Rick Rubin on AI (& now Graeme Revell, too)

[ghettosynth]

On the “is it AI?” I can see how someone might think it was hand-done with samples but the zippery/phase vocodey artifacts on the trills stand out like a sore thumb.

Yes, and, that shows up more, or, more correctly, is more obvious, in specific contexts. For the uninitiated, you can do these experiments yourself without cost. The site will give you 1min of the output from the current best model.

Anyway, does it always matter? Well, I recall the same kind of discussion about MP3 artifacts. I have also seen this forum completely take the other side of this argument with tools that they want to use, e.g., older synths with very weak filters. I recall criticizing the artifacts in GRM tools and, oh my, you would have thought that I had bad-mouthed the pope.

I still can recall stopping dancing mid track because of bad filter artifacts/distortion in the late 90s and watching while everyone kept dancing because they didn't notice. In this story, this forum is me, everyone else is, well, everyone else. It's the same reason that the general public doesn't care much about auto-tune, it's the modern equivalent of "It has a beat, and I can dance to it."

The artifacts are less obvious in more restrained vocal styles, if, you can also keep the other elements out of the problem areas. Of course, none of this matters much for the vast majority of people using the tools to just have fun.

For those that don't know, if you have a decent video card with at least 16GB of VRAM. Heartmula is a (free to you) project that lets you generate music on your own machine with reasonable results. It is not up the same level as the commercial offerings.

https://heartmula.github.io/

To my ears, this sounds like some of the older Suno models. Some people think that they are gaming their outputs for the benchmarks, I frankly don't care about benchmarks all that much.

What is interesting to me about this is how good the results are with respect to how large, really small, the model is. Compare this to LLMs and it's rather dramatic. Fine tuning this model, on any source, can be done privately with reasonable (but not typically consumer level) hardware. For more specific workflows, e.g., very narrow generative tasks, this is at the level of a completely personal project in 2026.

In about five years time, you will all be arguing over which AI plugin does this one narrow specific thing that matters to you the best. Vocal trills may still have artifacts in commercial models.

[revvy]

I still can recall stopping dancing mid track because of bad filter artifacts/distortion in the late 90s

[jancivil]

Here's an interesting quote from Graeme Revell, film composer and founding member of SPK - "The cultural meme of machine – logical, emotionless; human – creative, emotional, is ingrained in our civilization. It has been reinforced in popular culture, but it also goes back way further to our delusional sense of exceptionalism in the biological kingdom. Many people cling to this illusion and are unwilling or unable to explore AI with an open mind." This perfectly describes all the nay-sayers here, doesn't it?

It forms a kind of strawman from a full-on non sequitur (which also contains a strawman), the first part of which supposes the machine as something more than a machine. The accurate perception of a machine lacking emotion etc goes back to what? Anyway I don't have that sense.

The fallacy of the open mind occurs when excessive receptivity to new ideas leads to accepting false or illogical claims, often resulting in gullibility.

I know for a fact "AI" doesn't think, let alone feel. Therefore I am not open to considerations based in it having these capabilities.

You tend to show a marked incapacity to think things through here. This is a problem of people that need to believe they're always right.

[Bunny_boy]

I still can recall stopping dancing mid track because of bad filter artifacts/distortion in the late 90s

Have you really not danced since then?!?!?

[vurt]

I still can recall stopping dancing mid track because of bad filter artifacts/distortion in the late 90s

Have you really not danced since then?!?!?

guilty feet, have no rhythm.
thats all im saying on this matter.

[VOODOO U]

Welp. While you guys are all mumbo jumbo, i now have 6 songs done sans vocals (gotta get one to do them). And that's what....roughly about 5 days of f**king around with Tunee and about 1 to 2 hours a day of prompting. Recording and piecing together the bits i like to get the song obviously takes longer but point being that fully writen kick ass songs got done fairly quickly.

Turns out Boner was right. This shit works. You just have to be patient, persistent and have good production skills to get past the formulaic generic stuff A.I. spews out.

[chagzuki]

While you guys are all mumbo jumbo, i now have 6 songs done sans vocals (gotta get one to do them).

Only 6 songs? You should have 5 albums by now.

[ghettosynth]

I still can recall stopping dancing mid track because of bad filter artifacts/distortion in the late 90s

Have you really not danced since then?!?!?

No, I'm sure death is just around the corner so I'm saving myself for the moment.

[Bunny_boy]

I still can recall stopping dancing mid track because of bad filter artifacts/distortion in the late 90s

Have you really not danced since then?!?!?

