On Vibe Coding

At this point, the toaster is about the only thing that hasn’t told us how much easier AI has made building software.

In January 2022, around two thousand new apps with a monthly subscription model were launching worldwide. In January 2026, that number was already above 14,700.

Roughly a 7x increase.

At first glance, it sounds great: programming has finally become accessible to almost anyone. Anyone with an idea can open Cursor, Claude or Codex and have a working product a few days later. What once took a team, months of work and a decent budget can now sometimes be done by one person over a weekend.

But there’s a catch…

Apps launched before 2020 still take in 69% of all subscription revenue.

And apps launched in 2025 or later — in the age of vibe coding — get just 3%.

A year after launch, the median new app earns around $72 a month. Only 4.6% reach $10,000 in monthly revenue within their first two years.

In other words, we have seven times as much software, but not seven times as much value.

Apple is seeing much the same picture. In 2025, the company reviewed 9.1 million app submissions and rejected more than two million. Of those, 371,000 were rejected for copying other apps, spam or misleading users. Apple itself specifically notes a sharp rise in submissions thanks to new development tools.

Which brings us to a funny conclusion.

AI hasn’t democratized software development. It has democratized the ability to put something out there.

And those are very different things.

I’ve been programming for most of my life, and the better the models get, the less I feel that they ask less of us. If anything, it seems to be the opposite.

As the time we spend writing code shrinks (though don’t quote me on that), the craft moves into other tasks:

— understanding what problem we’re solving;
— breaking it down properly;
— designing the architecture;
— knowing what a correct result looks like;
— testing our own assumptions;
— noticing when the model starts talking nonsense;
— and, finally, having taste.

These are all skills of the highest order. They take years, sometimes decades, to develop.

A weak developer with a weak tool used to write a small amount of bad code.

A weak developer with a very powerful model can now produce an enormous amount of quite plausible-looking bad code in the same time.

I’m not sure that’s progress!

Google’s DORA research describes AI in software development as an amplifier: it amplifies the strengths of good engineering systems — and the weaknesses of bad ones. You know how music makes happy people happier and unhappy people unhappier.

Much the same happens with an individual. The more powerful the model becomes, the more the person directing it matters.

And perhaps this is where we’ve confused two things: an abundance of code and an abundance of meaning.

We’re being given more than enough of the first. We’ll have to figure out the second ourselves.

Vibe coding has democratized execution. That has only made mastery scarcer.

Here’s to good tools — and a head that knows what to do with them! 😎