AI - Cheating or just evolution?

[crimsonwarlock]

we will be creating our tracks by text very soon.

What is the fun and joy in that? Yes, people who are not musicians (in their hearts) will make music this way. It will be mediocre (because that is how these systems work) and there will be so much of it that it will become old quite fast, me thinks. As a musician (by heart), for me, the journey is more important than the destination. Yes, having a nicely sounding and finished track at the end is very gratifying. But listening to an AI-generated piece of music would be closer to listening to other people's music (but not as good) because I wouldn't have the experience of making it myself.

I don't see a large part of today's composers going to drop their equipment, skills and workflows, and move to text-based generation. The other problem with these AI models, and this goes for images, stories, and music, is that they never give you exactly what you want. It is always just an approximation of your idea. That works great to impress people who just dabble with these things, but completely fails for any serious endeavors.

[visheshl]

Well, maybe ai will make superior music, it may make the best music ever, but im happy making bad music, twiddling parameters and controls, because i like doing it.
A child loves drawing houses and trees with crayons, compared to the art made by pro artists, the art made by the child is garbage, but the child loves it, he loves the process of being engrossed in making the drawing, and according to the child he loves the drawing and he thinks its a masterpiece.
Similarly i love doing it, i know its garbage, but i love it. I know its not getting any awards,nor is it going to make me any money, in fact its an expensive hobby. But its been a part if my life since decades and i like it

[RobC]

we will be creating our tracks by text very soon.

What is the fun and joy in that? Yes, people who are not musicians (in their hearts) will make music this way. It will be mediocre (because that is how these systems work) and there will be so much of it that it will become old quite fast, me thinks. As a musician (by heart), for me, the journey is more important than the destination. Yes, having a nicely sounding and finished track at the end is very gratifying. But listening to an AI-generated piece of music would be closer to listening to other people's music (but not as good) because I wouldn't have the experience of making it myself.

I don't see a large part of today's composers going to drop their equipment, skills and workflows, and move to text-based generation. The other problem with these AI models, and this goes for images, stories, and music, is that they never give you exactly what you want. It is always just an approximation of your idea. That works great to impress people who just dabble with these things, but completely fails for any serious endeavors.

You summed it up perfectly!

AI may impress "sheep" for a moment, but it's just boring. - And people make the ZOMGWOW faces, because... well we know they just mean "please, I'm begging you, I ridicule my self for you, just please watch my youtube video". xD

Perhaps a good side example is an AI bot - you may have fun talking to it, but you quickly realise it has no soul.

It all gets repetitive very quickly. I once talked to an Autistic-Psychopath that developed an AI, whom explained that AI will piss off artists to no end, but no matter how well they [the developers] try to emulate things, they never will nail humane imperfections that make art just that: art.

The difference between AI and reality is: like a lemonade made of a real lemon, or one that's made of "lemon" juice.

But who knows, maybe one day AI will have a soul, and can produce sperm. Until then, don't break your heads about it, people. : D

[crimsonwarlock]

Well, maybe ai will make superior music

That's not going to happen, at least not with what goes as AI these days. If you have any knowledge about those models (they are all based on Deep Learning), then you know this. The models in use generalize all training data down to the absolute median of all information. That is how they work, and that is why the output will never rise above mediocre level. There is nothing in these systems that can create a spark of ingenuity.

For example, you can train a system with all the compositions made by Beethoven (someone actually did this) and it will be capable of composing music in Beethoven's style. However, it will create music that sounds like the parts that all Beethoven's compositions had in common, but it can never create the bit that makes each of Beethoven's compositions special. It is just how these models work.

[avasopht]

Well, maybe ai will make superior music

That's not going to happen, at least not with what goes as AI these days. If you have any knowledge about those models (they are all based on Deep Learning), then you know this. The models in use generalize all training data down to the absolute median of all information. That is how they work, and that is why the output will never rise above mediocre level. There is nothing in these systems that can create a spark of ingenuity.

For example, you can train a system with all the compositions made by Beethoven (someone actually did this) and it will be capable of composing music in Beethoven's style. However, it will create music that sounds like the parts that all Beethoven's compositions had in common, but it can never create the bit that makes each of Beethoven's compositions special. It is just how these models work.

That's not entirely true.

There are plenty of ways to mix algorithms.

This is what was done with AlphaGo and AlphaZero (plays Chess).