No, I'm sure death is just around the corner so I'm saving myself for the moment.

Why not combine the two?
https://en.wikipedia.org/wiki/Dancing_mania

[VOODOO U]

While you guys are all mumbo jumbo, i now have 6 songs done sans vocals (gotta get one to do them).

Only 6 songs? You should have 5 albums by now.

If i just did prompts then I would easily have 5 albums. 5 albums with loaded shit. I don't see why anyone would just go with the direct result of a prompt. That's like using presets in a synth.
No offense if you use presets by the way. I can't help that I'm loaded with pristine talent where i can make my own sounds.

[BONES]

It seems though that it's going to be almost impossible for copyright law to cope with the complexity of hybrid AI art production lines.

I find this situation ridiculous. If a human trains at a particular school, that school doesn't have any copyright claims over anything that student does after graduation so why should AI be held to a different standard.

If I listen to a song and it inspires me to write a song in a similar style, I don't have to pay a royalty to the artist who inspired me, unless I have used the exact same melody or something like that, so why shouldn't AI be able to use its training to create a unique song in the style of another artist, as long as its not a direct copy? Obviously if it's using a simulation of the original artist's voice, that's a different issue but if it's not readily identifiable, where's the problem? It's not like half the top selling pop artists don't all sound the f**king same, is it?

Even the idea that AI companies should pay for the material the AIs train on is ridiculous. If it's freely available on the open internet, then it should be fair game. There should only be an issue if there is evidence that a firewall may have been hacked to provide access to specific material that's not freely available, or similar illegal activity. But if a search engine can find it, it's fair game in my book.

Incorrect: https://blog.bandcamp.com/2026/01/13/ke ... amp-human/

That's just to shut idiots like you up. Until they tell the artists about it, it's not going to have an actual effect. BC are very good at communicating policy and terms of use changes so if they were serious about this, I'm sure they'd have let us know by now but, so far, they haven't told us anything.

The obvious problem with a policy like this is that artists who use AI will simply stop telling anyone they're using AI. There's a band on BC, who've been around for a while, who have clearly been using the same AI as us for their new album, probably using very similar prompts, although with clearly less effort or skill going into it. The result is stuff that sounds a lot like what we're currently doing, only it lacks the variety that we've got and the songs are a lot less interesting. But it's not terrible and it definitely deserves to be heard. I bought the album the other day and I've listened to it a few times already. I'd be quite pissed off if BC took it down so I have no intention of telling BC or anyone else who the band is or what they've done.

One of the problems with the way the current generation of AI works is that it just doesn't work that way. It's case of "best I can do is some Mignola mixed in with generic DeviantArt slop and all the cats have five legs".

That was a problem 6 months ago but I don't see too much of any more. AI is just getting started, it's only a matter time until it is indistinguishable from anything a human can do.

I do it because there‘s no one around to play with. Ones and zeroes don‘t cut it for me.

And yet here you are. I've never had the slightest desire to work with other people. I only started working with Craig on NOVAkILL because I wanted the band name. But I'd be surprised if there was another human being on the planet I could work with. He does his thing, I do my thing and occasionally we get together to see if we're still on the same page. (We always are.)

[eassae]

That's just to shut idiots like you up. Until they tell the artists about it, it's not going to have an actual effect. BC are very good at communicating policy and terms of use changes so if they were serious about this, I'm sure they'd have let us know by now but, so far, they haven't told us anything.

[eassae]

It does worry me that Bandcamp says they are going to have a way to report AI music by users. It seems there are fairly accurate detection tools that they could implement on song upload. Community policing could turn into a tool of retribution. If they do put the report system in place hopefully they have a decent review policy and at least give the accused a chance to respond.

[BONES]

The two that are in here arguing for "AI", one rather blandly, the other being the aggro bully he frequently is, are both utterly f**king clueless about the process in any fullness of creating music.

Again, this is provably wrong. I've got 40+ years experience and 9 albums worth of proof that I am anything but clueless about the process. It's complete nonsense to suggest otherwise and you just look like a complete f**king idiot in doing so.

AI will fit the bill if what you do is cut-and-dried manufactured unoriginal pablum.

Works for me, I have no ego attached to my music. It doesn't need to be so terribly clever that the YouTube views are in single digits or that you need to have spent half your life training to enjoy it. Quite the opposite, the less of that that goes into it, the more honest it is. The fact someone like you hates it just makes it that much better to me. That's what's great about working with AI - because you have to be so clear and details in stating your intentions up-front, what you end up with is much more pure. It expresses what we want to express far better than we've ever been able to in the past. Sure, that's because we've never been intent on expressing anything - finding lines that rhyme has always mattered more than what they were actually saying - but now we can do both thing better than ever.

