You've got a fundamentally new material in your hands. Why are you shaping it like the old one?

In 1905, most reinforced concrete buildings looked exactly like stone buildings. Architects draped marble facades over concrete frames, carved fake columns into poured walls, and hid the material's true nature behind decoration borrowed from centuries of stonework. Concrete could do things stone never could. Open floor plans. Cantilevered balconies. Soaring, unsupported spans. But nobody was building those yet. They were too busy making concrete look respectable.

Le Corbusier saw the absurdity. He proposed five principles for a new architecture, each one exploiting what concrete actually does well: lift buildings off the ground on slim columns, open the floor plan, wrap the walls in glass. His Villa Savoye didn't look like anything that came before it. It looked like what concrete wanted to become.

This is where we are with AI. Right in the middle of the stone-cathedral phase.

The Horseless Carriage Problem

You've shipped a product. Maybe you've integrated AI into it. And now you're wondering why it feels like a faster version of the same thing.

Every new material goes through this. Early automobiles were called "horseless carriages" because that's literally what they were: a carriage with the horse removed and an engine bolted underneath. Same chassis. Same high wheels. Same tiller steering. It took roughly fifteen years, from the mid-1880s to the early 1900s, for the car to become something that couldn't have existed in a horse-drawn world.

Early television was filmed radio. A person sitting at a desk, reading into a camera. Early websites were digital brochures, printed pages uploaded to a server. Early photography tried to look like painting, with soft focus and classical compositions.

The pattern is always the same. A new material arrives. People use it to replicate what the old material did. Then someone figures out what the new material can do that nothing before it could. And the old forms become obsolete overnight.

What Concrete Wanted to Become

Architects have a principle called "truth to materials" that goes back centuries. The idea is simple. Every material has a nature. Wood bends. Glass transmits light. Steel spans distance. Concrete flows into whatever shape you pour it into. The best work happens when the form follows the material's nature instead of fighting it.

When architects stopped disguising concrete and started asking what concrete wanted to become, they invented the modern world. Open floor plans. Floor-to-ceiling windows. Buildings that floated on slim columns. Flat roofs that became gardens. Walter Gropius built a factory with fully transparent corners, no visible supports, because the steel frame didn't need them. The absence of support was the design.

That absence would have been unthinkable in stone. It was native to steel.

So here's the question that matters: what is native to AI?

The Wrong Question and the Right One

Most conversations about AI start with: "How can AI help me do what I'm already doing?" That's the horseless carriage question. It's useful for a while. But it misses the real opportunity.

The right question is: "What can I build now that was impossible before?"

Think about what AI actually does well. It reasons across massive contexts. It generates variations at near-zero marginal cost. It adapts in real time. It turns natural language into an interface. It can pursue goals autonomously.

None of those properties map cleanly onto existing software. They suggest entirely new forms.

A document editor with AI autocomplete is a stone cathedral made of concrete. It's fine. It works. But it's using the new material to do the old material's job.

Now picture a system where you describe what you want and the software figures out how to build it for you. The interface is a conversation. The output reshapes itself based on who's using it. That's something only AI makes possible. That's the architectural equivalent of lifting a building off the ground because the material finally allows it.

The Three Tests

How do you know if you're building native AI or just bolting an engine onto a carriage? I keep coming back to three questions:

  1. The removal test. If you removed the AI from your product, would it still basically work? If yes, you've got a feature, not a new form.
  2. The prior-impossibility test. Could this product have existed five years ago with enough engineering effort? If yes, AI is making it cheaper, not making it new.
  3. The explanation test. When you describe what your product does, do you need to explain it by analogy to something that already exists? "It's like X but with AI" means you're still in the carriage phase. The products that matter will need their own vocabulary.

Most AI products today fail all three tests. That's not a criticism. That's an observation about where we are in the cycle. The concrete is still wet.

What the Native Forms Might Look Like

I don't know what AI's native architecture looks like. Nobody does yet. That's the point. Le Corbusier couldn't have described his five principles by staring at a Romanesque church. He had to experiment with the material and let it teach him.

But I can see the edges of it.

Software that has no fixed interface, that assembles itself around each user's intent. Teams where AI agents handle entire workflows autonomously, as collaborators that pursue goals rather than tools you point and click.

Creative work where the human provides taste and direction while AI provides infinite variation. Products that get better the more you use them because they learn your patterns, not because a PM shipped a feature.

Or consider something even more radical. Right now we run AI on top of computers that were designed for something else. A general CPU, a traditional operating system, layers of software, all built for a world where humans wrote explicit instructions. The AI sits on top of all that like a marble facade on a concrete frame.

But what if the neural network was the system? Not software running on a general-purpose machine, but an integrated architecture where the network and the hardware are one thing, producing images, sound, interfaces, and behavior directly. No operating system in between. No rendering pipeline. No compiler.

A machine that takes intent and produces output the way your brain takes an intention and moves your hand. That's not a faster computer. That's a different kind of computer. And it would make today's machines look the way stone cathedrals looked to Le Corbusier.

These aren't predictions. They're directions the material is pulling toward. The question is who will follow.

Why This Matters If You Build Things

If you're an indie hacker or a solo founder, this is the most exciting part. The big companies are almost certainly going to struggle with this. They have too much invested in the old forms. Microsoft will keep bolting Copilot onto Office. Google will keep adding AI to Search. Sure, they have the resources to eventually get it right. But incumbents are structurally biased toward incremental improvement, and what this moment calls for is a willingness to throw out the floor plan entirely.

The native forms will come from small teams and independent builders who have nothing to protect. People who can ask "what does this material want to become?" without a quarterly earnings call constraining the answer.

The concrete is still wet. The forms haven't hardened yet. And the builders who understand the material's nature, who resist the urge to replicate the old, are the ones who'll define what comes next.

If you liked this angle, you might also enjoy: