Artificial Intelligence and Music

[bender]

I can see how it would be useful for creating gingles for ads, or even movie score type stuff, but that’s just taking away another human’s job.

There’s definitely a market where the client doesn’t really give a fuck about what the music is actually doing- they’re just looking for wallpaper. That’s where AI generated music will hit hard first. Jingles, maybe… but in my experience the clients can get incredibly granular (and often ridiculous) with their requests and I could see it either being perfect (because writing jingles often feels like you’re just being fed a bunch of random buzzword prompts), or utterly useless (because they genuinely want something that feels unique and identifiable).

Movie scores… maybe- there’s been a trend that’s homogenising them, but that seems to be on the wane. I could definitely see editors using AI generated music to temp score a film (because that tends to need to be done fast, and just has to sound convincing to a casual listen), but movie scores really do thrive on finding new sounds (before rehashing them as infinitum until the next “thing” comes along.

I’m definitely not saying that AI won’t replace these things, but I do think we’d potentially lose something in the process.

Production library music is going to face huge upheaval for sure- The libraries seem to be consolidating in anticipation of this (smaller ones selling their catalogues to huge ones, and heading in the direction of only having one or two mega libraries).

It really feels like the blander and more homogenised the music gets, the more of a threat AI will be to it - look out pop music I guess.

I’m really hoping that legislation, developers, and consumers lead it more in the direction of creating tools for creatives, rather than methods to literally dispense with the human input.

All this might be wishful thinking, of course… it’s all getting very close to my profession.

[bender]

That’s hilariously brilliant!

Agreed. I snorted my tea when I read it

[jeremyb]

One of the issues with more and more AI generated content being out there on the web is that the generative models are sourcing their input data from the web so its going to erode it over time, especially when models are already hallucinating more as they grow in size and complexity, garbage in equals garbage out...

[NippleWrestler]

One of the issues with more and more AI generated content being out there on the web is that the generative models are sourcing their input data from the web so its going to erode it over time, especially when models are already hallucinating more as they grow in size and complexity, garbage in equals garbage out...

Given that bots can be used to jack up streams/listens/plays on streaming sites, and given that AI can make the content to furnish those streams we're going to be living in a world where bots are playing stuff for bots.

Then there'll come AI avatars to 'perform' this music.

That reminds me of something else actually - journalists in Mexico and Venezuela etc are using AI avatars and AI generated voices to hide their real face and voice so they don't get killed by cartels.

[bender]

One of the issues with more and more AI generated content being out there on the web is that the generative models are sourcing their input data from the web so its going to erode it over time, especially when models are already hallucinating more as they grow in size and complexity, garbage in equals garbage out...

Given that bots can be used to jack up streams/listens/plays on streaming sites, and given that AI can make the content to furnish those streams we're going to be living in a world where bots are playing stuff for bots.

Then there'll come AI avatars to 'perform' this music.

That reminds me of something else actually - journalists in Mexico and Venezuela etc are using AI avatars and AI generated voices to hide their real face and voice so they don't get killed by cartels.

Amazing

[Litterick]

Pop will eat itself.

A paper published in Nature shows machine learning models are fundamentally vulnerable. As the authors write in the paper’s introduction: 'We discover that indiscriminately learning from data produced by other models causes “model collapse” — a degenerative process whereby, over time, models forget the true underlying data distribution.'

As TechCrunch puts it: ''Basically, if the models continue eating each other’s data, perhaps without even knowing it, they’ll progressively get weirder and dumber until they collapse. The researchers provide numerous examples and mitigation methods, but they go so far as to call model collapse “inevitable,” at least in theory.''

[jeremyb]

Fundamentally its not Artificial Intelligence as its not from the AI region of France, its just Sparking Machine Learning

[Litterick]

San Francisco Standard: 'Dozens of AI workers turn against bosses, sign letter in support of Wiener AI bill; Thirty-seven signatories are employees of companies opposing SB 1047, including OpenAI, Google DeepMind, and Meta'

...
The brief letter warns “that the most powerful AI models may soon pose severe risks, such as expanded access to biological weapons and cyberattacks on critical infrastructure.”

[hamo]

San Francisco Standard: 'Dozens of AI workers turn against bosses, sign letter in support of Wiener AI bill; Thirty-seven signatories are employees of companies opposing SB 1047, including OpenAI, Google DeepMind, and Meta'

...
The brief letter warns “that the most powerful AI models may soon pose severe risks, such as expanded access to biological weapons and cyberattacks on critical infrastructure.”

I misinterpreted this as the AI rising up against its creators...

[sty]

Pop will eat itself.

A paper published in Nature shows machine learning models are fundamentally vulnerable. As the authors write in the paper’s introduction: 'We discover that indiscriminately learning from data produced by other models causes “model collapse” — a degenerative process whereby, over time, models forget the true underlying data distribution.'

As TechCrunch puts it: ''Basically, if the models continue eating each other’s data, perhaps without even knowing it, they’ll progressively get weirder and dumber until they collapse. The researchers provide numerous examples and mitigation methods, but they go so far as to call model collapse “inevitable,” at least in theory.''

This is why the really big players in AI are and have been spending an enormous amount of money and effort producing and curating the training data for their AIs.This is why they've been buying data from Redit, our very own IRD, etc.

It's going to get harder to augment their current learning data now so much AI content is "in he wild", which will require more and more cleansing to ensure it's not AI generated - so the irony is that the AI companies will likely become experts at detect AI content (and will likely keep that tech very secret). Also following the usual pattern those players with the deepest pockets will likely have the best trained AI and will squeeze out the poorer companies.

We will see a load of crap AIs that will have been trained on leaked data and less well cleansed training material, but they'll do a decent enough job to make a living I guess.

[hamo]

I still maintain that the best and safest use cases for AI will be smaller models trained on specific and well curated content.

[jeremyb]

I still maintain that the best and safest use cases for AI will be smaller models trained on specific and well curated content.

Celebrity nudes for example.

[hamo]

I still maintain that the best and safest use cases for AI will be smaller models trained on specific and well curated content.

Celebrity nudes for example.

Not exactly where I was going with that...