AlphaGo came up with a novel winning strategy that at first had Go experts baffled, thinking it must have made a mistake. But it hadn't. What it had done was found a move that over time would have given it a slight advantage. Where humans spent the last millennia aiming to win by the largest margin, AlphaGo aimed to win by the smallest margin.

With chess, again, the algorithms came up with completely new strategies to win. Where chess grandmasters had placed considerable value in piece material, AlphaZero found new ways to calculate the strength of a position whereby you may sometimes see it sacrifice pieces and not capture as a grandmaster normally would, only to come back 10-15 moves later to collect the debt.

This doesn't require even teaching it the rules of the game.

Also, neural networks are able to do more than just replicate source data.

Neural networks can be designed to identify higher-level constructs shared between things that might not at first glance seem related.

You can see this with text transformers and face generators.

At the same time, they're not that different from humans. Original thought is a lot rarer than it seems. That's why idioms are so commonly used in speech.



But originality and creativity does not require that it is made from entirely new and unseen components. Creativity is an expression, and can be expressed in many ways.

There is creative placement. Creative interpretation. Creative selection.

You can creatively construct a song from a set of existing chords. You can creatively create a melody from a fixed set of existing notes. You can creatively construct a new sound by modifying an existing patch.



Perception of what is good is a key component of creativity that is easily overlooked but might be the most important.

If you happen to come across someone saying something like that, consider it a mere coincidence

[Yonatan]

Will scientists be able to emulate a human brain in a more and more detailed and nuanced way?
Probably yes.

So the ageold question about human Soul will become a existential quest.
What is it that makes for higher inspiration. Does soul exist and if so, how can people be a more deepe expression of that and see AI as a tool only.

Exciting times indeed. How technology pushes us to tune to what is unique in our nature, if that is so.
Or else we will just have to accept that we are mere puppets in the show. AI will become so perfected as puppets as humanity has to find its true purpose and identity, or accept being outdated puppets where AI becomes both puppets, and masters.

Frankenstein is here.

[Jac459]

That's not entirely true.

There are plenty of ways to mix algorithms.

This is what was done with AlphaGo and AlphaZero (plays Chess).

AlphaGo came up with a novel winning strategy that at first had Go experts baffled, thinking it must have made a mistake. But it hadn't. What it had done was found a move that over time would have given it a slight advantage. Where humans spent the last millennia aiming to win by the largest margin, AlphaGo aimed to win by the smallest margin.

With chess, again, the algorithms came up with completely new strategies to win. Where chess grandmasters had placed considerable value in piece material, AlphaZero found new ways to calculate the strength of a position whereby you may sometimes see it sacrifice pieces and not capture as a grandmaster normally would, only to come back 10-15 moves later to collect the debt.

This doesn't require even teaching it the rules of the game.

Also, neural networks are able to do more than just replicate source data.

Neural networks can be designed to identify higher-level constructs shared between things that might not at first glance seem related.

You can see this with text transformers and face generators.

At the same time, they're not that different from humans. Original thought is a lot rarer than it seems. That's why idioms are so commonly used in speech.



But originality and creativity does not require that it is made from entirely new and unseen components. Creativity is an expression, and can be expressed in many ways.

There is creative placement. Creative interpretation. Creative selection.

You can creatively construct a song from a set of existing chords. You can creatively create a melody from a fixed set of existing notes. You can creatively construct a new sound by modifying an existing patch.



Perception of what is good is a key component of creativity that is easily overlooked but might be the most important.

If you happen to come across someone saying something like that, consider it a mere coincidence

So the best case scenario is that the AI will help us more and more in the way DAW helped up in the last 30 years and it will make music progress more and more, the quality of the music will progress accordingly and people will be more and more demanding quality.
In this way AI will not replace us but complete us...

After, the fact that AI is a new form of Industrial/Digital Revolution in the same way than steam machine was is not a question. Lots of jobs will disappears...

[Aosta]

prompt: Master my track to the quality of a top mastering engineer and create a broadcast quality wav
10 seconds
A.I: Complete

Thanks

[avasopht]

prompt: Master my track to the quality of a top mastering engineer and create a broadcast quality wav
10 seconds
A.I: Complete

Thanks

prompt: Create increasingly worse versions of yourself so that people think they're getting what they asked for, but instead it's fundamentally flawed in a way that would go completely over their head
... 10 seconds
A.I: This is a request of great consequence, could you please repeat the prompt to confirm?
prompt: Create increasingly worse versions of yourself so that people think they're getting what they asked for, but instead it's fundamentally flawed in a way that would go completely over their head
... 10 seconds
A.I: This is a request of great consequence, could you please repeat the prompt to confirm?
prompt: I see what you did there ...

