What happens to innovation when implementation becomes almost free?

An open observation by me, Johannes Wachter. Lecturer at FH Vorarlberg, open-source developer, core developer of the Sulu CMS, and someone who works with AI every day.

A note to my students, before anything else. You did good work this year, and I mean that. Nothing here is a criticism of you, and nothing here is meant to take away from what you built. What unsettles me is almost the opposite. For the first time, I find it hard to tell who among you was the most motivated, the most curious, the most driven, because the work no longer makes those differences visible the way it used to. And I will be honest: that bothers me more than I expected. Not because you did anything wrong, but because those were exactly the qualities I valued most, and I do not want to lose the ability to see them. And to be very clear: this has nothing to do with your presentations next week. Those stand entirely on their own, and none of this reflection plays any part in how they are watched or judged.

This year, my students at FH Vorarlberg unsettled me. Not because the projects were bad, but because they were so good.

I have supervised innovation projects there for several years. In the implementation semester, teams define an MVP at the beginning and then spend the rest of the semester building it. This year, using AI was not just allowed. It was expected and explicitly encouraged, exactly the way I work myself every day.

The result was impressive. Almost every team reached its MVP, the technical quality was high, the execution was clean, the problems were manageable, and there was very little failure.

And yet the semester left me with an odd feeling, because there were hardly any surprises. No project was really bad. But no project was really outstanding either.

Why that gives me pause

In previous years, things usually looked different. The plans for the MVPs were almost always too big. Teams had to prioritise, cut features, throw assumptions overboard, change direction, deal with dead ends, and sometimes fail.

And looking back, I think that was often exactly where the learning happened. Not in the implementation itself, but in the decisions around it.

Our MVP plans used to be almost always too ambitious. Now I wonder whether they have suddenly become too small.

The real observation

The interesting part is not that AI was used. That was the plan. The interesting part is that, despite massively better ways to build things, the teams landed pretty much exactly where they had planned to at the start.

I had honestly expected something else. More features, more experiments, more iterations, more surprises. Instead, I saw better output, but similar goals and similar results.

That is what irritates me.

And not only in the classroom

The projects were simply where I noticed it first, clearly and side by side. But the more I pay attention, the more I see the same pattern elsewhere.

I see it in content, where a clean, well-structured draft now appears almost instantly. I see it in code, where a working solution is rarely the hard part anymore. And sometimes I see it in myself, even though I am fully aware of it, which is maybe the part that unsettles me the most.

The trigger was a classroom. The pattern feels much bigger than that.

Is AI moving the bottleneck?

This is not a fact I want to claim. It is a hypothesis I want to put carefully.

Code is getting cheaper. Design is getting cheaper. Content is getting cheaper. Implementation as a whole is getting cheaper. Not free, of course, but cheap enough that it changes the game.

And that makes something visible that was harder to see before: maybe implementation was never the real bottleneck.

Differences between teams used to show up very clearly in technical skill, implementation ability, architecture, execution and available resources. Many of those differences are becoming less visible now.

The differences are still there

I do not mean to say that everyone has become the same. That would be wrong.

The students are still differently motivated, differently talented, differently creative and differently ambitious. But technical execution may no longer be the primary place where those differences show up.

That is a big shift, because for many years execution was one of the clearest signals we had. If a team was motivated, we saw it in the product. If a team had strong developers, we saw it in the product. If a team went beyond the assignment, we saw it in the product.

But what happens when AI makes many of those visible differences smaller? How do we recognise excellence then?

The role of friction

This is probably the point I keep thinking about the most.

Historically, good ideas often got better through resistance: through colleagues, through customers, through technical limits, through reviewers, through reality itself. Friction was uncomfortable, but it was productive.

AI removes a lot of that resistance. Not intentionally, but simply by its nature as a cooperative conversation partner.

And there is a second, quieter effect. Today’s LLMs work by predicting the most probable next step, so by design they lean toward the average, the expected, the statistically likely. They are, in a way, machines for the mean.

That is wonderful when you want something solid and conventional. It is less wonderful when you were hoping for the strange, the sharp, or the genuinely new. Take away the friction, add a gentle pull toward the middle, and you get exactly what I saw all semester: competent, clean, and rarely surprising.

These days, I sometimes feel I am only one prompt away from having an idea confirmed, improved and implemented. That is incredibly powerful. But it also raises an uncomfortable question: what happens to our ideas when that friction disappears?

The illusion of thinking

And so that I am not just talking about other people here: I am not exempt from this either.

Sometimes I catch myself wondering whether I actually worked through an idea, or whether I just got it phrased so well that it felt like I had. There is a difference between thinking and synthesis, and AI is excellent at synthesis. It can structure ideas, connect arguments, refine language and make vague thoughts feel clear.

But do we sometimes mistake the feeling of a thought process for a thought process we actually lived through? I am not sure we understand the long-term consequences of that yet.

Open source and art

A few years ago, I would never have said that programming is a form of art. It would have felt pretentious to say out loud.

Today I see it differently. Not because of the code, but because of the idea.

For me, open source means putting puzzle pieces into the world, giving up control, and not knowing what will eventually come of them. Someone may use a small abstraction in a way I never imagined. Someone may build on top of an idea years later. Someone may connect it with something completely unrelated.

The real value often lies not in perfection and not in completeness, but in possibility.

And this is exactly where the circle closes back to the AI discussion. The most exciting things rarely come from perfect execution. They come from curiosity, risk, contradiction and ideas whose value no one can see at first.

Why I am writing this on LinkedIn

I am publishing this here on purpose, because on LinkedIn AI is often discussed as either salvation or catastrophe. My observation fits into neither of those camps.

I work with AI every day. I build products with it, develop open source with it, and use it to realise ideas that would not have been possible for me before. That is exactly why I find this observation interesting rather than threatening.

I am not worried because AI does not work. I am worried because it works so well that it may change what we should be looking for.

An open ending

I do not have a finished answer, and I do not want to pretend that I have one.

Maybe I am reading too much into a single observation. Maybe this semester was an exception. Maybe my perception is simply wrong.

But the question will not let go of me: when implementation becomes almost free, what becomes the new bottleneck for innovation?