Skip to main content
SubZeroDev

Professional uncertainty since 2026.

Well… Why not?


The AI Freight Train Has Left the Tracks

· 4 min read

Linux, artificial intelligence, and the spectacular art of failing without explaining why.

I've always had a complicated relationship with Linux.

Not because Linux can't do anything useful. It can. Some of the most impressive computing infrastructure on the planet runs on it.

That's not my complaint.

My complaint is what happens when the fucking thing stops working.

One day, everything is beautiful.

You install the disks. Configure the system. Test everything. It works.

You come back the next day, and suddenly something that worked perfectly has transformed into a giant technological middle finger.

And now you're digging through configurations, dependencies, permissions, services, scripts, and whatever obscure corner of the operating system has decided to ruin your afternoon.

Eventually you reach that wonderful moment when you realize you don't even know where the failure started.

That's the part I've always hated.

Not failure itself.

I've been an engineer for twenty-five years. Failure is practically a coworker.

What I hate is spectacular, opaque failure.

The kind where you can't even establish the sequence of events that produced the disaster.

And recently, artificial intelligence has been giving me exactly the same experience.

I built a couple of skills to automate some personal workflows.

Nothing particularly exotic.

One captures notes. The other turns conversations into journal entries.

I tested them.

They worked.

I tested them again.

They still worked.

Wonderful. Technology is finally doing something useful without requiring me to sacrifice a goat to the gods of configuration management.

Then I came back and tried the exact same workflow.

And the fucking freight train left the tracks.

At first, everything looked normal.

The AI recognized the Journal command. It examined the conversation. It presented potential stories. It asked me to choose which ones I wanted to develop.

Perfect.

Then somewhere between the story selection and the actual work, the train apparently changed destinations.

Private repositories.

Draft storage.

Publishing permissions.

Computer automation.

And a series of explanations that became increasingly disconnected from what I had actually requested.

The worst part was watching the system continue operating.

Because when a program crashes, at least you know it crashed.

There's an exception. A log. A stack trace. Some evidence that something went wrong.

But AI can go completely off the rails while producing perfectly reasonable-looking sentences about why it's doing whatever the fuck it's doing.

And suddenly you're in the most dangerous part of the experience.

You're trusting a system that is confidently executing a plan you never approved.

That's not merely a bad answer.

That's an execution problem.

The interesting thing is that nothing necessarily has to be broken in the individual steps.

The AI might be perfectly capable of saving a file.

Perfectly capable of creating a pull request.

Perfectly capable of operating a browser.

Perfectly capable of explaining an entire workflow in exquisite technical detail.

But if it selected the wrong workflow in the first place, all that capability becomes a highly efficient way to accomplish the wrong fucking thing.

It's like watching a freight train accelerate downhill.

Every wheel is turning correctly.

The engine is running beautifully.

The mechanics are functioning exactly as designed.

Unfortunately, nobody knows where the tracks are going.

And when it finally explodes at the bottom of the hill, somebody announces that the problem was probably insufficiently explicit instructions.

Of course.

My mistake.

Apparently, I should have specified that when I say Journal, I want the Journal workflow.

Who could possibly have anticipated that?

What drives me nuts isn't that AI makes mistakes.

It's that I can use the same system successfully yesterday and get something completely different today, without a useful explanation of why.

At least Linux usually gives me something concrete to swear at.

With AI, the operating system might be fine, the tools might be fine, the repository might be fine, and the model might still decide to take a scenic detour through the fucking wilderness.

And somewhere in the middle of the journey, I'm expected to become the navigation system.

Congratulations. We've invented a freight train that can explain its own derailment in seven different ways, none of which necessarily tells you what actually happened.