Jukedeck

[Aleatoriac]

AI making music. I found one reference on a forum search, and that was with the comment: "scary" - though it may have been in jest.

https://www.jukedeck.com/share/34ed83ee ... c819febfa5

It's a bit... a bit like modern background cinematic music, no real melody. But I like it.

I want AI to make things for me. To help me. To fix my (and all Humans') sloppy heuristic approach to logic.

I love it. I'll indulge in the apex of Maslow's hierarchy of needs as long as I can, you f**king self-abusing sadistic world.



But what do YOU think about AI?

[woggle]

I think a lot of AI is really over-rated. You can make really simple software that will make music that sounds like music in all sorts of styles. Here is an example
https://soundcloud.com/greghooper/interludic-1

Doing something clever is much harder. The most interesting stuff i have heard is from an Australian composer Luke Harrald, who models performers rather than music and then lets the performers improvise together

[Aleatoriac]

Luke Harrald, who models performers rather than music and then lets the performers improvise together

Thanks! I'll check into that, that's the kind of thing gets me right off.

Computers are still sophisticated idiots... But we are musicians... ( ) so when it comes to music... we... we can't... we can't admit an AI can "do" it...

[jancivil]

In terms of music, I don't have any real interest in it. Music, to be anything for me involves a deep set of memories in the human experience. The most rigorous abstractions of say Anton Webern were still informed by being a human. *Being* human is not going to be easily reduced to an algorithm. So, the algorithm that's cleverly built enough to analyze for tendencies is reductive to a surface that a machine lacking a history of experience can manage, so the music is going to be superficial by definition.

When the algorithm is showing signs of self-destruction tendencies from its particular damage, or different versions of a model show neuro-atypical manifestations, then we'll talk.

[woggle]

Luke Harrald, who models performers rather than music and then lets the performers improvise together

Thanks! I'll check into that, that's the kind of thing gets me right off.

Computers are still sophisticated idiots... But we are musicians... ( ) so when it comes to music... we... we can't... we can't admit an AI can "do" it...

ha ha - not sure what Harrald is up to now, that was a few years ago I heard that stuff but I thought his approach was very interesting and the music was also good.
re AI - why should we like AI music at all, if the AI is any good it will write music for other AIs, and they may have different taste than us anyway. Who are we to judge

[Functional]

These sort of tools are very interesting and magical... until you demystify them and the principles which they work with. All they they do is essentially one of the two: follow blindly rules of music theory to always "sound good" or have bit more open rules that allow for non-cohesion BUT they require some sort of midi input where they analyze patterns or otherwise input such as "does this sound good? Y/N" with lot of abstractions (which still conforms to our expectations and would likely converge on boring). As they are right now, they're great for muzak, but not so much for music. Maybe useful if you lack inspiration and really want something to work with, but all the stuff they come up with don't take all that much effort.

And all other aspects of music are still complete mysteries to them. Sound design? The nuances of an instrument? Mixing related (innovative) decisions? Nope. They won't necessarily ever be able to do it because maths don't really work well with qualitative concepts. The best maths can currently do is "fuzzy maths" and that only helps against binary way of thinking about logic, yet it's such a trivial problem for us as humans.

However, I do have some hopes up for these algorithms. I hope that they will be able to encapsulate music such as "bigroom house" and just spill it out on a button press. It shouldn't be impossible. To understand what I mean:



This stuff gets way too much attention. So maybe the algorithm could be used creatively and break the illusion. I find that maths are pretty good at disillusioning.

[Functional]

Luke Harrald, who models performers rather than music and then lets the performers improvise together

Thanks! I'll check into that, that's the kind of thing gets me right off.

Computers are still sophisticated idiots... But we are musicians... ( ) so when it comes to music... we... we can't... we can't admit an AI can "do" it...

ha ha - not sure what Harrald is up to now, that was a few years ago I heard that stuff but I thought his approach was very interesting and the music was also good.
re AI - why should we like AI music at all, if the AI is any good it will write music for other AIs, and they may have different taste than us anyway. Who are we to judge

AI is capable only to like what it is designed to like, this is how computers work generally. They can't just develop a sense of taste out of nowhere. And really, to be able to "like" anything at all, you'd first have to define what it means that it likes. Does it give a smiley for you through some interface? Because if so, that would be equivalent of someone telling you to just smile whenever you hear a specific tone. That's the way you will "like" music. Sounds kind of odd doesn't it? Okay, so it means something else. But what exactly? Well, that's really hard to tell. You use neuropsychology to establish some sort of a picture of how brains respond when you hear something that you like... but even then, the meaning is lost. When you say you like something and you're presented a picture of your brains with some highlighted areas, you won't think "hey, exactly how it feels like!"

And here we're only concerned with one person. Oh boy, when you try to establish a picture of group, the differential equations will truly be a sight to gaze at.

Why am I saying this? Well, because it sort of sounds comical for AI to develop music for itself or another AI. Why on earth would anyone really want to do this, unless we're talking about some weird transhumanist future which I won't even describe because it will sound so weird and out of touch (and more importantly, far beyond us).

[ghettosynth]

Luke Harrald, who models performers rather than music and then lets the performers improvise together

Thanks! I'll check into that, that's the kind of thing gets me right off.

