"AI is going to have a profound impact on your business and your art"

[Tricky-Loops]

The more computers make artificial songs, the more people will love handcrafted songs with real instruments again. There's always a contrary movement... When everyone made synthesizer-based disco music in the 70ies, more and more people hated it and rather preferred rock music (again).

[arkmabat]

Things that sort of embrace this idea:

Jamstix and other drum machines
Melodo
Vocaloid
Arpeggiators
Virtual Guitarist
Vielklang

Microsoft's working on MySong, a Jam Machine that writes music to vocal input: http://research.microsoft.com/en-us/um/ ... an/mysong/

Actually, a beta came out: research.microsoft.com/en-us/um/redmond/projects/songsmith/download.html

[arkmabat]

I wonder what Band In A Box is like now? http://www.pgmusic.com/

[ras.s]

I guess you have never toyed with Jamstix?


http://www.rayzoon.com/jamstix3_vid.html

Well it's been some time now since I tried it and I didn't think of it being useful back then. Looking at the videos now and listening to the demos of the various additional packs, yea, sure it's really cool. I'd enjoy having that as a rehearsal/improvisation partner, but I think I would feel dirty using it in songs that I'd call my own (I guess that was what didn't convince me to buy it when I tried it). Plus I'd have to teach it to drum the way I want it to, which might take quite some time (though it might very well be rewarding).

But yea, it's really a cool looking thing. And sounding, obviously. A good example of a 'intelligent' software.

[jancivil]

jancivil,

I mean that part "whether a recording is "real" or made with a computer" more about the use of samplers, etc. It's relatively easy/common to fool a person to think a drum part is actually recorded in a studio, or that the string section is actually made of people's blood and sweat. It no longer takes years of playing an instrument to make a convincing recording with that instrument. So that "fooling" is already possible. Whatever it's going to be possible to use an algorithm to compose those parts convincingly (or in a manner that atleast produces interesting results), I don't think it's impossible, and some of it is already done, I think.

I think you have confounded sound with performance now. I saw that you were talking about sound, but I don't think, concrete example, Spitfire making really rough-sounding 'shorts' through itself convinces, it's the faking of a performance that does, and I do not think this is a trivial matter. My whole trip is realizing a score obviating hiring musicians and unless you're looking for it, you do not think about the process, it's perfectly real in effect.
I'm sure that the convincing people have years of playing instruments. I'm sure that samples that are recordings of say a master drummer beat the hell out of recordings of just anyone coming in and bashing about.
Additionally you have posed 'convincingly' but then 'interestingly'. I think a person falling down on a set of drums could be 'interesting'.

Another thing about AI as a musical vehicle is that it can be used to create music that isn't based on currently existing genres or musical traditions (or the history and personality of the composer). Take xoxos stuff for an example; he has made drum sequencers that produce drum parts that are independent of genres, they sound like their own personality and as such, are very creative. Whatever those results can be merged with others results is still something that is up to the composer.

Well, the statement I made was that AI as a composer depends on a number of things that haven't happened and are seriously more complex than advocates are dealing with at all, and you apparently agree as you conclude 'it's up to the composer'. I'm actually fairly aware of where algorithm is used to replace manual workflow.

[Sendy]

My music, my flaws, my timetable, my racket, my therapy.

I ain't gonna be farming it out to the cloud.

[jancivil]

I guess you have never toyed with Jamstix?

I think I would feel dirty using it in songs that I'd call my own (I guess that was what didn't convince me to buy it when I tried it). Plus I'd have to teach it to drum the way I want it to, which might take quite some time (though it might very well be rewarding).

I think the latter exercise could well be rewarding. But there you have it, it isn't AI, it's algorithm.

[Sendy]

Also, I still can't believe that people buy the hype from AI companies, they have shit to sell, so when they tell us "OMG AI YOU NEED IT BRO!" I personally take it with a pinch of salt. The Kasparov/Deep Blue match is a commonly cited example of machines besting humans yet anyone who actually bothers to research that incident will find all the evidence they need to see how flawed AI is as a concept and how skewed and blown-up the whole thing is in the media machine.

That's not to say AI is worthless, it's getting better all the time and does some very clever stuff. I still think there's more contextual understanding and "adaptive smartness" in a slug than anything we've ever made, but we may just get there in the end, after at least 1000 years. As it is I don't think we have the computational power, we need a paradigm shift and something like quantum computing to be fully developed, blowing the lid off of notions of what computation can and can't do.

[Frantz]

Current AI is pretty dumb and so music composed by it tends to be dull. Once AI becomes truly intelligent, maybe in 100-200 years, there is no reason it couldn't create art. I doubt it would be interested in creating music or paintings as we know them since they are geared towards human input devices - ears and eyes. I imagine AI art would exist as data structures in the digital domain and would be complex and incomprehensible to humans.

[ras.s]

jancivil, I'm not really confounding performance with sound. I specifically said that the sound part is there already and then I wondered about whatever the performance part is there yet or not.

We'd have to draw the line somewhere though -- is telling the computer what to do and then the computer deciding what's the best way to do that, is that artificial intelligence or just regular algorithm thing? To fool that performance part is really essential, but is it AI if we tell it what kind of articulations we'd like to hear. That Jamstix really seems to make articulations, rudiments and such, quite a trivial thing. Are there such programs for other instruments as well? Does BIAB know how to do specific Kontakt stuff, for instance?

I have no reason to believe that it really is hundreds of years away or that we need specific kind of hardware to reach that level, it's more about someone investing the time to program that.


What would be the equivalent of the Turing test in the world of music?

[Sendy]

What would be the equivalent of the Turing test in the world of music?

Being able to compose music and explain the thought process behind it, what inspired it, what was difficult, what was easy, etc. Then to be able to take constructive criticism and choose to accept or reject it based on an overarching artistic vision.

See, just being able to output a convincing piece in a certain style isn't enough. A Turing test requires feedback and recursive input/output cycles, in my estimation. We want to be sure there's a point of view in the machine, and not just a sophisticated causal system. It has to transcend it's station. It must also probably be terrified of being dismantled and acutely aware that this is inevitable.

[RunBeerRun]

Jamstix is very cool-I should check it out again, reminds me of the similarities like with Xoxos' midi sequencers. It won't affect my "art" though, as I couldn't do anything drawing wise, not even write legibly .

[jancivil]

jancivil, I'm not really confounding performance with sound. I specifically said that the sound part is there already and then I wondered about whatever the performance part is there yet or not.

I was referring to

It no longer takes years of playing an instrument to make a convincing recording with that instrument. So that "fooling" is already possible.

I'm confused, then. To make it more concrete, what is 'a convincing recording' here? A simple sample? That was availble with a mellotron many years ago. To some extent it's convincing. Who is convinced? If you make a shite performance with the best samples in the world, who is 'fooled'? If you're making a recording of a gesture on an instrument, I am sure the recording of a musician-for-life is more convincing. A buzz roll on a snare, even. SO:

We'd have to draw the line somewhere though -- is telling the computer what to do and then the computer deciding what's the best way to do that, is that artificial intelligence or just regular algorithm thing? To fool that performance part is really essential, but is it AI if we tell it what kind of articulations we'd like to hear.

I'd have to say no.

[jancivil]

Sendy has defined sentience there. I think that's a start.
So, given just how much more a human brain has in terms of computing power, I do tend to believe that what we have today, in terms of hardware is seriously insufficient for the task of building a creative musician.

[jancivil]

Maybe you could build an avicii on a Pentium 4 machine with 500MB RAM.