People tend to use things in the way they think they’re supposed to.

It’s why film cameras inspire slowness, intentionality, and an affinity for specialty coffee.

In fact, it’s the principle of marketing as was so well shown by Nathan Fielder’s sale of the DoinkIt , that people buy products to fit an identity.

If marketing sells products by associating an identity, and people use products roughly how they are told to, what identity is AI connected to and how are people told to use it?

When I say AI here, I don’t mean the amazing and niche world of ComfyUI, Colabs, and Ollama. I mean the billion dollar top-of-appstore apps and the flagship surfaces for LLMs & generative media.

Most of these tools are pitched as time savers, job doers, and work skippers. The key here is skippers not accelerators.

An ad I see in the daily in New York’s subway roughly goes

🚉 Our product: data → drink coffee → slide deck
Your life: data→ draft → revision → boss yells at you → … → slide deck

the idea is that your input, your direction, and your ideas are of little value. The model is smart, the model is capable, it’s basically magic, just trust it to do the thing please.

In my field of generative media, I see the same sentiment. I feel as though many creative tools powered by AI went and turned every knob into a black box without asking which knobs are a chore and which are a creative tool, and without giving users a way to graduate from auto mode to manual as they play.

The promise is concept in, movie out. Concept in, photo out. It’s so reductive of the art forms that it is a miracle anyone consumes AI content.

<Insert image of Sora app’s massive failure here>

Ok maybe no miracle. No one likes slop.

It’s not an issue of image quality or performance either. It’s a UX problem through and through.

ComfyUI is not a tool for making slop. It gives you the knobs and the tools and exposes these models for what they are, which I am confident will grow to be a mainstream and accepted creative tool like the sequencer and synthesizer before it.

Legitimate fantastic art has been made with Stable diffusion (not even XL), or with hyperGAN, or with anything else you can think of. It could also be made with the latest closed proprietary models.

The difference is in how we tell the user to create with these models.

A text box promising to do the creativity for them and blow them away does not inspire art, it farms out slop. A UI that inspires exploration, iteration, play, and learning results in art.

LLMs are the same way, and we are learning to embrace this. Harnesses like Claude Code make us more intentional about how we structure tasks, how we prompt, and how we guide these models to being genuinely useful tools for coding. Art, writing, everything else is the same way and needs the same level of intentionality baked into the user experience.

Art cannot be automated. No image worth looking at will ever be achieved with one button, unless it is a subversive dadaist readymade which spits in the face of the slop factories.

If you spent the last 4 hours in ComfyUI or Flow making a piece of art, I will take my time to appreciate it and give it my attention. If you farted out some image of a landscape from a text to image model, I’ll wash my eyes out in hopes of unseeing your abomination. That’s true even if the pixels underlying the 2 images are identical.

Anthropic and OpenAI seem to be well aware of this. They market to creators, makers, and tinkerers. Some of their products reflect that design vision and some famously did not. I contend this is why claude code wins and sora goes to burn in a silicon pit somewhere under Google+ and Vine