I usually start my day or my lunch break with perusing Hackernews. Sometimes to my amusement, sometimes to find new cool stuff and sometimes to my absolute detriment.

As of late, the amount of AI-glorifying comments is insane. If you are even in the slightest against its usage, the usual comeback is one along these sentiments:

• You are prompting it wrong
• Get with the new tech, bruh
• But have you used agentic coding? (i.e. have you thrown copious amounts of money at it?)
• Well, you must be using a language that it's not well trained on

and many, many others.

But you know what? If AI is as good as some people claim it to be, then why the fuck should it matter which language I am using? If it needs more training on a certain language and just can't find enough on Github then it means it cannot apply knowledge-transfer. And let's be honest here, 95% (random wild guess) of code on Github or other platforms is mediocre at best. And that's what the AI systems have been trained on. So what they'll produce is also going to be prone to be mediocre.

One common reply I get is "but, have you actually tried it?". Yes, yes I have.

When I am not on a time crunch I entertain myself by randomly throwing problems at it, even simple ones. The amount of fails is endless.

AI coding tools promise to make you faster. Type a comment, get a function. Describe what you want, get an implementation. But what you're actually getting is a black box that produces code you didn't think through, don't fully understand, and can't properly maintain. You're outsourcing the exact part of programming that makes you better at it.

The deeper issue I have with AI is what it's usage does to the profession and/or hobby. We're training a generation of developers who can prompt an AI but can't read a stack trace. They can maybe describe what they want but can't explain how it works. They think programming is about getting code to appear on screen rather than understanding the system you are creating.

Sure, for Javascript it might kinda work because there are enough examples. But, let's take something like the Linux- kernel, or =OpenBSD or FreeBSD. There are not many code examples for an AI to learn from in these regards. It would be prone to pjust repeat exactly what exists.

I dare you to let AI write a GPU driver. Honestly, if you get a working and performant one, I wanna see it.

Some other comebacks I see are: "well, as a senior engineer I am able to work without any juniors as AI takes these tasks.".

Ok…cool….but how the fuck do you think you became a senior engineer? Years of being in the trenches, debugging shit for hours. It's frustrating in the moment, but also rewarding.

The new generation will look at people who can actually use GDB like some of us are looking at people who can write correct COBOL code.

I will leave out the environmental impact of AI in this post, because that train has left the station long time ago.

But if you think the world is already in a bad state if a single AWS or CloudFlare outage can bring half the internet down, wait until most companies rely on AI and wait for OpenAI or Anthropic to have an outage. Nobody will be able to troubleshoot anything.

"Sorry boss, I can't finish the task because Claude is down."

And these companies are operating at a huge loss. You think agentic Claude is expensive now?! Wait until they actually need money back or else they'll shove ADs down your throat until you choke.

So yes, I could probably write code faster with AI assistance. But I'd understand it less. I'd be a worse programmer. And for what? To ship features slightly quicker? To avoid the hard work of actually learning my craft?

I'll pass.

I'll keep reading source code. I'll keep debugging my own mistakes. I'll keep building that deep, hard-won understanding that makes you good at this work. Because programming isn't about making code appear. It's about knowing what you're doing. And no autocomplete feature, no matter how impressive, can give you that.