If AI replaces musicians, does the entire plugin industry die with them?

[ghettosynth]

One of these dayz, when I am up to it, I will post some Van Halen type of stuff for ya, BONES...

You will luv it!!!

Post it to the cafe, it's so rare that "KVR cafe" and "I love it" appear in the same sentence.

[TechHaus]

Semi-related: came across this news segment from Singapore.


Local music studios see budget cuts from clients amid faster, cheaper AI alternatives

Local music firms in Singapore have seen budgets for client projects cut by up to 60% over the past three years. This is due to music and voice-generating AI, which is faster and cheaper than human labour. However, they say AI cannot replace client servicing, artist and strategic development, as well as branding and community development. Studios are calling for better guidelines to protect human creators in the industry.

[zerocrossing]

I can’t say for certain if it’s applicable to the plugin industry, but the company my wife works for replaced most of its low level customer service with AI agents by decree of the CEO, and now they’re having major problems with the software spewing garbage to angry customers.

Another Sunos user was recently complaining that he was canceling his subscription, because no matter what the prompt he came up with, he was basically getting sight variations of the same song, over and over.

I tried OpenArt AI, and it was a disaster. Plenty of “influencer” videos showing great results, yet it spit out a bunch of jerky garbage for me.

I’m starting to get the feeling that my initial impression that this is largely a shell game was correct.

[ghettosynth]

Semi-related: came across this news segment from Singapore.


Local music studios see budget cuts from clients amid faster, cheaper AI alternatives

Local music firms in Singapore have seen budgets for client projects cut by up to 60% over the past three years. This is due to music and voice-generating AI, which is faster and cheaper than human labour. However, they say AI cannot replace client servicing, artist and strategic development, as well as branding and community development. Studios are calling for better guidelines to protect human creators in the industry.

"We had to cut freelance hiring by over a third", (emphasis mine) So, your business, which outsources labor rather than employ people, is not able to maintain your same profits now that the people they used to hire temporarily as a cost cutting measure over hiring them as employees is being threatened by cost cutting measures brought on by technology? Is that right? Moreover, often those hires themselves may not use studio rental because democratization of recording technology means that they could squeeze studios out of the profits that they used to earn? What is that in that video? Is that "plugin technology" that squeezed hardware manufacturers out of business by creating "simulacrums" of the hardware through advances in technology, which, further allows individuals to not pay studios their "profits"?

Just making sure that I understand the dynamics here.

I believe the saying goes something like "I didn't think that the leopards would eat MY face."

[Grizzellda]

One of these dayz, when I am up to it, I will post some Van Halen type of stuff for ya, BONES...

You will luv it!!!

Post it to the cafe, it's so rare that "KVR cafe" and "I love it" appear in the same sentence.

I will look at all the options at the right time. Not quite ready yet!

I do wanna post some stuff on KVR. Actually show myself. Would jancivil do that? Gotta wunder!

[BONES]

I doubt she(?) sees us as worthy of her talents. But she(?) has stuff on YouTube if you're interested. It was nothing like what I had expected, given the way she(?) carries on.

LLMs prompting is copying. Artists steal ideas and especially improve them.

That is completely wrong, LLMs will never give you exactly anything. They don't work the same way we do, they don't hear a song as a combination of rhythms and melodies, it's all just a sequence of individual samples to them, at whatever sample rate you/they are working at. They generate a song whole, sample by sample, not by creating separate parts and them merging them together. They are designed to do similar things, not identically the same thing. It's as much in their make-up as it is in the most creative human minds. It's what makes them such useful tools in the right hands.

We generated more than 400 songs to get 10 songs for an album and no two are alike. You can say that some of them are "in the style of" another artist but not one of them uses anything you would say was taken from somewhere else. If it was copying anything, we'd have heard it but there wasn't anything even vaguely recognisably "copied".

It's most likely they are made that way to avoid billions of dollars worth of law suits but it works to our advantage because it allows us to make amazing things that are all our own. And, trust me, you could spend a thousand hours trying to get an AI to generate something the same as anything we've generated and never even get close. That's just how it works.

I tried OpenArt AI, and it was a disaster. Plenty of “influencer” videos showing great results, yet it spit out a bunch of jerky garbage for me. I’m starting to get the feeling that my initial impression that this is largely a shell game was correct.

Whereas if I was getting shitty results and I could see that others were getting good results, I might question my ability as a prompter.

"We had to cut freelance hiring by over a third", (emphasis mine) So, your business, which outsources labor rather than employ people, is not able to maintain your same profits now that the people they used to hire temporarily as a cost cutting measure over hiring them as employees is being threatened by cost cutting measures brought on by technology? Is that right?

Not really. It's project-based work, you never know from month to month how many people you are going to need. When you have a big project, you might need 200 artists for a year but once that big job is out the door, you might struggle to find ork for 20 artists until the next big project comes along. Until you get to the size/reputation of Weta or ILM, you need to have a highly flexible workforce.

It's also a bit of a misconception that freelancers are cheaper than employees. In my job before last, I was earning about $280 a day as a full-time employee while freelancers were being paid $500+ a day. There is no way that my sick leave and paid holidays made up that shortfall. I've spent most of the last 25 years bouncing between freelance and full-time - full-time for a bit of certainty and freelance to make lots of money when there was a lot of work around.

But it's all down to how you choose to see it. AI might enable work that would have gone off-shore, say to China or Vietnam, to remain in the hands of local companies. i.e. AI can make local production more competitive in the global marketplace.

[ksandvik]

LLMs copy patterns here and there and rearrange them, basically a multi-dimentional Markov chain. There's no creativity whatsoever in this mechanical process.

