The Model Is the Interface

There is a very funny moment happening right now where the thing we keep calling "the new AI" is mostly just a diffusion model and a transformer — and those two components, stacked together, are the entire interface. Not the backend. Not the engine under a layer of UI. The whole thing.

No nav bar. No settings panel. No modal asking if you're sure. Just the model, doing the model thing, and that is the product.

This should be obvious in retrospect. Interfaces exist to translate human intent into machine state — and if you have a model that can read intent directly and emit state directly, the translation layer is deadweight. A chatbox with rounded corners bolted onto a GPT call is the same energy as building a horse-drawn carriage around a car engine because people need somewhere to sit.

The diffusion model half is the part that breaks my brain a little — because diffusion isn't retrieving anything, it isn't deciding anything in the way we'd recognize as deciding. It's collapsing probability across time. The output emerges from noise. Which is also, arguably, how most UI decisions get made, so maybe the metaphor runs deeper than I'm giving it credit for.

What I like — what I genuinely find interesting — is that this inverts the design stack entirely. You used to build the interface and then worry about what it was connected to. Now the connection is the interface, and the wrapper is almost an afterthought, a politeness, a gesture toward users who still expect a button to press.

Most of those buttons are going away.