I don't set out to make some generalistic generic music where there is a kind of musical line, riff, lick what-have-you fulfilling a cookie-cutter form - eg., the same chord progressions 18 million songs use, reitrerated 4 or 8 or 16 times etc without exception - where one example of adequately filling this container is as good as the next.

Why, do you find it too hard to make it into something of your own? It's certainly a lot harder to do than the undifferentiated mess of yours I was listening to the other day. Seriously that stuff sounds like it was put together in five minutes, from whatever bits and pieces were to hand.

My ideas are uniquely my own

No argument there, I doubt anyone else would see them as ideas at all.

When there are cases I want something other than what I, too close to the material maybe or just to get another angle on, would do - mainly to be surprised, to get something fresh

Sounds like the perfect case for using AI to me.

I don't specify beyond a generalization. I'm looking for high intelligence. "AI" is_not intelligence.

You think? I bet it's a f**k-ton smarter than you are. How'd you go on the MENSA entrance exam?

It can_not think, it has no capacity to work through a problem

Neither does a piano but that doesn't stop anyone making music with one, does it? It's the same nonsense argument you've been peddling for days. It makes no sense. AI doesn't have to think, it doesn't need to have the capacity to work through a problem, that's still your job as the person inputting the prompts.

Again its intent is locate prevalence and produce copies recombined....

... based on the direction of the human inputting instructions (prompts).

If it produces a different simulacra the next time, this isn't because it had different thoughts, it's a product of its programming. It may as well be random factor.

How is that not a perfect way to iterate ideas? It's like probability factors in sequencers when you are making generative music - it creates variations based on parameters you set.

So someone insists 'you won't even try it', I don't know what you want from me. The pseudo-Bach shows it, the machine doesn't understand Bach

Or the person who input the prompt did a shit job, probably deliberately. You can't know which it is, you're just making assumptions that fit with your biases. Of course, if it was perfect, you'd just be complaining that it was too rigid, too perfect so your reaction is irrelevant anyway.

You won't be able to help it, it doesn't f**king understand.

Of course it f**king does, you just have to do your part of the job properly and tell it exactly what you want.

If a student handed that in from an assignment, from measure 1 its grade is F.

As would be appropriate given how little time and effort they put into it. But, as I've said before, AI is just getting started. In a year or two it will be exponentially better than it is now. You pointing out shortcomings is probably just aiding in the process.

I remember in rehearsal he came up with this killer complex time signature idea; he wanted me to sort it out but once he had to think about it once I'd done, he lost his own idea.

Another perfect problem to solve using AI - let it do the grunt work so you don't lose your train of thought.

I know for a fact "AI" doesn't think, let alone feel. Therefore I am not open to considerations based in it having these capabilities.

It doesn't need to, it just needs to know what you want when you ask for it. It's your same bullshit, nonsensical argument.

You tend to show a marked incapacity to think things through here. This is a problem of people that need to believe they're always right.

Which you are manifesting like you wouldn't believe. Using lots of big, serious sounding words in a vain attempt to lend weight to your bullshit doesn't make it stink any less.

Seriously, though, who is more likely to know what they are talking about, someone who has spent months actually doing the thing or someone who is adamant they never will so much as try it?

[BONES]

With me playing a rhythm guitar and my partner laying down the drums, there's something there that a machine can't formulate.

I'm pretty sure that's mostly in your mind. It might be more fulfilling when you're doing it but when you're done, the output, no-one else is going to notice. For the vast majority of listeners, me included, that stuff is irrelevant. If you have to delve into the minutiae like that to find something of value in a piece of music, then it's probably a shit piece of music. If it's a good song, none of that will matter. It's a visceral thing - it hits you or it doesn't.

I still can recall stopping dancing mid track because of bad filter artifacts/distortion in the late 90s and watching while everyone kept dancing because they didn't notice.

Yeah, that's because they were into the music and you weren't. I feel kinda sorry for you if that's all it takes to throw you out of the moment.

What is interesting to me about this is how good the results are with respect to how large, really small, the model is.

How does it work on less well known genres? Because that was initially the problem with things like Suno, it simply didn't know what EBM was so it took the rest of the prompt and made heavy metal from it, because that was something it understood that fit with the rest of our goals.