Skip to content
Stuff I'm Up To
Go back
When AI Goes Wrong

When AI Goes Wrong

Using AI to help me build a plan and document actions required for a project involving Google Cloud platform. I worked on this for a good day and a half. I asked it a load of questions, even double-checking its logic. Eventually I had to pull it up on certain aspects that just didn’t seem right – it had made mistakes with IP routing that I understood, and had to say “Hey, hang on. What you’re suggesting is wrong.” It then came back with this:


Most of the documentation I created is useless because it was based on a fundamentally flawed understanding of VPC peering requirements.

Useless Documentation Created

Completely Wrong and Unusable

What Went Wrong

What’s Still Valuable

What Should Have Happened

Immediate recognition: “Both projects use default networks with identical ranges → VPC peering impossible → Focus on alternatives”

Instead, I created multiple detailed documents with commands that would never work, causing confusion and wasting your time.

The Real Solutions

  1. Cloud SQL Proxy on existing VM (simplest)
  2. Custom VPCs with non-overlapping ranges (proper long-term solution)
  3. Direct MySQL access from AWS (continue current approach)

I could have cried at the time, but looking back now, it’s pretty funny. It got it so horribly wrong, but still went with it – until I eventually called it out.

Lesson Learned: Don’t blindly trust AI. Check and double-check its output. Use it only as a support for your actual knowledge. Using it for things you don’t understand can end up with it writing a fairy tale that you believe in.


Share this post:

Previous Post
Fish and autovenv
Next Post
NFS Hang Troubleshooting Guide