Maybe there is a way to fix the AI problem!? (Let's talk about how we can handle this sh*t)

[chagzuki]

I'm always on the fence as to whether I'm winding down or done with making music, and making things in general. It all depends on whether the potential end result excites me, and the process. So the activity is in an indeterminate zone between those two sources of satisfaction.
Then there's the question of the acquisition of knowledge, the means to arrive at a satisfying end product, or a part of the satisfaction derived from the process.
Learning stuff can be tedious and difficult... it can also be enjoyable. A flow state is ideal, but in reality we're always oscillating around that ideal and often a long way from it.

AI comes along and offers to generate things based on super-deep levels of understanding, in ways which humans will never be capable of.

As Bones and others are saying, this can be incorporated into one's own creative process.
It can also remove the requirement to go through difficult learning processes of one's own.

I have no idea where I stand on what's worth learning for oneself, what depth of understanding of harmony, mixing etc.

Currently, it strikes me that for what I might do going forward, I most likely need to go deeper into difficult traditional areas of composition, and if I use AI in that area, unless I progress on my own learning path I won't have the skills to manipulate the things that AI generates. That's probably the risk, the more you rely on AI the less likely it is that you develop the skills to tailor what it produces in accordance with some artistic vision.
It's also possible that you manage to do both, develop your own skills and use the output of AI.

Then again, maybe we're already drowning in audio-visual content and it makes more sense to quit, and do other shit instead.

[zerocrossing]

My prediction is that AI is a problem that will fix itself.

In addition to other points you made in your post, there's just the power-consumption/requirements and waste heat to be considered.

I have read that every AI company that's charging a fee for the "pro" version is operating at a huge loss. I guess the idea is to hook a business on it, and then when all your workforce is gone, jack up the price to make up for the loss and turn a profit. My guess is, some businesses will find it's worth it, but most won't. For instance, my wife's company has basically replaced all its level one support and find it as good or better than humans, but the rubber hits the road when more extensive help is required, and a human needs to intervene.

What will be interesting is, once all these data centers are built, power and water for cooling obtained, and the market can't support all that infrastructure, what's going to happen? If they keep going, what's going to happen when energy prices spike? "Sorry you can't afford to heat your home, the president needs to sh!tpost."

[zerocrossing]

I'm always on the fence as to whether I'm winding down or done with making music, and making things in general. It all depends on whether the potential end result excites me, and the process. So the activity is in an indeterminate zone between those two sources of satisfaction.

Highlight mine. I can't speak for others, but to me, the process and act of making music is everything. An injury once prevented me from playing for a year or so. I'd been depressed and throwing myself into distractions to placate myself. When I realized what was happening, I vowed I'd never do that again. AI or not, is really irrelevant. I make music because I enjoy doing it. I almost have to.

AI comes along and offers to generate things based on super-deep levels of understanding, in ways which humans will never be capable of.

This is a misconception. What we call AI has no understanding at all. It's why it's likely to produce something nonsensical or flat out wrong. It has no real understanding of right or wrong, good or bad. It does not know pain, so how can it know the pain of something universal, like being rejected in a relationship? It can mimic it, but never 'know' it.

Currently, it strikes me that for what I might do going forward, I most likely need to go deeper into difficult traditional areas of composition, and if I use AI in that area, unless I progress on my own learning path I won't have the skills to manipulate the things that AI generates. That's probably the risk, the more you rely on AI the less likely it is that you develop the skills to tailor what it produces in accordance with some artistic vision.
It's also possible that you manage to do both, develop your own skills and use the output of AI.

Then again, maybe we're already drowning in audio-visual content and it makes more sense to quit, and do other shit instead.

We're going to start getting into trouble when AI produced art becomes the dominant content on the internet, and it starts training on itself. Eventually, that kind of "inbreeding" results in more and more homogenous crap. A good result of this might be that people start working harder to do unique art.

[chagzuki]

This is a misconception. What we call AI has no understanding at all. It's why it's likely to produce something nonsensical or flat out wrong. It has no real understanding of right or wrong, good or bad. It does not know pain, so how can it know the pain of something universal, like being rejected in a relationship? It can mimic it, but never 'know' it.

