On the gap between building and watching

Two ways to use AI

Among people who use AI tools, a divide is emerging. I don’t think it runs between techies and laypeople, or between enthusiasts and sceptics. It runs between people who understand what these tools make possible and people who (still) don’t see it.

That second group uses AI as a search engine on steroids. You type a question and get an answer that is reasonably right, and there is nothing wrong with that. It just stops there. They have no idea that someone else has meanwhile built a complete web application, implemented a booking system or designed a data model, and worked deep into the night to do it, only to start the ordinary work early again the next morning.

The sneak peek paradox

You see that problem most sharply in online communities and group chats around this kind of project, and it has a paradoxical side.

People who build something often want to share it, to show others what is possible, and less to show off. Look, you can do this too. So sneak peeks get shared: a first screen, a working feature, an early prototype. The intention is to help others get started.

What happens is different. The reactions are enthusiastic, but soon impatient too. Interim updates suggest fast progress, so the audience expects the end result even faster. Meanwhile well-meant advice pours in that ignores the plan the builder already has in his head. Because from the outside nobody sees what has already been decided, that advice feels more like noise than help.

You could then stop sharing sneak peeks and show only the end result, but that undermines what the builder wanted to achieve: showing that this tool is accessible to others too. Share too much and you get unfounded expectations and interference you did not ask for. Share nothing and the gap in understanding stays just as big.

Prompting is not building

That gap goes deeper than impatience. It sits in the frame of reference people use to look at this kind of project. Because they don’t see what is happening, they fill the hole with what they know: you just write a few questions and the system does everything else, right?

That is partly true, and to a large extent it is not.

The outsider thinks you type something into ChatGPT and then it works. What he does not see is that people who build something with these tools are constantly making decisions, solving problems, testing and rebuilding parts. The AI is the tool in all this.

That distinction is almost invisible to anyone who does not know it from the inside. The builder knows what it costs in time, energy and mental bandwidth. The spectator only has the comparison with a search query: a question, an answer, done. They cannot see for themselves how far apart those two pictures are.

What those outsiders miss is the difference between prompting and building: between asking ChatGPT a question and a development process in which you keep deciding on architecture, scope, error handling and user experience, and in which AI does not take the thinking off your hands. The AI writes the code. You decide what has to be built, why, in what order, and what you do when it turns out differently than expected.

You don’t see that work, and that is what creates those unfounded expectations.

Where we stand now

The result is commentary from the sidelines that does not do justice to what is going on. Taking long. Is it done yet? While the builder is debugging an edge case until midnight that nobody saw coming.

I don’t think the difference is technical. It lies in the picture people have of AI. Anyone who knows AI only as a question-and-answer machine does not see what happens on the other side of that screen, and I expect that only changes once they have experienced it themselves, in a project of their own and not in a demo.

I don’t mean that as a reproach. That is simply where we are.

Jan-Willem Rodenhuis is managing partner of Ductus B.V. and writes on miliarium.nl about the practice of AI adoption in organisations.