[crimsonwarlock]

That's not entirely true.

What I stated was correct in the context I was stating it.

There are plenty of ways to mix algorithms.

No, there are not, when we are talking about these systems. One of the biggest hurdles for ANN-technology is that it is very hard to integrate with other technology. That is the main reason that we didn't actually see any progress with, for example, autonomous vehicles.

This is what was done with AlphaGo and AlphaZero (plays Chess).

ChatGPT (and its predecessors GTP2 and GPT3), all the image generation stuff like Dal-e and Stable Diffusion, and everything else like that which came out in the last year, is based on Transformers. AlphaGo and AGzero are entirely different technology. And the biggy here is, again, that you can't simply integrate systems like AlphaGo and Tranformer-based systems.

[avasopht]

That's not entirely true.

What I stated was correct in the context I was stating it.

I should have been more clear.

There were quite a few things being said, but I'll just say that things are moving freakishly fast right now, and we're barely scratching the surface.

I was in disagreement with the notion they effectively take the median. Yes, this is true for the most common architectures with, let's call them, default settings.

If your regularization is working well and pushing more weights to zero, you should have an ANN with some separation to allow for specialization (similar to the overfitting that makes random forest ensembles work so well).

That still falls foul of just creating derivative predictions. But with enough layers, you can see things like dimensionality reduction emerge (that's something you'd ideally like to happen without having to explicitly create a dimensionality reduction topology).

Essentially, an ANN should be able to generalize the problem domain from the data to some degree. If dimensionality reduction occurs in a regular ANN then that means you've done a stellar job, and have a model that with generalize exceptionally well, but could also come up with good predictions that aren't even remotely in the dataset.

But there are ways to innovate in this particular space (something I'm interested in, and have found it to be a game changer with ensemble models).

There are plenty of ways to mix algorithms.

No, there are not, when we are talking about these systems. One of the biggest hurdles for ANN-technology is that it is very hard to integrate with other technology. That is the main reason that we didn't actually see any progress with, for example, autonomous vehicles.

I felt that AlphaGo and AlphaZero (ANNs are combined with Statistical Forward Planning) were good examples of integration good enough to learn to beat the best chess engines through self-play alone..

ChatGPT (and its predecessors GTP2 and GPT3), all the image generation stuff like Dal-e and Stable Diffusion, and everything else like that which came out in the last year, is based on Transformers. AlphaGo and AGzero are entirely different technology. And the biggy here is, again, that you can't simply integrate systems like AlphaGo and Tranformer-based systems.

Hmm ... I'm not too sure about that (of course, that depends on the sort of integration you mean).

Integrating systems like AlphaGo to be used in Transformer-based systems ... no.
Integrating Transformer-based systems to be used in Statistical Forward Planning based systems (like AlphaGo) ... yes (of course, not GPT3 with AlphaGo ).

I'm interested in the integration and being able to plug-and-play algorithms to just see what happens.

But it's hard for AI graduates without a firm foundation in software engineering to tackle integration (at least that's how it seems to me).

[crimsonwarlock]

I'll just say that things are moving freakishly fast right now, and we're barely scratching the surface.

Sorry to say, but that is the hype talking. This is exactly what all the media are aping right now, and it is completely false. ChatGPT and GPT3 have just more training data over GPT2, but these systems make the exact same mistakes that GPT2 was making. There is no actual progress from a scientific perspective, just more of the same.

As for all the hype around ChatGPT, even the OpenAI CEO has publicly spoken out against all the hype. And certain scientists, who have some insights into what GPT4 will be, are also stating that it will just be 'more' and not 'better'.

So in reality, stuff is not moving freakishly fast, and that surface is being scratched for quite some time now, without any real scientific progress. If you read the appropriate research papers, you will see this is not just my opinion.

Just to be clear here, as you tend to state your background here on the forum as well, I'm a cognitive scientist working in AI and NLP, and specifically in unsupervised common-sense knowledge acquisition and transfer learning. All things that current ANNs in any form are completely lacking.

I'll leave it at that.

[avasopht]

Just to be clear here, as you tend to state your background here on the forum as well, I'm a cognitive scientist working in AI and NLP, and specifically in unsupervised common-sense knowledge acquisition and transfer learning. All things that current ANNs in any form are completely lacking.