We're going to start getting into trouble when AI produced art becomes the dominant content on the internet, and it starts training on itself. Eventually, that kind of "inbreeding" results in more and more homogenous crap. A good result of this might be that people start working harder to do unique art.

Well, it has no understanding in the sense of the experience of the meaning. But it does in terms of the codification of a structure.
I know nothing really about how models are created, but I gather it's based on the input of human preference?
If AI music uses AI produced music as the source upon which to be trained, so long as there's some human steering the selection of output then there is an experience of meaning in the loop.

I had to ask mr Grok about the human training of music models, and he told me this:

"Yes, many modern AI music generation models incorporate human feedback—often in the form of voting, ranking, or rating the quality of outputs—to align their generations with what humans find enjoyable, coherent, or aesthetically pleasing. This is typically achieved through techniques like Reinforcement Learning from Human Feedback (RLHF), where humans evaluate generated music samples (e.g., on aspects like melody, harmony, rhythm, and overall appeal), and this data trains a "reward model" that fine-tunes the AI to prioritize human-preferred outputs. How This Works in PracticeRLHF for music generation generally follows these steps: Pre-training on Large Datasets: The base model learns patterns from vast audio corpora (e.g., songs, MIDI files) without direct human input, similar to how language models learn from text.

Human Evaluation Phase: Annotators (humans) listen to pairs or sets of generated clips and provide feedback, such as pairwise comparisons ("Which sounds better?") or ratings on scales for qualities like "catchiness" or "emotional fit to a prompt."

Reward Model Training: The feedback data trains a separate model to predict human preferences, acting as a proxy for "what humans like."

Fine-Tuning with Reinforcement Learning: The main model is optimized using this reward signal, often via algorithms like Proximal Policy Optimization (PPO), to generate outputs that maximize the predicted human approval.

This process ensures the model "embodies" human tastes, making outputs more engaging and less random. Without it, generations might mimic data statistically but lack subjective appeal."

[zerocrossing]

This is a misconception. What we call AI has no understanding at all. It's why it's likely to produce something nonsensical or flat out wrong. It has no real understanding of right or wrong, good or bad. It does not know pain, so how can it know the pain of something universal, like being rejected in a relationship? It can mimic it, but never 'know' it.

We're going to start getting into trouble when AI produced art becomes the dominant content on the internet, and it starts training on itself. Eventually, that kind of "inbreeding" results in more and more homogenous crap. A good result of this might be that people start working harder to do unique art.

Well, it has no understanding in the sense of the experience of the meaning. But it does in terms of the codification of a structure.
I know nothing really about how models are created, but I gather it's based on the input of human preference?
If AI music uses AI produced music as the source upon which to be trained, so long as there's some human steering the selection of output then there is an experience of meaning in the loop.

I had to ask mr Grok about the human training of music models, and he told me this:

"Yes, many modern AI music generation models incorporate human feedback—often in the form of voting, ranking, or rating the quality of outputs—to align their generations with what humans find enjoyable, coherent, or aesthetically pleasing. This is typically achieved through techniques like Reinforcement Learning from Human Feedback (RLHF), where humans evaluate generated music samples (e.g., on aspects like melody, harmony, rhythm, and overall appeal), and this data trains a "reward model" that fine-tunes the AI to prioritize human-preferred outputs. How This Works in PracticeRLHF for music generation generally follows these steps: Pre-training on Large Datasets: The base model learns patterns from vast audio corpora (e.g., songs, MIDI files) without direct human input, similar to how language models learn from text.

Human Evaluation Phase: Annotators (humans) listen to pairs or sets of generated clips and provide feedback, such as pairwise comparisons ("Which sounds better?") or ratings on scales for qualities like "catchiness" or "emotional fit to a prompt."

Reward Model Training: The feedback data trains a separate model to predict human preferences, acting as a proxy for "what humans like."

Fine-Tuning with Reinforcement Learning: The main model is optimized using this reward signal, often via algorithms like Proximal Policy Optimization (PPO), to generate outputs that maximize the predicted human approval.

