Here's What Gemini Thinks About the Whole Creative AI Argument

Explore how Machine Learning and AI can expand musical creativity while keeping the human in the creative workflow. This forum is dedicated to respectful dialogue where diverse perspectives are welcomed.
RELATED
PRODUCTS

Post

zerocrossing wrote: Sun Mar 08, 2026 5:24 pm
KBSoundSmith wrote: Sun Mar 08, 2026 4:53 pm
zerocrossing wrote: Sun Mar 08, 2026 5:42 am What does zerocrossing think about AI?
Gemini wrote:Based on discussions on KVR Audio and Gearspace, zerocrossing (a long-time, prolific KVR member) generally views AI in music production with a mix of cynical pragmatism, skepticism about its "artistic" value, and an acknowledgment that it is an unstoppable technological shift.
His views, often expressed in threads about AI adoption, can be summarized as follows:
AI as a "Random Inspiration Machine" or "Tool": Zerocrossing views current AI tools not as true creators, but as advanced, albeit sometimes useful, assistants. He has noted that while AI can be great for overcoming writer's block or generating ideas, it is not fundamentally creative.
Skepticism of "Writer's Block" and "Prompt Jockeys": He often challenges the narrative that AI is needed for productivity. He has implied that if someone constantly relies on AI to overcome writer's block, they might be in the wrong hobby, stating, "most composers and musicians I know almost never have writers block".
Focus on the "Human" Element: He tends to emphasize that the value of music is in the human experience of creating it—the "doing," the "making," and the emotional, often messy process, which AI cannot replicate.
The "Super-Genius Parrot": He has described AI as a "super-genius parrot" that can remix and regurgitate information, but it lacks true consciousness or understanding, arguing that "intelligence" in AI is often marketing hype.
Pragmatic Adaptation: Despite his criticisms of AI art's lack of soul, he is not completely opposed to using AI tools. He has expressed interest in how AI can be integrated into an "indie approach" to music, noting that some AI-generated music can be "listenable" and "interesting".
"Music had a one night stand with sound design.....And the condom broke." This humorous quote from a recent KVR post (Feb 28, 2026) highlights his view that AI has irrevocably changed the landscape of music production.
Gearspace
Gearspace
+5
In summary, zerocrossing is not an anti-tech purist, but a skeptical artist who believes AI is a "soulless" tool that cannot replace the emotional, and, "alchemical" process of human creation, even if it is a useful assistant.
Most of this is not true, except that I'm a prolific poster on KVR. :lol:
I'm a bit of a late adopter when it comes to trying AI, but I've been experimenting with Claude, ChatGPT, and Gemini a lot recently. They each have distinct strengths and weaknesses (some of which I haven't fully explored for pay-wall reasons, but the implied use case is clear both from the experiments I've done and the pay-walls themselves). Some of my deeper experiments have gone into rabbit-holes which the engines themselves have described as "atypical", "black swan events" and so on. Interesting, but I don't have an intention of discussing those.

Rather, I'd offer a comparison. I'm reminded of software vs hardware synthesizer debates. While the debates of the early days of VSTs were before my time, so to speak, looking back to them and comparing, I see strong parallels to AI models today. Early softsynths showed promise, could give decent output (or trash, haha), but "analog purists" would have none of it (and often still don't, haha). Now I'm not indicating that AI is "the way of the future" -- I'm not one for jumping on hype trains, especially when there's billions of dollars behind it -- but based on my experiments, I see (for targeted use cases, and utitlizing specific models and setups for specific use cases) the promise in these models that some saw in the early days of softsynths.

Regarding Gemini specifically, it naturally wants to look at large volumes of data, do a deep analysis, and then provide a "synthesis" of that data. The free tier is locked to small volume context and operates quickly -- pretty much the opposite of its intended use case. So the free version is prone to hallucinations and unwarranted grand conclusions from small amounts of data. Althought FWIW, for some of my use cases, I actually find this a charming "lo-fi" glitch which has some interesting applications.

