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

[mi-os]

So ignore the replicators but utilize and encourage the development of the tools rather than discarding the technology entirely.

Where did i "discard the techology entirely"?

[guitarzan]

Where did I say you did anything?

[El°HYM]

Where did I do anything you said?

[guitarzan]

The developers of Band In A Box should train an AI agent on their complete software package and have it reorganize that 40 years of bloated mess into a coherent new experience with user prompt AI collaboration ability… is another potential use for AI maybe.

[mi-os]

Where did I say you did anything?

Are you for real? You wrote "rather than discarding the technology entirely". What do you think this insinuation means?

[guitarzan]

Lighten up, Francis — I post of generalities. It’s not about you.

[mi-os]

Um, yeah, whatever.

[Zeisner]

Instead of rejecting AI, the industry should be scrambling to develop the killer apps that utilizes AI in a way that musicians can appreciate, either by outputting MIDI or perhaps designed to help build user specified AI sample libraries, etc

The first already exists and doesn't need LLM, not even machine learning. The latter is simply out of reach because you need:

-> A lot more data centers for processing. Imagine multiplying the current amount with 1000. At least. Just to reduce artifacts.

-> A lot of new and primarily unprocessed data. More than what has ever been saved online and offline so far in human history. That's not even possible with turning every OS and DAW into a data miner.

-> A lot of professional musicians, instrument makers and audio engineers for supervised learning and model finetuning.

The costs for access to such a service - of course there's no way to run it locally - would be insane. There's no way to generate revenue with such a product. Absolutely impossible.

What is possible and realistic: A simplified machine learning program you can train yourself with your own samples. Of course without LLM, you would control it like a synth. The output would be full of artifacts and very glitchy. Could be nice for genres like industrial though.

[guitarzan]

I think a local LLM focused on music theory would be possible now, or in the near term. A not so extremely large LLM.

For something like Band In A Box, all the samples and theory and everything else is already in there, just let the AI agent take it all and organize it into something coherent and accessible to AI prompt based collaboration. That approach could probably apply to tons of old generative software and involve more human input than previously with results that more closely match the musician’s intent. The AI element would actually make the outcome more human.

And I don’t really see why it would be harder for AI to produce a set of samples than it would be for it to produce completed songs. “Make me a tuba sample set with the following attributes…” should be easier than “produce a rock opera for me”, right?

[Zeisner]

I think a local LLM focused on music theory would be possible now, or in the near term. A bit so extremely large LLM.

You don't need a LLM for that. Music theory can be perfectly described with simple math. Harmony helpers and similar tools have been around for a while now. Example: http://www.midiplugins.com/Plugin/298

For something like Band In A Box, all the samples and theory and everything else is already in there, just let the AI agent take it all and organize it into something coherent and accessible to AI prompt based collaboration.

Something like that could be done indeed. But you would still need the online connection for the LLM function to work which means another company would be permanently involved. It's questionable if the additional afford would pay off in the long run.

[guitarzan]

Yes, great generative software has been around almost as long as computers, but AI could make interfacing with the music theory algorithms a much richer experience.

It could help you decide which scales, chords, etc, and humanize the results. It can collaborate. If AI can now do entire songs with complete orchestrations in seconds it can certainly do killer arpeggios and sequences in MIDI that are individually crafted through prompts to fit in a specific tune, fitting ideas of how the tune will progress and such.

And I think soon local LLM will become common, handling a lot of what needs to be done in the box, and then the local LLM can reach out to the big LLM online only when necessary, reducing online demand.

[BBFG#]

'AI could help you humanize the results'...

[guitarzan]

'AI could help you humanize the results'...

Yes, because the results are arrived at through a conversion consisting of exchanged prompts rather than just having a robotic quantized scale exercise based on the key selected spewed out like has been done for 40 years with generative composition software, like the arpeggiators used by Jan Hammer that I mentioned before.

[Zeisner]

Yes, great generative software has been around almost as long as computers, but AI could make interfacing with the music theory algorithms a much richer experience.

The tool I linked to couldn't be any simpler, you only have to press a button - instead of writing a short story. It doesn't get more efficient than that. Like... do you want to make music or do you want to chat?

It could help you decide which scales, chords, etc, and humanize the results.

You're outsourcing composition if you let an AI do that stuff for you. You don't make music anymore then. Aside from that, AI can't humanize anything. Only humans can with their choices regarding composition and arrangement. And of course performance, something AI can never replicate or assist with.

it can certainly do killer arpeggios and sequences in MIDI

But such functions already exist in DAWs and plenty of plugins. Without the need for machine learning. All that's needed is some MIDI routing. Which is less work than writing prompts. And more precise too.

[Zeisner]

Yes, because the results are arrived at through a conversion consisting of exchanged prompts

That doesn't solve the averaging issue. Artificial neurons react the most to the modal value of a given data set. The modal value has the highest influence on probability. But this is not how humans perform, there is no such thing as an average human with average performance. It doesn't sound good or realistic because it's too average or "smoothed out". If you want a good human performance you need a good human player. There is no way around that, not with all computers in the world. It's just too complex statistically.