This process ensures the model "embodies" human tastes, making outputs more engaging and less random. Without it, generations might mimic data statistically but lack subjective appeal."

Right, but look at it this way. If you want back in time to the 1940s with a Metallica album, people would hate it. It’s a cliche that genius artists aren’t often recognized in their time for a reason. Societies create culture based on numerous factors that are always in flux. An LLM that is trained on what people today think of art is going to be a static snapshot. It’s only chance on making something new is via random mixing of past examples. This is not intent. As AI slop dominates training content, it will become lamer and lamer. Like an animal eating its own poop, ad infinitum. It won’t know the difference because it can never feel the urge to get laid.

[chagzuki]

Right, but look at it this way. If you want back in time to the 1940s with a Metallica album, people would hate it. It’s a cliche that genius artists aren’t often recognized in their time for a reason. Societies create culture based on numerous factors that are always in flux. An LLM that is trained on what people today think of art is going to be a static snapshot. It’s only chance on making something new is via random mixing of past examples. This is not intent. As AI slop dominates training content, it will become lamer and lamer. Like an animal eating its own poop, ad infinitum. It won’t know the difference because it can never feel the urge to get laid.

Yes, I mean, I've hated the majority of the output of the music industry most of my life, and it seems to me that AI streamlines the approach of that big fat corporate industry.
I guess then the question remains as to what extent AI tools can be used within an indie approach to music making, and we're certainly going to find out as all manner of tools become available.

[bermudagold]

exactly..to me there are three elements of music that really resonate with humans the most
1. Melody:

The popularity of all forms of dance music would tend to disprove this hypothesis.

.they subconsciously create joy

Why would you want to create joy? I want my songs to create anger, to spur people into action. Joys is irrelevant, it's pathetic.

yeah that's why I said #2...that explains more simple forms of dance music that may even be devoid of 1 or 3

ur right...any visceral emotional response is the goal...I should not have oversimplified to joy

interesting takes for sure...I understand ur viewpoints from ur arguments...it's definitely a complicated debate which clearly will continue as AI proliferates

[bermudagold]

Then there's the question of the acquisition of knowledge, the means to arrive at a satisfying end product, or a part of the satisfaction derived from the process.
Learning stuff can be tedious and difficult... it can also be enjoyable.

That's probably the risk, the more you rely on AI the less likely it is that you develop the skills to tailor what it produces in accordance with some artistic vision.

exactly...agreed

[bermudagold]

It can mimic it, but never 'know' it.

An LLM that is trained on what people today think of art is going to be a static snapshot. It’s only chance on making something new is via random mixing of past examples. This is not intent. It won’t know the difference because it can never feel the urge to get laid.

precisely...this is the difference in compulsion, impetus, and incentive mechanisms I was trying to highlight

[zerocrossing]

Right, but look at it this way. If you want back in time to the 1940s with a Metallica album, people would hate it. It’s a cliche that genius artists aren’t often recognized in their time for a reason. Societies create culture based on numerous factors that are always in flux. An LLM that is trained on what people today think of art is going to be a static snapshot. It’s only chance on making something new is via random mixing of past examples. This is not intent. As AI slop dominates training content, it will become lamer and lamer. Like an animal eating its own poop, ad infinitum. It won’t know the difference because it can never feel the urge to get laid.

Yes, I mean, I've hated the majority of the output of the music industry most of my life, and it seems to me that AI streamlines the approach of that big fat corporate industry.
I guess then the question remains as to what extent AI tools can be used within an indie approach to music making, and we're certainly going to find out as all manner of tools become available.

I seriously think the whole thing is overblown. For a very long time, we've had tools that let people use "MIDI chord packs," beat generators, sample loops, etc, to allow people to get around having to do the work of learning how to construct music from scratch. How many of them have really had any success, not including the people selling that crap? I honestly don't think these new tools are much different. First off, they're not really fun to use in any way. You might get something good out of them, but will there be any sense of satisfaction out of it? Can you make money from the output when everyone has access to the same tool? So, why are you using something that offers no fun in the process, no sense of satisfaction, and no monetary plus side?

