With AI, yes, one prompt can generate a full working, polished UI. That is the version people show off. The way I actually do it is different. It takes 400 prompts, not 2. I iterate, fix bugs, test each thing by hand, sometimes read every line it produced. I didn't type all the code, but I had to stay up late after my daughter went to bed to iterate on the ideas — sleep-deprived, working across multiple projects, with family duties on top. When you live like that, you cannot sit and debug a long piece of code in one sitting. Sometimes my daughter does not let me type two words before she needs me — which is actually a very good thing — and that is when AI helps me execute and try ideas out. Things that otherwise would have had no outlet at all. They become possible. So when somebody dismisses what I built and says it was just AI, I think they are missing the bigger picture of what is happening, and where the shift is going. They also don't see why I am using AI in the first place.
What makes this work is that writing code is cheap now. Cheap enough that you can put together a rough version of an idea just to see if it works, instead of arguing about whether it might. You try a few, keep the one that holds up, throw the rest away. That used to be too expensive to do. Now it isn't. You can fail faster, and fail cheap.
And when the cost of trying drops that low, things you used to shelve come back. Work that used to need pyramid-scale coordination — generations of architects, armies of builders — now sits within reach of one person, or a small team. A new compiler. A simulator. A research tool. A platform. Those got shelved because trying was expensive and failing was expensive. AI doesn't build those things for you. It just makes failing cheap enough that you can try.
So the bottleneck moves. It used to be: can we build it? Now it is: which of the things you tried is actually good? Typing got cheap. Taste — knowing what an experiment is telling you — got scarce.
There is something else happening underneath all of this. When code can be produced faster than any one person can read it, you start losing the ability to understand what is being built. I keep wondering if this is the same kind of shift as Assembly → C++ → JavaScript, where each layer let us build bigger things by quietly abandoning the one below. A JavaScript developer doesn't need to read assembly. Maybe AI is the same shift. But there is a difference, and it is worth being honest about. Those older layers were deterministic and well-tested. You could trust the one below even without reading it. AI-generated code isn't like that. Which makes the same trade more risky.
All of it comes back to one thing about AI. It is a synthesis of what is already written. It is good at combining what is known — including ideas from smart people you will never meet — but the genuinely new isn't in its data yet. The part that is left for humans is the part that isn't recombination. Picking which of your experiments is actually good, and judging what is worth attempting at all, turn out to be the same thing. Both are about reading something that is not in the data. So when you ask an LLM whether an unfamiliar approach is going to work, or can this disease be cured? Will I fail at this? — you are asking it about something it cannot see. And it can bite you. Humans are known to push the boundaries of what is not possible. Right up until somebody does it.
None of this is what most people are arguing about. The debate has been: will AI take the coding jobs, or just let us ship more of the same CRUD apps faster? Both assume we are going to keep building the same things. People are looking at AI wrong. And because of that, they are using it wrong. The tool was never the question. What we decide to build with it — and what we decide not to ask of it — is.
One more observation I keep coming back to. When everyone on a team can ship code fast, the work shifts. If each person is putting up dozens of PRs with new features in a day, the bottleneck is no longer writing the code. It is reading it, reviewing it, keeping track of what is being shipped, and deciding what is worth keeping. Cheap to produce. Expensive to verify. My observation is: when output grows like that, knowing what is good takes longer than producing it. I do not have a clean answer for what to do about that yet.
And I think we are going to need more people in this kind of work, not less. The skill is shifting from typing to judgment — knowing what AI is good at, what to ask of it, and what not to. That takes practice. People will need to learn it, not be replaced by it.