I am not a professional software developer. I’ve worked in product management for the past 36 years, but have written a reasonable amount of code for myself. I decided to try co-writing a some software using AI. I chose Copilot because it was available. This is my experience. Please don’t comment on how Copilot sucks. I think most AI software has similar limitations.
Part 1 — AI Accelerates Execution, Not Direction
What happens when assumptions go unchallenged.
I started a recent project with a simple goal: build a clean front‑end interface for a mesh‑networking device. I hadn’t touched modern React or Node.js in years, so I leaned on AI to generate the scaffolding.
Within minutes, I had a working UI. The productivity boost was real.
But the first lesson came quickly: AI accelerates execution, but it doesn’t validate your assumptions.
When I tried connecting the UI to a real device, nothing worked. I asked why, and only then did the AI explain. The device didn’t speak WebSockets or HTTP at all. It used Protobufs over raw TCP/IP. The AI knew this from the documentation, but it never surfaced the limitation until I explicitly asked.
That moment re-framed how I think about AI:
1. It won’t investigate a problem unless prompted.
2. It won’t challenge your assumptions.
3. It follows your direction—even when the direction is wrong.
AI can move you faster, but only if you’re asking the right questions.
How are you making sure AI isn’t just accelerating the wrong assumptions?
Leave a Reply