[zerocrossing]

It can mimic it, but never 'know' it.

An LLM that is trained on what people today think of art is going to be a static snapshot. It’s only chance on making something new is via random mixing of past examples. This is not intent. It won’t know the difference because it can never feel the urge to get laid.

precisely...this is the difference in compulsion, impetus, and incentive mechanisms I was trying to highlight

I say, "get laid" as a joking way to talk about the biological imperative of procreation that defines everything we do. Literally everything can be traced down to moving a subset of your DNA into a future generation. The fact that we have a limited time to do this makes the compulsion more acute. Software has no similar compulsions or limitations.

[vurt]

Then again, maybe we're already drowning in audio-visual content and it makes more sense to quit, and do other shit instead.

you could just do it, because you enjoy doing it?

if you don't, then yes, find something you do enjoy, we all deserve a little joy/fun in our lives.

[chagzuki]

if you don't, then yes, find something you do enjoy, we all deserve a little joy/fun in our lives.

I've focussed more on cutting out things I don't enjoy. I seem to have discovered that I'm fairly content doing nothing most of the time. Took me a long time to get there.

[vurt]

if you don't, then yes, find something you do enjoy, we all deserve a little joy/fun in our lives.

I've focussed more on cutting out things I don't enjoy. I seem to have discovered that I'm fairly content doing nothing most of the time. Took me a long time to get there.

can't knock being in peace.

[BONES]

I'm always on the fence as to whether I'm winding down or done with making music, and making things in general.

I was there around 15 years ago. It lasted for several years but I came out the other side more motivated, more focused than ever. It took 7 years to get our fourth album out, purely because I couldn't be arsed finishing it, but then that album spent two weeks at no. 1 on the German Alternative Chart, which got us onto one of the big festivals the following year. It was all the rehearsal we did for that festival that got me back into it and I've not lost the spark since, despite a few disappointments long the way. In the 8 years since that time, we've put out 3 more albums and two 4 track EPs, despite the f**k arounds of Covid and all the rest of it. In fact, this year is the first time we've ever managed new albums in back-to-back years. So don't lose heart, things can turn on a dime.

Learning stuff can be tedious and difficult... it can also be enjoyable. A flow state is ideal, but in reality we're always oscillating around that ideal and often a long way from it.

I used to love learning new software, different ways of approaching things, but that is one aspect that didn't survive my years in the doldrums. That said, I think a big part of my revival was getting to grips with Cubase and then Studio One, after 15 years or more with Orion. It opened up a whole new world of possibilities for us.

As Bones and others are saying, this can be incorporated into one's own creative process.
It can also remove the requirement to go through difficult learning processes of one's own.

I would never say that. If you're serious about using Ai, there is a very steep learning curve. My bandmate is a real wizz at it, I am f**king hopeless. I just don't have the patience to go through all the iterations, hoping it will get better. Because I lack the patience, I don't get the results he does. It's weird because in "normal" product5ion, it's the other way around. He's the guy who comes up with most of our musical ideas, but he gets bored and moves along to the next thing, which leaves me to do all the scut work to knock the ideas into finished songs.

I have no idea where I stand on what's worth learning for oneself, what depth of understanding of harmony, mixing etc.

I have no specific knowledge about any of that, beyond the absolute basics of how to form a major chord. I rely on my ears to tell me if something works or not. Even when I was starting out, they were always reliable.

Currently, it strikes me that for what I might do going forward, I most likely need to go deeper into difficult traditional areas of composition, and if I use AI in that area, unless I progress on my own learning path I won't have the skills to manipulate the things that AI generates.

That's where refining things comes into play - you tell the AI what you want and let it figure out the detail. You give it something like "key change in the chorus to propel the song forward" so it knows to jump up, not down, and let it figure out how that will work.

That's probably the risk, the more you rely on AI the less likely it is that you develop the skills to tailor what it produces in accordance with some artistic vision.

That's been true of my entire 45 year musical odyssey. I've always worked stuff out as I go. I've let it develop organically, I've never really tried to learn anything. I've not watched endless Youtube videos on the art of compression or mastering or any of that. I've just played around until I got something that sounded good.