AI can code, but it can’t build software

Have you noticed that quite a few people are looking for technical cofounders
or CTOs right now? I, for one, get a surprising amount of these queries; most
of them along the lines of “hey, I have this vibe-coded app, would you like to
make it production-ready”. I have sort of a profile for these people. Think
someone who knows their business but has always lacked the technical skills to
make their ideas happen — a legal counsel, perhaps, or an account manager.

Why would these people need me? That’s what I’ve thought about a little bit,
and I think there is an important signal here: What is it exactly that they
can’t get done with GenAI alone? This is something everyone is trying to
understand, right? Everyone wants to know what these models can do. Or, to be a
little blunt, everyone wants to know which jobs are soon to become obsolete.
The fact that I get these requests says something about software engineering. I
mean, if software engineering was automated, no one would be looking for
technical cofounders.

Well, I think I know why we get these proposals. The thing is that AI can
code, but it can’t build software. This is the conclusion I’ve come to after
spending a significant amount of time writing AI-assisted code and watching
demos by other people.

There is old wisdom that says: Coding is easy, software engineering is hard.
It seems fair enough to say that LLMs are already able to automate a lot of
coding. GPT-5 and the like solve isolated well-defined problems with a pretty
nice success rate. Coding, however, is not what most people are getting paid
for. Building a production-ready app is not coding, it’s software engineering.

The way I see it is that coding becomes software engineering around the point
where you try to turn your demo into a real product — which happens to be
exactly the point where these people reach out to you with their pitch.

I don’t really know why AI can’t build software (for now). Maybe it has to do
with the nature of the job. When you write software for a living, your main
task is to deal with complexity. The average production software only does a
bunch of easy things. The challenge is doing hundreds of these easy things at
once, and keeping the whole thing maintainable. Or, to rephrase this in the
present context: It’s one thing to demonstrate a feature. It’s a much more
difficult thing to build that feature in a manner that supports integration,
expansion, and long-term maintainability.

When you look at the code these people send you, you realize that “making the
app production-ready” really means torching the whole thing and starting from
scratch.

I think this says a lot about where we are at right now.