Google's advanced music generation model and two new AI experiments

I believe it will still be worth it to become musically trained; otherwise, you’d just be adding to the pile of worthless garbage on redundant storage somewhere. My only fear is how to preserve the idea with kids of how awesome and enlightening it is to learn to play a musical instrument, be able to master it, even hacking it, and maybe improving it. That’s why over time I’m appreciating more and more analog music producers like Vulfpeck, and their success might be a lesson to learn from, even though I don’t expect many to know them in this high-tech context.

That would be the equivalent of automating out the middle class. Surely the prodigies will survive, but everyone else's dollars will only result in autotuned musical frozen dinners.

I like your optimism. In my opinion, popularity is not a great selector of raw skill/talent/quality. In practice, it does select for greatness certainly much more than random noise, but it's also quite arbitrary too, it both misses a lot of quality and selects a lot of low-quality crap.

You could call it "taste", though it doesn't exactly matter what you call it. You know it when you see it. Every so often, someone sells me on why something is awesome and suddenly it 'clicks' for me in a way that it hasn't, and now I'm enjoying something on another level. Thanks to curation and filtering by picky people, you get to experience a lot of things that would be hard to discern on your own. In general, you still have to rely on your own untrained senses to judge things, so curation can be very helpful to find the best stuff. Sometimes you don't know why something is better, but that doesn't mean you can't tell.

So then, a problem with generative AI is that it will mainly be optimized to mimic the appealing end results and then probably also tuned to prefer things that people respond positively to. That's cool and it can definitely produce results that people like, but I think it misses a lot of dimensions of artistic endeavors. Stable Diffusion today isn't drawing; it's forming pictures out of noise that look like they were drawn. No reason, imo, to believe an AI model won't some day actually be drawing; it's just not what's going on today. But because Stable Diffusion skips all of the details, it does things that don't make stylistic sense at all, like putting vastly overly photo-realistic details/rendering onto a cartoon drawing and other such bizarre things. On one hand, this is probably something that can be fixed, but I also think that fixing it is patching around the fact that it's generating images rather than creating them the way a person would. (To me, diffusion models feel like they could be the "imagination" of some AI agent that draws; but, ultimately, I have no idea what the future will hold.) The other thing that's awkward about the way this works is that it's a bit deceptive: looking at the end results, you can see things that are extremely impressive at times, but no doubt part of the problem is that it's difficult for us to actually evaluate it. For any of these very impressive generative models, I think that a surprising ability of being able to produce novel things has been demonstrated in many cases, but the truth is that even some less novel things tend to still be hard. Kind of like how you can generate a pig with a cowboy hat in space using one of these generative models, but ask it to generate two people holding hands and you may be treated with man-made horrors beyond our comprehension. (Or maybe it will work, but either way, the failure mode of these models is very interesting.)

OK, so I droned on a bit there. But what I'm trying to say is, I think these generative AIs are basically all going to be tuned to be as appealing as possible to the average person. Because of that, they will short-circuit our dumb ape brains into thinking "Wow! This is surprisingly good!" even though they are often lacking much substance and a closer inspection may show plenty of issues and general strangeness.

I think that more advanced AI agents might some day overcome this on a fundamental level, by simply not skipping "the hard part" and going right to the end result. That would help limit a lot of weirdness that just fundamentally does not make sense. Even when that happens, though, I suspect the domain of AI stuff will wind up being more limited than it initially appears, as the generative models will prove to have unintuitive limitations, and as a result of that, we may see floods of very same-y garbage stuff.

I'd be glad to be very wrong here. In the past, the availability of powerful tools definitely did what you said: it raises the bar. I mean, seeing what kids are doing in Blender and Inkscape feels like proof of this. AI does change some variables in a way that is scary, though: it produces things effectively infinitely fast and infinitely voluminous compared to humans, and at least by my measure it often achieves superficially good results, and the flaws may not be apparent quickly and maybe not at all to the untrained senses.

So maybe we're in for a flood of same-y garbage. Or maybe not. Guess we'll see.

If you have music that's playing in your head, genuinely playing in your head, as in you can actually hear the physical melody then you should be able to trivially whistle it. So in that respect, this tool which aims to combine pitch detection with instrumentation synthesization would be a pretty big net win for you.

The not so subtle difference is that synthesizers and drum machines don't play by themselves, or when they do (arrangers) they're 100% recognizable from a real band. AI will change everything.