unknown

AI coding assistants are fundamentally changing software development, with some developers already using AI to handle most implementation work while focusing on product strategy and user needs. The industry is splitting between 'experimenters' who embrace AI tools and 'guardians' who prefer traditional coding. As technical implementation becomes commoditized, success will depend on understanding users, making strategic product decisions, and effective distribution rather than coding skills alone.

The Last Programmers

Anja Petry

I am honestly a guardian over anything and I don’t plan to change. Yes they are getting better but that doesn’t mean you are getting any better. Who would want a mediocre dev that eventually doesn’t understand what AI is churning out? How will you as an engineer make tough decisions and take ownership if you don’t fall and learn from mistakes, only blindly taking into consideration everything the AI assistant is throwing out?

Thank you for sharing your experience. If what you call people who look at Vibe coder lazy as “guardian” then I am fine with being a “guardian”.
The problem is not about “not learn AI tools” but its about if that users has a good coding foundation to use AI tool. Dive head in without learn how to program, how to design architecture is like giving the computer to a monkey, and the application is bound to break at some point.
I have work in an project where the team only have junior dev and everything was written with AI. Well, feature come out extremely fast. But did we deliver? None, the project was like a scam, the AI fail to build the user required feature and no one understand what is going on, it was nightware.
As a guardian, I do agree that understand the users is the most important thing. And I am also not against using AI since I also use it.
What I am against is the suggestion to use AI and trust it. Imagine calling your company a software company without understand how your application work.

Millions are poored into iphones, laptops, operating systems, etc. They didn’t really evolve in the last 5-10 years, the author assumes that gpt will become good just because millions are being poored into it. That’s not how it works, you eventually hit a plato and you need new breakthroughs and different technology. We are already seeing gpt slow down innovations between major versions, sometime they are even worst.

Writing code is much more rewarding than watching code.

My main concern on these approaches is not the “Good enough code” or “Good enough quality” and even “Good enough performance”. It’s the good enough correctness.
I know every software has bugs, of course. However, a good software is built towards the direction of correctness (or it’ll fail to meet user needs). If your field can tolerate good enough correctness, that approach might work. If your field cannot tollerate good enough correctness it’s the receipt to failure.