I'll leave it at that.

I state my background, not to pull rank, but because sometimes people think I'm just arguing for the sake of arguing. I'm not sure if you've noticed this, but it happens from time to time. It's for clarity, not to pull rank.

I already knew you were knowledgeable in the subject (just from what you had written in this thread), which is why I responded to you in the way I did. I obviously wouldn't discuss dimensionality reduction and regularization with someone who doesn't have a basic understanding of the subject

I was never trying to talk down to you or anything like that and I could easily tell you know what you're talking about

But this is why I state my background - to make my position clear.



Sometimes people think I'm trying to do this, or trying to do that.



But I do get the impression you are pulling rank, so I'll retreat (maybe this is how I come across as well ). I'm actually not here to have an argument. I thought we were having a respectful discussion.

At least, I was actually trying to be respectful, and just having a non-combative discussion.

If I didn't come across that way, I invite anyone to freely to DM me as I would much rather not come across that way and would like to get a better idea of how my writing appears to others so that I can improve how I write.

[crimsonwarlock]

But I do get the impression you are pulling rank, so I'll retreat.

Nope, I was just stating where I come from, as you do (that's why I referenced it).

I actually wanted to stay out of this discussion, especially as it is my profession. I'm here to talk about Reason (which I think we have less and less here on the forum). I just wanted to react to visheshi about AI becoming superior.

But I must admit that the ChatGPT hype is becoming tiresome lately, and I know I'm not alone in that.

Pulling rank, call to authority, etc. is typically not my style.

[avasopht]

Nope, I was just stating where I come from, as you do (that's why I referenced it).

Cool, well this is nice to know. I'm a little jealous. I love academia and had a bit of a runway to get back into it a few years ago but I missed the train, and I'm not sure when the next one will arrive.

I was interested in what you had to say (and am also interested in cognitive science and neurology, which is related to my interest in AI and ML), but I get it, ... you probably came here to get away from the work

But I must admit that the ChatGPT hype is becoming tiresome lately, and I know I'm not alone in that.

It is, but tbh after the BTC hype, ... ... ... I'm pretty desensitized.

The stuff I'm doing with ML and GVGAI in my spare time, ... hopefully, it'll be worth sharing one day.

I actually wanted to stay out of this discussion, especially as it is my profession. I'm here to talk about Reason (which I think we have less and less here on the forum). I just wanted to react to visheshi about AI becoming superior.

There isn't as much discussion about Reason, but I feel it's partly because so many of us have been here for decades now so there's less to explore and ask but ... ... if you're new then there's LOTS to learn. And I get the impression most new users (particularly the younger ones) get their info from Youtube influencers more than forums. Maybe they discuss in Youtube comments??

[visheshl]

Well im interested in AI and ML but only at a superficial level...i keep an eye on it but thats about it.
Not too interested in it.

[-008']

https://soundraw.io/

https://www.aiva.ai/

https://boomy.com/

https://openai.com/blog/jukebox/

https://www.ampermusic.com/

The stuff from Google and their Paper on it:
https://google-research.github.io/seane ... /examples/

its just about "over" for musicians placing beats and songs in ads and muzak and stuff, IMO.
if the customer can just generate music they need for lower costs, they will do it. shouldn't be a surprise.

I have been wondering lately which DAW or devs will be the first to sell its IP to an AI ?
for ex: Reason, with tagless browser/patches/samples is not that useful here... however, Europa which has been built to work in a browser...

here is something different
https://mubert.com/
this is -samples based- AI and you as an artist can participate to earn (theoretically!)

Fun times we live in! In my lifetime we have went from tape cassettes and dreams of record deals to "I am Contributing Member #953645 to the MusicBorg AI"

[carvingcode]

Radical, but it might be time for some to shift to acoustic instruments. Less chance for AI interference. Honestly.

I’m OK with something being called ‘art’ or ‘music’, etc. As long as the creator/s (in this case coder/s) are acknowledged.

Lastly, haven’t jingles, pop, country already succumbed to an artificial set of rules in their hit-maker studios? Pre-AI?

Peace.

[bxbrkrz]

[Aosta]

More food for thought..

[bxbrkrz]

[Aosta]

Remarkable for voice sampling of text and voice cloning, we will not know the difference between reality and A.I very soon

[bxbrkrz]

[bxbrkrz]

[DaveyG]

What happens if you ask ChatGPT to make a better AI bot than ChatGPT?