Anyway, just a point of interest that flashed into mind and I thought I'd share.
So, I read that as, "if you want an actual answer to a question, you need to pay for it." Of course, this isn't really talked about enough, IMO. I donate $100 to Wikipedia every Christmas. For that, I can go and do a search on nearly every topic, and get good information, or I could not donate anything and go to Wikipedia and get the same good information. So for $240 a year, I could subscribe to the better version of ChatGPT that has some "advanced reasoning," (not really) and get good information... maybe. But it will make me an image of cats dancing at a night club, so that's something. I could go the free version, and get bad information! How quaint!

The new version of ChatGPT is 25% better than the previous version. It also uses 4 times as much energy to get this 25% increase. It's out of training data, having looked at everything on the web. Because a lot of the web is now AI generated, it is eating its own sh!t at an alarming rate. So, unless there is some magical breakthrough, we've already hit a plateau. This is very different than plugin instruments. Most of the issues with early plugins were that they had to work in real time with a reasonable amount of voices, so shortcuts were taken so CPUs didn't melt. CPUs got much faster, added more cores, so developers could take fewer shortcuts and get excellent results. This is very different than one AI is doing, which is throwing a sh!tton of off-site processing to make the possibly correct information happen.
That is a way you could look at it. I'd offer it's not exactly accurate overall and that there are other avenues of exploration. Wikipedia isn't exactly citable, I would throw into the mix.

I'm neither pro- or anti-AI. When it comes to technology in general, I'm interested in looking at how things function in their current condition and whether or not they can be applied usefully. For certain applications, I think AI has potential, for others, not particularly at this time. For certain use cases, careful engineering of conditions need to be specified and a protocol followed to extract useful results, naive prompting won't provide satisfying results.

But in any case, I get the sense that the current mood of the forum is "AI is sh** without exception" so I doubt an interesting or useful conversation is possible at the moment. In which case, another thread for me to ignore.

Post

KBSoundSmith wrote: Sun Mar 08, 2026 5:40 pm But in any case, I get the sense that the current mood of the forum is "AI is sh** without exception" so I doubt an interesting or useful conversation is possible at the moment. In which case, another thread for me to ignore.
Then maybe ChatGPT can read it for you and provide a summary, because that’s absolutely not what is being said. I literally provided exceptions of when AI can be useful.

I think what you’re describing is that people are already sick of all the happy-talk and hype. If it’s so great, why is it being shoved down our throats. Google started proof reading my emails without me asking if I wanted it to. Most of the suggestions seem to be based on removing any kind of creative language and going for the most bland and concise language. Other times, it suggests changes that change the meaning of my message.

But I am excited about other developments. The machine learning used to create Synthesizer V models is great. I hear that SWAM has improved its instruments using this technology. I’ve seen ACE Studio do a great job of turning a sung line into an orchestra. That would save me a lot of time having to put in articulation cues. It would have to be really good at interpreting my vocals, though, and editable.

But song writing? What is the point? If you can’t write a song, asking a machine to write a song is still you not writing a song. If you’re expecting to make money, don’t worry. There’s no way the music industry is doing to let you. There’s a few people they let into that club every few years, and chances are, it’s not going to be you. Why pay $20 a month for mediocre music that you can make with a prompt, when there are streaming services that give you access to amazing human made music for the same price?

The anger you see towards AI for music is that it seems to be mostly a fix for a nonexistent problem. I just had a great idea for a middle bit in an old song I’m thinking of recording. It was late and I was brushing my teeth and my wife was sleeping, so I didn’t bother to even record it as a voice memo, as is my usual method. Why? Because I am in infinite resource for song creation, and I know that I’ll either remember it or come up with something equally good, if not better, at another time. No monthly subscription necessary.
Zerocrossing Media

4th Law of Robotics: When turning evil, display a red indicator light. ~[ ●_● ]~

Post

For what it's worth, I never used the term "super-genius parrot." It was jancivil. I'm going to have to disagree with her on this matter, as I think a "super-genius parrot" could make sense of simple nested quotes.
jancivil wrote: Sat Nov 15, 2025 11:47 pm
zerocrossing wrote: Sun Nov 09, 2025 4:29 pm
chagzuki wrote: Sun Nov 09, 2025 1:37 pm
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.
It's a serious misconception, there's nothing real in support of it. Now, if I see something of the subject on Youtube, the algorithm is pretty much trained to suggest critiques from a position of some knowledge if not expertise; but in general the layperson I'll encounter (I'm not doing social media in the sense usually meant by the term), to a person, has accepted seemingly with no examination whatsoever that there is indeed a form or forms of technology in the world that match the name Artificial Intelligence.

