
Vibe coding. Suppose I ask an AI to make a marble roll across an Apple Watch when I tilt my wrist. It writes some code. I try it. At the slightest tilt, the marble shoots across the screen and hits the far wall. On rereading my request, I have to concede the point.
Vibe debugging. “A little slower, please.” The AI has several ideas. Before trying any of them, I put the watch flat on the table. If the marble still rolls, speed can wait. I’d quite like to know which way it thinks downhill is.
A matter of taste. How fast should a marble roll on a wristwatch? Five minutes ago I had no particular opinion. Now I can tell you the first version felt like a pinball and the second like a grape. Somewhere in between is the game I wanted to make.
Would I install it? Would I install any app that happens to run? Good Lord, no. I’ve seen “it works on my machine” do quite a bit of duty as a security review. A favourite question of mine is what happens when somebody gives the app an input its author never imagined. OWASP has quite a bit to say about that. “I asked it to roll a marble” seems a rather thin answer.
Before it ships. I’m a mathematician and a computer scientist. I’ve programmed in assembly, Lisp, Pascal, C and Python, and I still write code today. When AI contributes code, I read it, work through the parts that surprise me and sometimes rewrite them myself. Every app on this site goes out only once I understand how it works and have decided it’s ready. If the marble rolls uphill after that, you know whom to write to.