Right. I'm also very much into this sort of thing, that is, if what we mean is using generative tools to explore artistic ideas that make "some sense" musically. I quote some sense here because I don't mean that in any strong sense. I mean that, given that you want to represent a particular style, that the algorithm makes sense with respect to that style.

I don't actually care to consume music that is purely generative and unending. I don't care about eno's latest work, for example. However, to use a generative idea to explore a particular interest that results in a piece or set of pieces that are then fixed is interesting to me, whether as a producer or consumer.

I still haven't found a (computer) language that is both expressive at the right level and has the right level of sophistication and simplicity.

[woggle]

[ However, to use a generative idea to explore a particular interest that results in a piece or set of pieces that are then fixed is interesting to me, whether as a producer or consumer.

that has been my approach with software - for example to ask "How important is note order in Satie's piano music" or "what is the role of rhythm in Philip Glass" and then use statistical techniques to explore that - the method of surrogate data in particular. I just use Matlab to write everything - I am familiar with it from my days in research and it has the advantage of a huge mathematical user base sharing lots of code. If I were younger I would use Python but I couldn't be bothered learning it for my small application

[harryupbabble]

If viewed as creator of music, sure, AIs are still primitive and can't compete with humans. But since music is subjective, maybe music made by AIs could fool a lot of people, if people don't know, beforehand, that the music was created by AI.

If viewed as tools for composers, AIs might not be totally useless.

AIs are like electric guitars, just tools? The creator of the electric guitar (Les Paul was it?) was probably shocked at how Jimi Hendrix used that tool. But who knows, for example, Chess AIs has already surpassed humans, maybe AIs that are designed to compose whatever music will surpass humans too, eventually. It depends on how good the creator(s) of the AI is/are? They get better? Maybe fiction is reality? Roy Batty was a better AI than Deckard (if Deckard is a replicant) because Roy Batty is next generation AI?

[woggle]

But since music is subjective, maybe music made by AIs could fool a lot of people, if people don't know, beforehand, that the music was created by AI.

I've been involved in little experiements where maybe a hundred people are asked which piece is humand which AI - can't remember the AI researcher but at the time was seen as the best and biggest thing around (ie was by a researcher in the US with all their great PR). It would write fake Mozart or fake Bach etc.
No-one confused the AI with the real composer. Not a single person. The main determinant was the lack of surprise or modulation of expectancy in the music from the AI. That's coz the AI rule extraction technique(s) by definition will produce a simplified model of the real works and the methods to try and insert novelty are not anywhere near as sophisticated as those used by people.

[harryupbabble]

But those experiments, the participants knew in advance that an AI is participating. Humans might have bias against AI. But maybe AI already was used on famous simple pop songs. And no one but the perpetrators know that the music is made by AIs?

Edit: Also, how long ago was those experiments?

[woggle]

But those experiments, the participants knew in advance that an AI is participating. Humans might have bias against AI. But maybe AI already was used on famous simple pop songs. And no one but the perpetrators know that the music is made by AIs?

true - but people did not know which pieces were done by the AI - the testing was "blind". It really wasn't that hard to tell for that sort of music. But you did need more than a couple of bars before you could tell.
As for pop music - I doubt any AI/bit of software will compose a believable Radiohead piece, but they could definitely make a fair go of a fragment of a Radiohead piece

(those experiments were maybe 8 years ago, but I don't think the field around learning systems for music has progressed that much since then - might have though - I still think there are fundamental limits to some of the methodologies of learning systems that are based around existing pieces of music - hence why I find the Luke Harrald approach more interesting. I definitely think there is a role in using computational systems to assist composers)

[Aleatoriac]

Right. I'm also very much into this sort of thing, that is, if what we mean is using generative tools to explore artistic ideas that make "some sense" musically. I quote some sense here because I don't mean that in any strong sense. I mean that, given that you want to represent a particular style, that the algorithm makes sense with respect to that style.

Yes! I see them as a sort of "non-judgemental musical sparring partner" - probably just another idea generator. The bit that I let jukedeck generate has inspired a bit of something, but I ended up removing what the AI produced completely in the end, and I don't like it.

But I can't tell you why.

Perhaps I'm anti-AI and I'm just too afraid to admit it? KILL THE MACHINES! Except no, they're not alive, so... oooh machine necrophilia... Song title!

I still haven't found a (computer) language that is both expressive at the right level and has the right level of sophistication and simplicity.

I wonder if that's why people make music. Other languages fail to convey what we're trying to express. Maybe?

[Aleatoriac]

In terms of music, I don't have any real interest in it. Music, to be anything for me involves a deep set of memories in the human experience. The most rigorous abstractions of say Anton Webern were still informed by being a human. *Being* human is not going to be easily reduced to an algorithm.

So... you're sayin there's a CHANCE?

So, the algorithm that's cleverly built enough to analyze for tendencies is reductive to a surface that a machine lacking a history of experience can manage, so the music is going to be superficial by definition.

Agreed, but I must ask, what exactly is it that Humans are doing that is so different? Merely sophistication? I really don't know. I *feel* something different. But, it occurs to me, my emotions are like a fuzzy *.jpg of whatever it is my sensory organs are showing me about my environment.

When the algorithm is showing signs of self-destruction tendencies from its particular damage, or different versions of a model show neuro-atypical manifestations, then we'll talk.

Wait...