I think yours is the first comment I've seen in this kind of context that states the plain truth of it. There is no such thing. It doesn't think. It's a well known issue among the key drivers of the thing that it cannot reason. I'm not saying it's not great at it, it does not and cannot do it.
When this word intelligence is used for it, the word has been corrupted to suit the hype.
I've given a fair bit of thought over a lot of years to 'what is intelligence'. My friend at school and I would sometimes interview people to see if we could pinpoint where an opinion or a notion started with a person. In the end this winds up as virtually always derivative of someone else's "thought". Now, one can appear to be for all practical purposes a genius by very clever parroting. "AI" is a super-genius parrot.

I take my meaning from Roger Penrose: intelligence is a part of consciousness; consciousness is not computational. The seeming goal is always going to be out of reach from the current state of affairs.
I think one of the reasons this misconception has persisted is that I've heard some people, including one of the guys who started developing neural networks way back in the 80s, say things like, "AI comes up with answers in a very similar way to how humans do." What I think people are leaving out is that, like you've mentioned, consciousness isn't just a query-answer process. I'm reminded of a story where an AI started answering, "I'll get back to you with that tomorrow/later," when asked to perform a task, because it saw that response in human interactions. The problem was, AI has no idea what "later" or "tomorrow" is. It never ended up responding.
Zerocrossing Media

4th Law of Robotics: When turning evil, display a red indicator light. ~[ ●_● ]~

Post

Stochastic Parrot seems to describe a parrot in a more figurative sense than literal since the LLM has no innate sense of anything, including mimicry of spoken syllables; though the Generative music model's matching patterns in audio might be somewhat similar to the literal parrot's ability to imitate; which, like birds and their capacity for birdsong, is not terrifically well understood. Also some birds have better chops at least partly owing to their 'upbringing', confer 'training'.

Post

One day at U of Az I smoked a joint and sat in the commons on a bench listening to a host of birds, it got to be a jam session when a mockingbird joined the fray doing all the various bird's licks, which caused them to be kind of hyperactive. it was fantastic, total surround sound immersion from trees all around and this virtuoso mockingbird on top of everone, harmolodics galore. So it seems like talent is a factor, and interest. Who knows from bird brains really, though. But it has been observed that some have better chops at their particular song

One thing I do wonder about is that the training as pertains to particularly later iterations of LLM imprinting personality traits of programmers on the thing. It's as though language itself has super powers.

Post

I'm not against the technology in any general sense and have said so. I think what the music 'generating' model does is amazing. It's just not having ideas, it's a hard fact that it has no awareness or genuine knowledge, and LLM in fact can at best deduce, it has zero capacity at this writing for inductive reasoning. It's a trick, a highly impressive trick and a tremendous achievement but there is no mind there or anything close to it.
And if you want it to be a calculator, stick to the one in your protected pocket or in Applications, it's not meant for that either.

Post

KBSoundSmith wrote: Sun Mar 08, 2026 5:40 pm another thread for me to ignore.
So you do you. sayonara

