can’t create anything close to what I can get out of Night Cafe or Playground AI. On the other hand, I like to think that I can write pretty well and I can play music pretty well, so I don’t have the slightest interest in having an algorithm do it for me.
Just messed with it for another few minutes. Instead of making it output crazy sounds, I gave it a text prompt of “hip-hop loop”, and that’s what you get (you can also specify a tempo, key, etc.). There is also a model where you can play it a melody and it will work with that. Problem is, it is far from working in real time, at least on the kinds of graphics cards I have lying around, so at least for now, it can’t be used for, e.g., micing in input and getting auto-accompaniment. Under the hood it is playing an imitation game: a transformer model trained on 20,000 hours of recorded music, so in that sense it is not creating “new things”, but it may be possible to screw with it (or a different or more complex model architecture) to get unexpected behavior 
, e.g., directly manipulate the encoding they use to generate sound to use as a novel form of sound synthesizer.
So, something similar to Magenta’s Continue or Interpolate functions, albeit with audio vs. MIDI?
Are you aware of Google Colab? Until two weeks ago I didn’t have any GPU available, and was using that to start to mess around with LLMs and then MIDI.
It’ll be interesting to see how things develop, that’s for sure. I’ve tried feeding MusicLM prompts such as “heavy metal saxophone solo”, and it went off the rails. On the other hand, it did better (compositionally) with more traditional forms, such as jazz or synthpop. Results are here, sampling rate limited on download to 22.5kHz by TestKitchen: MusicLM 01
I can see that, but what if you design the generative method/algorithm/AI, does that change? For me it does.
I’ve thought about it recently (due to this thread), and for right now, the only way I can see myself actually incorporating AI in a work, is an AI vocalist. However, that doesn’t mean one of those celeb voicechangers - I’m thinking something along the lines of Synthesizer V.
I have no idea how your singing skills are, or if you need vocals. On occasion, I do. Do you think it could be helpful being able to compose a melody line on a guitar and have it sung? Or, would you see the extra step(s) as onerous? I’m genuinely curious.
I don’t think there’s anything with using them. Especially for non-commercial use. In fact, I think AI tools will be really useful for musical education. But there is line, and I don’t know exactly where that line is, where it stops being the operator’s art. There is always some degree of human involvement because a human is coming up with the idea for the prompt, so there is something creative coming from the mind, but is it enough (also, a human prompt generator could be replaced by an AI)? A human has to curate the works and decide what to select? Is that enough? If a human gets a score and changes a single note is that enough? If not one, then how about if they change half the notes minus one? It gets very tricky.
I also like using AI art generators like Midjourney, but I would never call myself a visual artist based on it. My friend takes Midjourney images and improves them using his vast experience in Photoshop. And again, we enter into that grey area.
AI is really forking everything up. Personally, I blame the AI researchers. Curse their oily hides!
I had an idea last night, and due to the proper application of technology, it’s now available: The Hand Of Man
Today, I think I’ll make a Blender visualizer video for this…
Hehehe, and now I have a bunch of modules that I simply have to download to see what in the world they do.
I really enjoyed your song
Thanks! Here’s the Visualizer for it - I don’t think Pixar has anything to worry about! 
@scabpicker - I recommend starting with the Clockd and Meander modules, then going from there. If you want any ideas, I’m happy to share some of my favs.
How did you do the visualizer?
I miss all the cables even though there were not enough. 
I did that using Blender. I’ve been watching tutorials by a guy named Ducky3D to learn how to make them for myself. I’ve got animators and vid editors in my Rolodex (do people still say that? I’m might be old…), but I don’t like bothering them with my little experiments.
For this one, I took a sphere, applied some metallic paint to it, add a still image (factory) so there’d be something for it to reflect. Then, I imported the music (full-spectrum, I also can specify frequency ranges to use) and subdivided the sphere to get different segments. Then I used an F-Curve to modulate (for lack of a better term) the segments’ displacement on the Y-axis. Then, I render, which goes surprisingly fast on this new, re-purposed gaming box, and stitch the audio and video back together using DaVinci Resolve.
I’ll make sure the next piece has more cables…there’s never enough! 
It’s interesting applying modular generative and routing techniques to a ‘live orchestra’.
I’ve been working on a squirrel war song (yes you read that right) and it coming along well. I finished up for the night, and I thought, I wonder what AIVA could make for a squirrel war song.
Well, I wasn’t able to figure out how to use to create the song, but I did have it write a folk song for me.
Its ok. But now I understand the appeal for people. If you have absolutely zero ability to create music, then it will create something reasonably pleasant (although for it seems to LOVE long intros on the bass).
And for those inclined, if you pay for the premium edition then you get 300 songs a month that you can shovel up to Spotify, etc. and start collecting royalties. Having such a massive output greatly decreases the need for massive numbers of streams. I know there are AI vocal creators out there as well. I didn’t play around with that. Although I now understand why they’re supplementing the AI generated music with AI listen bots.
Could AIVA be useful as a source of inspiration? Maybe. There was one song it created where I liked the concept. It was a sad song and I’ve been a little sad lately so maybe that’s all it is. Because it matched my emotional state I “liked” it. I guess I shouldn’t say maybe. I don’t think there’s any doubt that it could be used that way. The music is good enough to give a starting point but you would be constrained to using it as a starting point if that makes sense.
For example, it is unlikely to ever generate a piece that would make me go “OOooooo squirrel war song!” So I would be forever writing whatever the generated piece inspired me to write, instead of writing the music for my inspiration.
Still, it is spooky how good it is, even if it is “only” ok. We are a LONG way from the Illiac Suite. Wow.
Does AIVA just do “whatever”, or can you tell it exactly what you want: what scale to use, what melodic intervals are allowed, what chord progressions, rhythmic constraints, etc.?
I didn’t play around with it for that long so keep that in mind. This is based on like 30-60 minutes of playing around with it.
You can certainly have it do whatever. That was the first thing I did. I just clicked “Create” a composition.
You can select things like genre, instruments, chord style (jazz, cinematic, etc.), percussion style, mood, etc. In my subjective opinion, it doesn’t quite rise to what would be considered algorithmic music. I definitely felt far more removed from the output than I do when I use things like VCV rack , which requires a lot of careful planning to get something decent.
For example, I selected, write me a sad folk song using piano, guitar and double bass, and it would do it. But the output, while sad, was not specific to anything I had in mind other than the very broad notion of a sad folk song. This is a key difference I think compared to other approaches to generating music.
Interesting, I’ll have to play with AIVA and see what I can pull out of it. But, I must hear the squirrel war song when you’re done with it!
In theory you could do the loadup Spotify thing much easier with AIVA, but can’t that already be done now with Splice or other loop sources? I get this is generated, instead of assembled, but I don’t see too much difference in those processes, myself, when it comes to mass-churning product for stream distribution.