Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

MuseNet

openai.com

Post reply on HN
Print view View original

111–120 of 189 posts

Re: MuseNet

#111

Post by erichocean »

Permit wrote: post #66
TheOtherHobbes wrote: post #62

Earlier quoted context omitted.

The results are meaningless. This isn't chess or go, where you either win or you don't. Chopin composed for a reason, and it wasn't just an excuse to throw a lot of notes at the page. It might be possible for AI to work at that level someday, but it's not just a technical problem, and you won't be able to solve it by throwing a corpus of compositions at it. Aside from that, this still sounds like aimless noodling. It…

>This isn't chess or go, where you either win or you don't. Chopin composed for a reason, and it wasn't just an excuse to throw a lot of notes at the page. I disagree based on your following quote: > Aside from that, this still sounds like aimless noodling. It's far more polished noodling with some awareness of genre cliches, but it's still essentially aimless - and so meaningless. This sort of problem can easily be…

> More concretely: would this music be able to convince you that a gifted human composer created it?

Would a computer? If it could, then in theory it could use that as an optimization goal...

Re: MuseNet

#112

Post by CuriouslyC »

leesec wrote: post #104
CuriouslyC wrote: post #81

I'm very into both music composition/production and ML, so this is neat. That being said, I think the "computer generated music" path is probably the wrong approach in the short term. Music is really complex and leans heavily on both emotions and creativity, which aren't even on the radar for AI. Being able to dynamically remix and modulate existing music is still really cool though. I would kill for a VST tool that…

>Music is really complex and leans heavily on both emotions and creativity, which aren't even on the radar for AI. I definitely think creativity is on the radar for AI, see: AlphaGo. Everything we think is based on emotions is ultimately learnable.

I don't think alpha go is a good example here. It play with strange brilliance, but I wouldn't call that creativity. If you care to elaborate on why you said this I'm curious to hear your rational.

To me, creativity is really about generation of "aesthetic novelty" which is hard to get from a ML algorithm that is trying to approximate patterns in training data. Eventually, there will be models trained on a wide variety of art, music, stories, etc that can recognize aesthetic and structural isomorphism between mediums (say between a grizzly picture and death metal), then we'll lose our competitive advantage. I don't think we're nearly so close to that as the singularity types would have us believe though.

Re: MuseNet

#113

Post by Toine »

CuriouslyC wrote: post #83
codelitt wrote: post #64

While this may not be perfect composition, this is a surreal (and almost sad) moment in my life to hear music that is passable created by a computer under it's own volition. I work at a company that works with a lot of machine learning so I generally understand its limitations and haven't ever been an alarmist. That being said, I generally always thought of it being applied to automate work. For some reason I had alw…

Music created solely by machines will probably remain derivative and simplistic for a long time. I expect the biggest result of this research in the near term is that we'll be able to create tools that lower the skill and time required to create good music, kind of like an audio version of templates/autocomplete/spellcheck.

Agreed completely, don't know why you're getting downvoted...

Re: MuseNet

#114

Post by bufferoverflow »

elihu wrote: post #106

I expect the next major copyright law legal battle will be around the question of whether content generated by some machine learning approach that was trained on copyrighted data is a derivative work or not.

Proving that a certain piece was generated from a certain dataset will be a huge obstacle.

Heck, proving a piece was generated would be hard.

Re: MuseNet

#115

Post by trexmix »

erichocean wrote: post #111
Permit wrote: post #66

Earlier quoted context omitted.

>This isn't chess or go, where you either win or you don't. Chopin composed for a reason, and it wasn't just an excuse to throw a lot of notes at the page. I disagree based on your following quote: > Aside from that, this still sounds like aimless noodling. It's far more polished noodling with some awareness of genre cliches, but it's still essentially aimless - and so meaningless. This sort of problem can easily be…

> More concretely: would this music be able to convince you that a gifted human composer created it? Would a computer? If it could, then in theory it could use that as an optimization goal...

Yes, that is what he is saying...

Re: MuseNet

#116

Post by TaupeRanger »

codelitt wrote: post #64

While this may not be perfect composition, this is a surreal (and almost sad) moment in my life to hear music that is passable created by a computer under it's own volition. I work at a company that works with a lot of machine learning so I generally understand its limitations and haven't ever been an alarmist. That being said, I generally always thought of it being applied to automate work. For some reason I had alw…

It's not passable...it's pretty obviously algorithmic.

The program does not have volition.

Why would you think that using statistics to generate a model of a piece of art (which is just data in the case of MIDI and pixels) would be "off-limits"? People have been doing this for decades.

No one knows the answer to your last two questions, but there is no indication that this program is leading there.

Re: MuseNet

#117

Post by uptownfunk »

What speaks to me here is this is one of the closest things we have to creativity in AI and while AGI is a ways away, this is a key step, even though it may not seem like some hard tech we can put to use in industry immediately.

I think it’s a subtle distinction, but imagine a model where we can throw in thousands of math proofs then give the model some initial assumptions and just let it run wild. I think getting a neural net to model the creative spark / the ingenuity is what has been missing.

A part of me doesn’t want to believe that it is possible, but a part of me is genuinely curious as to the consequences.

An exciting time to be alive, folks.

Re: MuseNet

#118

Post by TaupeRanger »

leesec wrote: post #104
CuriouslyC wrote: post #81

I'm very into both music composition/production and ML, so this is neat. That being said, I think the "computer generated music" path is probably the wrong approach in the short term. Music is really complex and leans heavily on both emotions and creativity, which aren't even on the radar for AI. Being able to dynamically remix and modulate existing music is still really cool though. I would kill for a VST tool that…

>Music is really complex and leans heavily on both emotions and creativity, which aren't even on the radar for AI. I definitely think creativity is on the radar for AI, see: AlphaGo. Everything we think is based on emotions is ultimately learnable.

Go is a game with a strict ruleset. The comparison is not valid.

Re: MuseNet

#119

Post by TaupeRanger »

dpcan wrote: post #109

As a web developer, all these "do it yourself" website systems make me sick. The end results look amateur and messy most of the time, and it's an insult to the industry to give people with no visual or code-based skills the tools to create sub-par websites. I'll bet talented musicians and real composers feel the same about MuseNet. The music probably makes them cringe to their core. Then there is the general public..…

Quite true. I've found this to be the case with almost every "composition AI" system ever created, with the exception of those that aid human composers, such as those designed by David Cope.

Re: MuseNet

#120

Post by Ericson2314 »

As a sometimes composer, the thing that makes me most sad about this stuff is it's all computer or no computer. Deep learning with symbolic stuff is still in it's infancy, so I don't expect to really mix this with hand-composed stuff anytime soon. As far as I understand it as a layman, the way the input is compressed for learning is also how it's fed the sample, so there's no way to "pay more attention" to the piece your augmenting than the pieces in the training data.
Post reply on HN