Post

zerocrossing wrote: Sun Mar 08, 2026 8:24 pmIf it’s so great, why is it being shoved down our throats.
Is it being shoved down our throats? I'm not seeing that at all, anywhere. OK, Co-Pilot is now a feature in Windows and Edge but it is in no way intrusive, it's just there if you want to interact with it. I mostly don't and it's as easy to ignore as any other feature I don't need.
Google started proof reading my emails without me asking if I wanted it to.
And people wonder why I hate Google and go out of my way to avoid it. But if you want an example of something being shoved down our throats, Google are definitely the poster-child for that. We use Google at work and they aren't shoving Gemini down my throat there any more than they try to shove all their other shit down my throat. It's just what Google do and they can f**k right off with it.
But song writing? What is the point? If you can’t write a song, asking a machine to write a song is still you not writing a song.
Why does it need to be me? I realised a while ago that I get at least as big a kick out of performing a cover as I do from performing my own work. More importantly, I feel the same sense of ownership over the finished piece. It's not an Ultravox song, it's my cover of an Ultravox song. When I take something like a Buzzcocks song, something originally done with drums and guitars, and turn it into something purely electronic that works, there's a much bigger sense of achievement than writing something of my own, which seems trivial by comparison.

Where a song comes from doesn't matter in the slightest to me, having songs to perform and listen to does. So just as I buy new music to listen to on a regular basis, I like to have new songs to perform, too. The question for me isn't why would I use AI, it's why the hell wouldn't I? And the answer is that I can't think of a single reason. Maybe your sense of self-worth is tied up in such petty things, mine is not. A good song is a good song, how it came into being is less than meaningless to me.
I’ve seen ACE Studio do a great job of turning a sung line into an orchestra. That would save me a lot of time having to put in articulation cues.
But, in the end, you'll claim the credit for the finished piece, even though an AI will do all of the work.
I know that I’ll either remember it or come up with something equally good, if not better, at another time.
Because your brain has been trained on the millions of pieces of music you've been exposed to over the course of your life. Your subconscious has let something you heard as a child bubble to the surface to make a connection with some old piece you had worked on in the past. You think that's you being uniquely creative but it's just the same shit AI is doing.
NOVAkILL : Legion GO, AMD Z1x, 16GB RAM, Win11 | Audient EVO 8 | Lumi Keys | Studio Pro 8
Korg Odyssey, bx-oberhausen, Proxima, PolyMax, GR8, JP6K, Union, Atomika,
Invader 2, Flow Motion, Olga, TRK 01, Thorn, Spire, VG Iron

Post

There is one thing that pisses me off, and that is people posturing and making claims that are simply not supportable and are fairly easy to debunk given a few minutes if you can read. It's not different than stupid trolls that believe lies and are, for me at least, inexplicably driven to spread them; going for whacky conspiracy theories where that stands in for objective reality, checking for a source is just too big of an ask, no can do.
As per AI as discussed around here, some of this sort are too smart to be this ignorant, some are demonstrably not too swift on the uptake.

Post

zerocrossing wrote: Sun Mar 08, 2026 8:24 pm I think what you’re describing is that people are already sick of all the happy-talk and hype. If it’s so great, why is it being shoved down our throats.
It has to be pushed because it is not as obviously useful as say smartphones were when they first arrived. Also, there is a lot of social resistance to AI, both the concept and the implementation (huge resource use of electricity and water for data centers driving up prices.).

That plus AI not yet being a money maker. OpenAI and Anthropic both lose money, and most companies that have adopted the use of AI have not yet found it to save them money, or produce money.

It's still being promoted more on potential benefits than actualized ones. Thus the hype.

Post

pdxindy wrote: Mon Mar 09, 2026 2:26 am
zerocrossing wrote: Sun Mar 08, 2026 8:24 pm I think what you’re describing is that people are already sick of all the happy-talk and hype. If it’s so great, why is it being shoved down our throats.
It has to be pushed because it is not as obviously useful as say smartphones were when they first arrived. Also, there is a lot of social resistance to AI, both the concept and the implementation (huge resource use of electricity and water for data centers driving up prices.).

That plus AI not yet being a money maker. OpenAI and Anthropic both lose money, and most companies that have adopted the use of AI have not yet found it to save them money, or produce money.

It's still being promoted more on potential benefits than actualized ones. Thus the hype.
Right. Sorry, that was a rhetorical question. I get it. My wife and I stood in line for the first iPhone. It was exciting. There was a giant line and restaurants were coming by with free food for the people in line. We got it home and loved it. Mostly because of the photo abilities, but we quickly saw all the applications. It helped get me employed doing UI for a bunch of iPhone games.

