ai products are boring
a short post
Back when Midjourney was first made available for anybody who wanted to sign up to use it, my experience—as somebody who didn’t try it out, because I was too lazy to make an account—was that everybody was suddenly posting slews and slews of images with “wow, check this out.” Suddenly anybody could make a picture of a circus seal in the style of Van Gogh, or whatever, just by typing in the prompts. All these people you knew suddenly posting pictures, and then variations on those pictures, all of which they found really interesting and / or hysterically funny.
But… were those pictures very interesting to you?
My guess is: not really. Even if you were generating your own images. What was actually fun (if you were having fun) was not the end product, but the process of prompting itself. Divorced from this process, nobody’s dump of their interactions with generative AI is all that interesting. Every conversation people post between themselves and ChatGPT has a similar problem. I believe that these conversations are, to the people who post them, interesting. But they don’t say much to me.
The same thing happened to me when I was playing around with DeepSeek. Even though I knew, from my own experience, that there was nothing interesting here to some other person, I still felt like I should share its attempts at H.P. Lovecraft prose.
Two things brought this topic to mind. A month or so ago, I was excited to discover that Speak For Yourself, the 2005 album by Imogen Heap, was getting a twentieth anniversary edition. I ordered a copy and kept an eye out for any other Heap news. And she released an EP of new music. Hooray! Except.…
The EP is mostly a “collaboration” between Heap and (from what I understand) an AI called “ai.mogen” she’s been training on her voice for years.1 Which is in character for her. She’s always been interested in incorporating new technology in her work. So, if Heap wants to experiment with AI trained on her own material, I’m not going to stamp my feet about it. It’s just that I don’t really want to listen, either.
I did listen, because I’m writing about it here, and I respect her as an artist. And it’s fine, but for me, there’s something very pointless about having something that sounds like Imogen Heap on an Imogen Heap track, even if getting the AI to that point took years of labor. If you are not interested in that aspect, I’m not sure what you can get out of the composition. As far as I can tell, through years of effort, Heap made a Vocaloid of herself. But unlike actual Vocaloid compositions, which can lean into the artificiality of the instrument,2 this one just sounds like… Imogen Heap. And I already have something that sounds like Imogen Heap. (It’s Imogen Heap.) So.…
Do I believe that she could do something interesting with ai.mogen eventually? I do, actually. But in this particular composition it comes off as a gimmick.
Less likely to produce anything good is the approach recently announce by Bayreuth, which will be incorporating an AI trained on their past productions into their 150th anniversary staging of the Ring:
For the first time in the history of the Festival, Artificial Intelligence will take part on stage—not as a character, but as an image-generating force.…
The singers are at the heart of the performance, in a calm, almost sculptural presence. Their bodies become the fixed point within a visually seething cosmos of light, texture, history, and association, amidst projections that erupt, constantly shift, and merge into one another. The projections are more than stage design—they are a reflective surface of a 150-year discourse. The AI that generates them has learned from countless images, voices, documents, and productions. It does not present a single “Ring,” but many: the national myth, the socio-political upheaval, the artistic explosive force, the romantic utopia, the deconstructed shadow. Every performance will be unique—because the images and associations never stand still.
Again, I can’t really imagine the appeal of this as a spectator. I assume that this AI, like Heap’s, has been trained only on data that Bayreuth either owns itself or has paid for. But if I go to see a production of an opera, I don’t particularly want to see an amalgamated product of every past production of that opera. I want to see the vision of particular people executed on the stage. If I wanted to experience past productions, I could get them in some other form.
I can see why an artist might find this sort of process useful to try, the same way feeding your work into something that generates a word cloud can be interesting. You see associations and concerns that you might not have seen otherwise.
But I also think that this process seems narcissistic in the most literal of senses—it’s about staring at a mirror of yourself and feeling like you’re making something new. I enjoy staring at my own reflection as much as the next vain person, but I don’t expect somebody else to get much out of looking at my reflection. Lots of great art will be, and has been, produced through means and out of motives that are vain, self-involved, and solipsistic. But this kind of mixing up your own past material and sticking it into the microwave is a different thing. I find myself dubious it is going to produce much of interest to any third party.
Somebody reading this might be thinking that I’m creating something of a “no win” scenario, in which Heap’s ai.mogen isn’t interesting because its contribution is so minimal, whereas Bayreuth’s use of an AI set isn’t interesting because its contribution is too big. But really what I think, or suspect, is that generative AI is interesting as a process but not as an end result. Somebody who uses it in their art successfully will figure out how to showcase this.
As far as I can tell, the AI has no role in composing any music here: “The song as well as this music video were created by dozens of human artists over the past few years.”
I think I’ve linked to this before… possibly in this exact context… but:
Vocaloid covers of things actually composed for the human voice, however, are another story:
(Maybe it’s gotten better.)
I’ve always felt like reading or viewing AI products is like listening to someone else’s dream. It’s inherently interesting to the person who produced it, but is largely meaningless to someone else.
I think Holly Herndon’s Proto album is probably the best and most interesting use of AI in music I’ve come across yet - and followed a similar path to Heap, where she spent years training her own algorithms. That said, she hasn’t released one since and I wonder if there are similarities with modular synthesis where a lot of artists get so lost in the process and possibilities, they more or less stop creating a finished article - AI as an anti-productivity tool.