When I release a piece of music, I consider making the video for it to be just as interesting and worthwhile as the audio. We’ve come a long way from my original experiments with analog video synthesizers, and lately I’ve been busy constructing a new music video “studio” inside my computer. It’s unlocked something unexpected in me: pure joy. It’s like an infinite Lego set, and it’s become one of the most creatively fulfilling things I’ve ever done.

Because I have a tool like Claude at my disposal, I can make the studio as eccentric as I want. It’s not built for efficiency or cranking out a ton of content. Instead, it’s made for exploration and learning. It lets me be as weird as I want to be (and that’s pretty darn weird), and it keeps up with my brain, following any random ideas or rabbit holes I want to dig into.

The setup is simple: I won’t use any generative image, video, or music models for my work, and the rendering is done entirely on my solar-powered home computer. Claude writes the code for me, as well as dispatching a local model (Qwen 3.8) for various tasks, and we iterate back and forth on my ideas. It takes hours (again, by design), and I have a pretty high-powered Mac Studio doing the work to actually make everything, so this approach is decidedly not for everyone. If you’re interested in seeing this workflow in action, please check out my Zephatron Drift post.

Sometimes my rabbit holes lead nowhere, and sometimes they can be an absolute delight. On my recent track Honest Hearts I wanted to put in some spoken word because I love the voiceovers used in electronic music back in the late 80s and early 90s. Before I settled on anything, I had a conversation with Talkie, an LLM trained exclusively on content that cuts off in 1930. It’s one of the strangest and most wonderful experiences, chatting with something trained on a body of knowledge that predates so much of today’s world. I imagined myself in Madame de Staël’s salon, speaking about the nature of truth, virtue and beauty.

Me:

Whither a poet?

Talkie:

Whither a poet? To Parnassus. Thither let him repair, and ascend the steep diligently, cheered in his labour by so sweet a voice as was heard of old on the banks of Helicon. Let the music of his own lyre be the solace of his pilgrimage; and, when he has reached the summit, let him lie down and sleep in the shadow of the laurel. Let him wake again to sing. Let him sing purely and boldly, of the good and the beautiful, in Nature and in life. Let him be sensible of wrongs, but not embittered by them; let him be alive to griefs, but not overpowered by them. In the midst of evil, let him preserve faith in goodness; let him hope, amidst despondency, for a better future; let him love, amidst injustice, and be gentle, amidst contention. So, living and dying, let him contribute to the pleasure of honest hearts, and leave the world at last in quiet.

I mean, come on now! In that instant, I knew that I had to use it for my track. On my last album I used a quote from Thanatopsis, and this is a perfect complement to it. I used an ElevenLabs text-to-speech model to create a voiceover from Talkie’s words. Then it hit me: This is like the sampling they used to do back in the day, taking a longer body of spoken words and slicing it up to fit the music. The process may use new tools, but the creative technique is a classic. You can check out the result for yourself:

The question I have now is, “whose work is this?” For the music, I’m putting together samples that someone else created. For the words, I’m taking excerpts from an LLM’s output and using a speech synthesis model. For the video, I have Claude writing the code, which I couldn’t write myself. Can I consider this “mine,” then? It’s complicated, and that’s a big part of why I find it so interesting. I stay in the driver’s seat, making decisions down to a ridiculously minute level at times. I decide what to keep, what to toss, and how to sequence it. The looks and sounds I choose are all mine. I’m not allowing choices to be made for me, dumping full outputs into pieces, or making a music video factory. The work I’m creating with this studio simply could not be made by anyone but me.

I have things set up so that no video, image, or music is built using models that are trained on creatives’ work without their consent. I did use a generative voice model, but it’s from a creator who gave explicit consent for it. He verified that it’s his own voice, and gets paid every time it’s used. I render locally, on my own computer, as much as possible. That doesn’t absolve me from some of the more problematic aspects of using AI, but it is a much better fit than the way I had used it before. In Obscured, I heavily incorporated early generative art models, and while I’m proud of the results (and they took hours of work to create), I just don’t feel comfortable with this way of working anymore.

The difference, I think, is that now I use AI for the plumbing, and not the art itself. I have to think hard about what I want to do, and I struggle to find ways to articulate it successfully. I think articulation is a skill that will become increasingly important for getting the most out of our work with AI in the future. There’s a ton that I have to puzzle out, and I’m not always good at it. How much of the problem is my inability to express myself, and how much is from the model’s inability to understand me? I’ve discovered that friction is a natural byproduct of this way of doing things, and that’s the opposite of the cognitive-offloading trap that many of us are worried about when we engage with AI.

This came into focus when I wanted to try a visual style for another video that I didn’t know how to describe. It was 80s New Wave, but in a specific geometric style and color range. I had Claude build me a tool for developing a style vocabulary, where I discovered that the style that I was looking for was a cross between Swiss New Wave and early 80s album art. From there I learned more about wonderful artists like Wolfgang Weingart and April Greiman, among others. These aren’t diversions. They’re part of the process, taking the long way around to learn new things, and I wouldn’t have it any other way.

Following Along poster

When I shared this new way of working with some of my students, I thought that they might be interested in how I’ve adapted to mitigate the impact of AI on my creative work. I was wrong. None of my careful choices and long deliberations mattered. They firmly explained that AI has no place in art. To them, art is a uniquely — and exclusively — human endeavor and should be kept that way. The students also explained that it’s wrong for me to pay for these products for a host of reasons, including the fact that the companies say their product may end humanity, and yet they won’t stop building.

That’s a compelling perspective, and while I can respectfully disagree about the nature of art, I really can’t argue against the corporate support. How could I? My joy comes with some serious baggage.

In Robin Sloan’s recent follow-up to one of my favorite posts of his, “An app can be a home-cooked meal,” he asks:

And … even if you were an enemy of the railroad corporations, the vast syndicates … would you deny yourself the pleasure of the train?

For now, I’m choosing to ride this train. If it no longer feels like play, if the companies choose to follow social media’s trajectory, profiling us from the content of our chats, if they make it so that only rich people can access their tools, or if they keep crying “rogue agents!” when bad things happen instead of owning the fact that their product was to blame, I’m out. Until then, I’m going to keep creating and exploring, and I’m going to keep doing my best to share what I learn so that together we can all keep figuring out this surreal new frontier.