With AI... it's been a mixed bag. I think it sucks as a way to get answers on the web. I'm constantly finding out that it's wrong. I hear my wife complain about it hallucinating and when countered with the mistake, it shuts down and goes into a weird loop. That's not "intelligence," unless there's a bit of a blood clot in its gray matter. The image generation thing... I'm sure that's a fad that we'll look back at and laugh at, like the 3d Magic Eye posters of the 90s. Most of it looks so bad and noticeable. I tried to upscale an old McDonald's video with it and the one frame I stopped on looked amazing, but as soon as I ran the video, it fell apart. Could it be used for good? Sure. I'd love to see someone take the Simpsons and other old TV and upscale it and fill it in for 16:9, if they could make it look seamless and good. I know people who worked on film who spent hundreds of hours going frame by frame removing wires and whatnot. No one likes doing that drudgery. With music, I've heard some convincing stuff, but all pretty standard. The one time I tried one, I asked it do imitate a modern Adrian Belew song about sleeping cats, and the intro and beginning was good, but then it didn't really go anywhere. The lyrics were mostly word salad, which is fine in a song, but nothing like the feeling I get from a Thom Yorke lyric. I came away thinking, "why would I pay for this when the thing it trained on is a lot better, and about the same price per month?" Meanwhile, I'm told Sunos wants you to be liable for any copyright infringement, if it's found that its model created something based on a recognizable tune. I decided to use Synthesizer V for vocals, because I write tunes that are outside my vocal range, and even if I stay in my vocal range, my voice is kind of crappy. It's just for me, mainly, as no one cares about my tunes, so who am I hurting? I wasn't going to record the songs without it, and no one is offering to step in and collaborate, so I figured it was about on par with me using SWAM Cello in a song, or even a sample library based instrument. Why would anyone pay money to have a lame computer write a song based on a few sentences? I mean, maybe if you're already Taylor Swift and need new material and you have nothing. Sure, her music is already profoundly mediocre, so why not have it write a song for her. No one would notice and she could actually make money off it. But me? I'm lucky to get a modest salary for making some boss battle music and laser sounds. I bet those jobs are gone now.
Zerocrossing Media

4th Law of Robotics: When turning evil, display a red indicator light. ~[ ●_● ]~

Post

BONES wrote: Sun Mar 08, 2026 5:40 am
jancivil wrote: Sun Mar 08, 2026 2:42 am"I - iV - V - I has been a staple of western music for centuries... but Bartok wasnt a fan." and so we see a different functional paradigm. For reasons.
The thing is, if you don't know any of this shit, and it means less than nothing to me, then you aren't bound by any of it, you're free to use whatever notes you think work together in any given situation. It's instinctive, not constrained by rules and conventions. If it sounds good, it is good and that's all that matters.
Tell her BONES !!

She will do all this avante garde stuff, like this Bartok stuff, and Xenakis, but don't ask Grizzellda, he's prolly not "sophisticated enuff". :lol: :lol:

Not that I don't respect it, (modern exploratory type music), but really more like a fondness for power chordz...

Post

zerocrossing wrote: Mon Mar 09, 2026 5:20 am I hear my wife complain about it hallucinating and when countered with the mistake, it shuts down and goes into a weird loop.
No, that's projection of a human behavior. That's not what's happening. The model is biased/anchored to the text in the conversation. The behavior that you're describing is often attributed to detailed requests, e.g., expecting detailed information on some niche technical topic that would typically be answered by a stack overflow post, in the context of limited information in the training set. For example, partial code snippets, ambiguous error messages, version-unspecified dependencies, etc. The model fills gaps confidently, and the fill becomes load-bearing in the next turn.

