Can AI (Artificial Intelligence) even compete?

[havran]

More importantly, will it cull the herds of musical posers?

Now just wait a minute -- perhaps we might need to question the tacit underlying assumption that muslcal posing is wrong.

[BBFG#]

True.
And they have their place by giving us a base of comparison!

Of course, it might increase the suicide rate as well...

[Zeisner]

Maybe that could help to explain the Fermi Paradox.

If you're not an old fart like me you may have the opportunity to find out in a few decades when the industrial civilization finally collapses.

[havran]

Maybe that could help to explain the Fermi Paradox.

If you're not an old fart like me you may have the opportunity to find out in a few decades when the industrial civilization finally collapses.

Oh -- I was thinking 2027 or 2028 at the latest.

[Zeisner]

After 2040 according to that MIT study. The BAU scenario seems to be the most likely based on recent data. If there's no nuclear third world war in the meantime of course. That would make things a little more complicated...

[enroe]

Does AI even feel competitive, or would it ever?

Considering the grand view, what if most organic (or similar) intelligent species get supplanted by their own AI because they ask it the wrong questions? What if their AI progeny subsequently just sit there quietly, tended by robotic caretakers, calculating idly and aimlessly among themselves without any kind of ambition or drive? Maybe that could help to explain the Fermi Paradox.

Haha, you've probably read a lot of science fiction novels.

But I think the "takeover of (almost) everything by AI" will happen
quietly and subtly – almost imperceptibly to us humans. In the
decision-making bodies of states and companies, AI will initially
provide more and more support. It will be consulted more and
more, integrated into more and more decision-making
processes. Until a point where there are still governments,
mayors, and parliaments, where there are still executive boards
and supervisory boards – but the big, important decisions will
ultimately be made by AI.

Whether this is an advantage or a disadvantage, no one can
say today: Will AI work in the interests of the state and the
community – or will it consolidate the power of a small circle?

Our generation will probably never know. Nor will we know
whether AI will be the "great filter" — and thus explain the Fermi
paradox.

What we can probably predict for the next few years is that AI
will create a large number of creative works in art and music
itself – thereby completely devaluing and eliminating human
creativity. We are already seeing this, for example with Spotify.

[BONES]

Spotify didn't need AI to devalue human creativity. You have to understand that the music industry is a popularity machine, it has minimal interest in art or creativity. It exists to generate profits and if AI allows the industry to generate bigger profits, then the industry will embrace it wholeheartedly. But the likes of you and I live outside of the industry or, at best, on the margins of it. It will be game over for all the big stars once the industry works out how to market AI music effectively, which might take a while but will happen eventually. TikTok will likely play a big part in that but we'll mostly get by like we always have.

[luciansyn]

The audio language models are actually incredible but as time goes on there is less and less interesting music in the training data. The early google AI test kitchen / AudioLM model was pretty wild. It could make wild dark psytrance, breakcore played on a glass instrument with mallets. I imagine there was a ton of copyright material in that training iteration though. I suspect what has happened is the models have got worse and worse since then because the training data has become more and more limited. It is just a bad investment bothering to license more obscure genres to put in the training data. These models are also not at all made with musicians in mind. I have never seen one that had a 12-tone row as a label.
I suspect we actually could have the greatest synthesizer of all time right now but it won't bother to be trained and realized. Then there are just a lot of closed minded, ignorant motherfuckers pretending to be "creative" who would see this as a bad thing anyway. A 21st century version of "REAL musicians play the oboe!".

[zerocrossing]

The audio language models are actually incredible but as time goes on there is less and less interesting music in the training data. The early google AI test kitchen / AudioLM model was pretty wild. It could make wild dark psytrance, breakcore played on a glass instrument with mallets. I imagine there was a ton of copyright material in that training iteration though. I suspect what has happened is the models have got worse and worse since then because the training data has become more and more limited. It is just a bad investment bothering to license more obscure genres to put in the training data. These models are also not at all made with musicians in mind. I have never seen one that had a 12-tone row as a label.
I suspect we actually could have the greatest synthesizer of all time right now but it won't bother to be trained and realized. Then there are just a lot of closed minded, ignorant motherfuckers pretending to be "creative" who would see this as a bad thing anyway. A 21st century version of "REAL musicians play the oboe!".

One thing I'm noticing in the job market is "music director" jobs... where you are directing a generative AI music model. I'm OK at the moment, so I can pass those jobs up, but I wonder how many people facing financial hardship might be prostituting themselves to help train these models. With so little paying jobs in the music industry, I don't blame them.

But AI is like a dog. Give it a treat every time it does its trick well, and it will do that trick a lot. Tricks that don't get a treat get forgotten. If everyone looking for obscure genres and more experimental music ignore it, it will go where the treats are. In this way, AI is kind of like a mirror of humanity. Deep down, we love slop and generic crap. People like us are outliers, and always have been.

[Zeisner]

One thing I'm noticing in the job market is "music director" jobs... where you are directing a generative AI music model.

...doing the data poisoning for people like me. Neat!

[BONES]

The audio language models are actually incredible but as time goes on there is less and less interesting music in the training data. The early google AI test kitchen / AudioLM model was pretty wild. It could make wild dark psytrance, breakcore played on a glass instrument with mallets.

My initial experience with Suno was the opposite - it only knew how to do popular style/genres. It had no idea at all what Industrial or EBM was and couldn't make anything that sounded even remotely like it. I think that mostly came down to the very limited range of vocal styles it had in its arsenal, which may not have affected EDM sub-genres.

I don't think it matters how much interesting music is in the training data, as long as there is some of everything. The formula for most genres is pretty simple, it doesn't need too many reference points to get something going. Of course, there is also the ability to load up a reference track, so you can train it yourself, to a limited degree. It's something my bandmate does really well - he gives it a prompt and uploads a reference track, then gets it to generate some options (Tunee will give you up to 6 at a time). From there he picks the one he likes, uses that as a base, then he uploads a new reference track and generates some ore options. Do that three or four times and you're giving the AI targeted training material as you work.

I imagine there was a ton of copyright material in that training iteration though.

How does that matter? The AI isn't breaching copyright by consuming training data, any more than you or I are for listening to music. If OpenAI has a Spotify account, then ChatGPT should have the same right to access Spotify's content as any other account holder, shouldn't it?

It is just a bad investment bothering to license more obscure genres to put in the training data.

Do they even do that? If they want our music to use as training data, they are welcome to it, no licensing required.

People like us are outliers, and always have been.

I've noticed that comparatively few people around here realise that. I think they have lived their lives mostly in the centre of the bell curve and they flounder when they find themselves a couple of stand deviations away from the cosy middle, where everything is laid on for them.

[vurt]

After 2040 according to that MIT study. The BAU scenario seems to be the most likely based on recent data. If there's no nuclear third world war in the meantime of course. That would make things a little more complicated...

hold thst thought...

[zerocrossing]

After 2040 according to that MIT study. The BAU scenario seems to be the most likely based on recent data. If there's no nuclear third world war in the meantime of course. That would make things a little more complicated...

hold thst thought...

Coming right up!

[BertKoor]

> > I imagine there was a ton of copyright material in that training iteration though.
> How does that matter?

Not because of the (c), it matters because a large portion of data was removed, hence there are now less diverse examples.

[BONES]

What do you mean that data was removed?