I am one of four software engineers who use AI every day — my three co-founders at Osparna, and me. Linnet States leads our design and works in UX and JavaScript. Carl Fyffe is a Ruby on Rails engineer and a former Navy sailor. Matt McKnight founded and ran a software company before we started this. Every one of us can build anything we need with no help at all — and every one of us uses these tools constantly. If you want to understand what AI actually is right now, that room tells you more than almost anything being said about it in public. Because the public conversation is loud, confident, and aimed at the wrong thing.

The conversation is about the model. Will it wake up. Will it take the jobs. My model versus yours. "The AI said X." All of it treats the model — the trained system, the weights — as the thing that matters. But the model is the commodity. Many labs are converging on the same capability, and the price is collapsing toward nothing. Arguing about the model is arguing about the cheapest, most interchangeable part of the whole arrangement.

Watch the four of us and you see what the model-talk misses. Each person's AI is different — and not because we run different models. We mostly run the same one. It's different because it takes the shape of the person using it. A designer's mind, a Rails engineer's, a founder's, mine — four different ways of thinking, and four differently-shaped tools out of one model. They are not digital twins; nobody has been cloned. They are projections: partial, angled casts of how each of us thinks, shaped by how each of us works. The tool does not have a mind. It borrows the shape of yours.

Which is why the two most common takes are both wrong. It is not a crutch for people who can't do the work, and it is not a replacement for the people who can. The engineers getting the most out of these tools are the ones who least need them — people who could build the entire thing by hand. A projection of a strong mind is powerful; a projection of a weak one is weak. The tool amplifies whoever is driving. It doesn't lift the floor for the unskilled; it raises the ceiling for the skilled, and it widens the gap between them. That is the opposite of the equalizer people imagine, and the opposite of the replacement they fear.

And the people using it well know exactly what they are dealing with. This is the part the public conversation can't picture, because it swings between two fantasies: that the AI is an oracle you can trust — "the AI said it" — or an autonomous mind that will act on its own. Neither is how any of us uses it. Carl reads every line of code his AI writes; I focus on the user experience — I check what it built the way the person who has to use it will. Same tool, two vantages — he verifies the implementation, I verify the experience — because we come at the work from opposite ends. Neither of us hands off the judgment. That habit — check the output against what you actually know, at whatever layer you know best — is the whole difference between using the tool and being used by the stories about it.

So what actually decides whether AI is useful, or safe, or valuable? Not the model. It's everything around the model: the context it's given, the system that routes and runs it, and the verification that catches it when it's wrong. People inside the field have a word for that layer now — the harness. The venture capitalist Tomasz Tunguz has spent months showing, with benchmarks, that the same model performs wildly differently depending on the harness around it, and that the harness matters more than the model.^[1] Hold the model still, change the room, change the result. The value was never in the weights everyone is arguing about. It's in the room — and in the person who built it.

Here is the fair part: the public isn't stupid. The model is the visible, marketable, nameable part. It has a personality, a launch event, a doomsday scenario. The room is invisible infrastructure — context, plumbing, verification, a person's own judgment. So the conversation fixates on the part it can see and name, and misses the part that actually decides the outcome. We argue about the mind of the machine and ignore the room it's standing in.

The literate version of the conversation isn't "will the model wake up." It's quieter and far more consequential: whose context is it running on, who owns that context, who verifies it, and who's accountable when the room is wrong. My mother — who reads constantly and understands computers — got the entire thing in about five minutes, once I showed her that my AI and my co-founders' are different not because of a smarter model but because each is a projection of a different mind, running in a different room. She is not an AI expert. She just asked what it actually is instead of what it might become.

The model is the cheapest, loudest, most interchangeable part of this, and it's the only part most people are talking about. The thing that matters — the room, the context, the verification, the mind kept in the loop — is quiet and invisible and doing all the work. Four engineers who never take the machine's word for it — one reading the code, one living in the user experience, each checking it at the layer he knows best — will tell you more about where this is going than any argument about whether the machine is about to think. Look at the room. Look at who's standing in it.

Sources

1. Tomasz Tunguz (Theory Ventures) on the harness — the software layer around a model — mattering more than the model itself: "Aftermarket Harnesses" (tomtunguz.com/aftermarket-harnesses) and "The Harness Margin Opportunity" (tomtunguz.com/the-harness-margin-opportunity).