Hence, the model isn't "resisting correction" due to anything like stubbornness. It's because the conversation history itself is in the context window and acts as a prior that the next token distribution is conditioned on. The longer and more consistent the prior text, the stronger the pull toward coherence with it. So "goes into a weird loop" is just high-autocorrelation in the output given a heavily anchored context, a sampling artifact, not a behavioral disposition. Now, I know that you and, I assume, your wife get that. However, your language betrays that you don't know how to act on it to get better information. The gap there is skill, not the model. The model is what it is.

The overconfidence can be a real problem. It's much stronger, at the moment, with OpenAI's models than it is with Anthropic's models. Some open source models handle this much better as well, e.g., GLM-4.7. However, the detail of why it's overconfident matters and can point to different solutions. Often the best way to deal with a biased chat, however, is to start a new chat with a very detailed prompt that combines the detail that you are now attuned to. One thing that you will see if you do this is that the very same model may now give you a completely different result because it's no longer biased by the full conversation.

The bottom line is this, however. The vast majority of the complaints that I see expressed here are people expecting what they believe that the hype promised and not learning enough about how models work to be effective with them. My background happens to be particularly useful in this area so they have been useful to me on a daily basis for some time now. The problem isn't the model, per se, it's what people expect from a model when it isn't necessarily a good fit for that problem.

They can be great at looking some things up, and terrible other times. A part of that is how you present the problem, and a part of that is vendor cost control and limits on models gathering live data. Learning what it's good at and what it's not is not that much different than learning the delta between a new employee and their resume.
Last edited by ghettosynth on Wed Mar 11, 2026 9:38 am, edited 1 time in total.

Post

ghettosynth wrote: Mon Mar 09, 2026 7:23 am
zerocrossing wrote: Mon Mar 09, 2026 5:20 am I hear my wife complain about it hallucinating and when countered with the mistake, it shuts down and goes into a weird loop.
No, that's projection of a human behavior. That's not what's happening. The model is biased/anchored to the text in the conversation. The behavior that you're describing is often attributed to detailed requests, e.g., expecting detailed information on some niche technical topic that would typically be answered by a stack overflow post, in the context of limited information in the training set. For example, partial code snippets, ambiguous error messages, version-unspecified dependencies, etc. The model fills gaps confidently, and the fill becomes load-bearing in the next turn.

Hence, the model isn't "resisting correction" due to anything like stubbornness. It's because the conversation history itself is in the context window and acts as a prior that the next token distribution is conditioned on. The longer and more consistent the prior text, the stronger the pull toward coherence with it. So "goes into a weird loop" is just high-autocorrelation in the output given a heavily anchored context, a sampling artifact, not a behavioral disposition. Now, I know that you and, I assume, your wife get that. However, your language betrays that you don't know how to act on it to get better information. The gap there is skill, not the model. The model is what it is.

The overconfidence can be a real problem. It's much stronger, at the moment, with OpenAI's models than it is with Anthropic's models. Some open source models handle this much better as well, e.g., GLM-4.7. However, the detail of why it's overconfident matters and can point to different solutions. Often the best way to deal with a biased chat, however, is to start a new chat with a very detailed prompt that combines the detail that you are now attuned to. One thing that you will see if you do this is that the very same model may now give you a completely different result because it's no longer biased by the full conversation.

The bottom line is this, however. The vast majority of the complaints that I see expressed here are people expecting what they believe that the hype promised and not learning enough about how models work to be effective with them. My background happens to be particularly useful in this area so they have been useful to me on a daily basis for some time now. The problem isn't the model, per se, it's what people expect from a model when it isn't necessarily a good fit for that problem.

They can be great at looking some things up, and terrible other times. A part of that is how you present the problem, and a part of that is vendor cost control and limits on models gathering live data. Learning what it's good at and what it's not is not that much different about learning the delta between a new employee and their resume.
:lol:

5 paragraphs justifying AI making sh*t up and blaming users for it. This is exactly why lots of people aren't impressed with AI.

Post

I just did a song called Disco Jesus.

Let's see AI come up with something like that. LMAO

https://soundcloud.com/steven-wagenheim/disco-jesus

(Vocals by Audimee)

Post Reply

Return to “Machine Learning and AI for Music Creation”