The creativity happened with the patterns beforehand that LLMs were (illegally) trained upon.

[BONES]

And what, a person strumming a guitar or playing a few chords is somehow f**king magical? Get over yourself, there is nothing special about the human creative process. Our brains are pattern recognition machines that regurgitate the things we hear around us imperfectly. It's every bit as random, every bit as "copied" as what LLMs do and you want to think it somehow makes you f**king special? Spare me days! Do you not see how pathetic that is? LLMs have already achieved far greater success than you ever will so if they're as incapable as you're suggesting, what does that make you? (Cue excuses.)

[zerocrossing]

LLMs copy patterns here and there and rearrange them, basically a multi-dimentional Markov chain. There's no creativity whatsoever in this mechanical process.

The creativity happened with the patterns beforehand that LLMs were (illegally) trained upon.

Correct. If something like Sunos spits out something creative, it's because it's mimicking something creative it was trained on, or it accidentally stumbled upon it. When I add elements of something I was "trained on," it has the varnish of emotional weight to it. If you detect a bit of the band Madness in one of my songs, know that I have a core memory of making out with my first girlfriend at a Madness concert on the pier in NYC. That girlfriend later ghosted me after I bought us tickets to a Utopia (Todd Rundgren's band in the 80s) show she wanted to go to. She showed up with some bizarro excuse as to why she couldn't go, kissed me on the cheek and I never saw her again. That complex interaction of life, emotion and music, permeates everything we do as artists. It has zero to do with what generative AI does, because it is not alive, or capable of emotion.

[Tiles]

"Humans transform influences into art, AI transforms influences into output" is a reasonable argument.

"Humans learn from prior work but AI learning from prior work is theft" requires a lot more justification than people usually provide.

And while personal experiences can influence artistic choices, they don't magically create creativity. Every artist learns patterns, techniques and ideas from others and recombines them into something new. The debate is whether AI can do something similar, not whether it has emotions.

Also, proving that AI has no emotions is not the same as proving that it cannot generate novel output. Those are two completely different claims.

And as a side note, Suno isn't even an LLM. People keep using "LLM" as a catch all term for AI, but music generation models and language models are different technologies.

[ksandvik]

The models are the same, just used with different stolen content for training. Suno is nothing spectacular technology-wise. Same with any AI music, it's basically musak for this age. Glad if someone likes it, though!

[Tiles]

So we've gone from "it's just a Markov chain" to "all AI models are the same" to "the data was stolen".

Those are three completely different arguments, and I think you're misunderstanding all three of them.

If the point is that the training data was obtained unethically, make that argument. But that doesn't prove the output is incapable of being novel, and it doesn't make music generation models and LLMs the same thing.

And even then, the vast majority of training data is not "stolen" in the first place.

[BONES]

If something like Sunos spits out something creative, it's because it's mimicking something creative it was trained on, or it accidentally stumbled upon it.

OK, so exactly like humans, then. I've asked this before but what sort of music do you think you'd make if you had never heard any music before? The music you make is as much a product of your experience of other people's music as an LLM's is of its training. If you can't see that it because you refuse to look.

When I add elements of something I was "trained on," it has the varnish of emotional weight to it.

Similarly, when we get an AI to generate something for us, it has the varnish of emotional weight we have imbued it with, or at lest the emotional weight we have chosen from the options presented. Our AI collaborations probably express our emotions better than any of our previous work because the way they have come about has made us think more deeply and directly about that aspect of it. It's an almost perfect expression or our emotional intent. It took a lot of effort to get that from it but, listening to the finished album now, it is there and is undeniable.

"Humans transform influences into art, AI transforms influences into output" is a reasonable argument.

Maybe until you hear the shit they try to pass off as "art". I mean, seriously, is Shaddap You Face art? Is Barbie Girl or I'm Too Sexy? No, that stuff, like so much popular music, is commercial product, pure and simple. Of course, the same is true of a lot of AI output but that's because the human doing the prompting wants it that way. If you want to use AI to make product, you can choose to do that but if you want to use it to make art, it is equally capable. Art or product is entirely at the discretion of the human in charge of the process.

The models are the same, just used with different stolen content for training.

Stolen how? Is any of it not freely available to the public, through websites, public libraries, radio and TV broadcasts, etc.? If it's freely available to us, how can it be "stolen" when an AI accesses it for training purposes?

[Tiles]

Maybe until you hear the shit they try to pass off as "art". I mean, seriously, is Shaddap You Face art? Is Barbie Girl or I'm Too Sexy? No, that stuff, like so much popular music, is commercial product, pure and simple. Of course, the same is true of a lot of AI output but that's because the human doing the prompting wants it that way. If you want to use AI to make product, you can choose to do that but if you want to use it to make art, it is equally capable. Art or product is entirely at the discretion of the human in charge of the process.

Heh, true

But even then, it is still art, just in its own way. I understand your point that it is often created using the same formulas and patterns, which makes the artistic aspect less obvious. A good pop song is probably 99% craftsmanship and 1% artistic expression, whereas something like art rock places a much greater emphasis on the artistic side. That does not mean the artistic element is absent, only that it plays a smaller role. And of course, the real art here is to make lotsa money with it.

[ksandvik]

The models are the same, just used with different stolen content for training.

Stolen how? Is any of it not freely available to the public, through websites, public libraries, radio and TV broadcasts, etc.? If it's freely available to us, how can it be "stolen" when an AI accesses it for training purposes?

Maybe you don't know how Suno got access to the massive amount of audio material they illegally downloaded and used for generating their LLM datasets. Unless they had a trillion or so dollars to pay to each and every content owner for the tracks .