Will Ai be the next big step in music?

[DCrown]

it started with the release of Pro Tools in 1990 ies when non-musicians started producing music, they were called DJs and I knew some who earned a lot of money. And they had no clue how to play an instrument.

[lobanov]

My response to all that replied to "Check this guy out. "
!st off, it's not about the guy. I couldn't care less about that annoying user, I mostly skipped ahead. My interest was the Ai music program designed for Turkish music. People are so negative here. Are you all depressed? You need to cut everything down and can't see the novelty in things. No one had a single comment about the Makyo Ai experiment I posted just above that. Maybe because it's harder to find anything negative to say about it. (that's not a challenge)

- I don't know about everyone here but I am not depressed at all.

- What can I tell about "Makyo Ai experiment' if you just posted several words as a very vague description of it? What should I tell about it?

- What novelty don't we want to see? There is nothing new in AI generated music. AI just reproduces what we all have heard many times before.

- You tell "for good". Is polluting the world with AI generated music a good? I don't think so. Why are you so positive about AI? Why don't you see problems? And how about "for better", not just "for good"? What does AI 'bring to the table'? I mean something new, except the ease of 'composing without composing', 'arranging without arranging', 'playing without playing' and 'recording without recording'. And 'being a musician without being a musician'.

[Bombadil]

Terrible waste of electricity and whatever other resources it eats up.

[Zeisner]

This guy exists but he does nothing.

Why would a real person use an AI-generated avatar? Hint: Zoom in and look at the eyes. Those are not scaling artifacts, such distortions are typical for generative AI. The whole image is full of artifacts but it might be easier for you to see it in the eyes.

[Zeisner]

What does AI 'bring to the table'?

There are special cases when deep learning can be useful, like researchers looking for a protein with very specific properties or finite element method enhancements. But this is completely different from generative AI because you have experts who carefully choose training data, adapt or even write models and test the results. Generative AI on the other hand has been created by idiots who don't know or understand what they're doing. Example: Using audio data compressed with lossy codecs as training material. Only an idiot would do that.

[lobanov]

What does AI 'bring to the table'?

There are special cases when deep learning can be useful, like researchers looking for a protein with very specific properties or finite element method enhancements. But this is completely different from generative AI because you have experts who carefully choose training data, adapt or even write models and test the results. Generative AI on the other hand has been created by idiots who don't know or understand what they're doing. Example: Using audio data compressed with lossy codecs as training material. Only an idiot would do that.

I agree with you. I don't think that AI isn't useful at all. AI is excellent as a translator, good for synopses. But we talk about generative AI in music. And even this isn't bad if we need background music. Results impress. But I don't understand why I should take this seriously and admire it. I want more cool real music made by real humans.

P. S. About real 'hacking'. What if we train AI on its own production rather than real music?

[Zeisner]

AI is excellent as a translator, good for synopses.

Only on material that correlates high with the modal value of the given data set (Most common). The more you deviate from that, the more hallucinations you'll get. This applies to any data set, that's why you need experts to pick high quality data, something that is simply not possible for music. It will never be possible. Hiring the top quartile of musicians and audio engineers from all around the world (!) to generate the required training data (because it barely exists or isn't accessible) and do the model finetuning would be too expensive, not to mention the additional processing power. It would also be a project for decades. Such a music AI would easily cost more than a thousand times than everything that has ever been spent on all AI before, including all research ever done - just to match the low error rate of "simple" projects like finding the right protein. And you would still not get an AI that could invent new genres or styles, only an AI that would produce musical/sound artifacts below the thresholds of human perception.

About real 'hacking'. What if we train AI on its own production rather than real music?

The result would be model collapse. For audio you'll get unweighted=white noise.

[jancivil]

Chat GPT trained on itself will result in gibberish within six generations, acc'ding to Michael Woolridge

[4damind]

I don't think this "AI slop" is here to stay it's mostly hype because it's new. The real "next big step" won't be prompt-based music, but smarter tools. I see AI being used for ingenious emulations and plugins where traditional circuit modeling hits a wall. Small, locally-run AI models acting as tools to speed up the creative process and that’s the future. At least I hope that "push-button" music isn't the final destination

[Bunny_boy]

Chat GPT trained on itself will result in gibberish within six generations, acc'ding to Michael Woolridge

It's weird, but that has been mentioned a few times in these threads but it's never become part of the main discussions. I mentioned in this or one of the other ones about the outcome of AI music trained only on AI music for a few generations.

[Zeisner]

Chat GPT trained on itself will result in gibberish within six generations, acc'ding to Michael Woolridge

OpenAI already had to do two rollbacks last year because ChatGPT put out gibberish. It doesn't take that many generations to make LLMs useless.

[Zeisner]

It's weird, but that has been mentioned a few times in these threads but it's never become part of the main discussions.

Because it scares the crap out of all the AI fanatics.

[Zeisner]

The real "next big step" won't be prompt-based music, but smarter tools. I see AI being used for ingenious emulations and plugins where traditional circuit modeling hits a wall. Small, locally-run AI models acting as tools to speed up the creative process and that’s the future.

You don't need deep learning for that though. Deep learning is only needed for very specific cases, just like quantum computing.

[jancivil]

AI fans WANT TO BELIEVE but they're not Fox Mulder investigating. Magical thinking rules OK.

Those arguing for it here as though from a musician's perspective seem unclear on the concept of musical creation or skill. It's no great surprise, it's the logical extension of your DAW tells you what chords are safe with a given scale (after it tells you what the scale is). The fun part is avoided in favor of WHAT?

"He sings his parts and [Eita] assembles it with instruments." Why? If he can indeed sing the thing, the next step is to write parts for it. If that's not part of his facility, this is where one obtains the facility. Is he lazy? Is he looking for an angle trying to get exposure with this gimmick? If he is not real, what sort of person thinks this is any kind of way to make music? Answer suggested in comments: morons. Who is pushing this? What are you doing pointing to this?
Nothing about this is clean.

[Zeisner]

AI is already having a big impact but in a very different way than expected. After RAM and GPU it's now hard disks getting very expensive. It won't make much of a difference if you buy a real 1176 or a plugin clone if this continues...