AI: Don’t Be Like Me
One morning last month I got the flu and COVID shots, and my midday I was back on my heels. Knowing I was not in top form, I thought I’d try letting Gemini have a loose rein and let is guide a rather messy bug hunt in Tinderbox’s JSON code. I’d sit back and see what Gemini proposed.
Gemini was completely at sea, but close enough to plausibility that I kept being tempted to let it try just one more patch, one additional escape sequence, one more fix. The results were worse and worse. After a few hours of this, I used the version control system to roll back to the morning and headed for the barn.
Both Claude and Gemini are remarkable assistants, but neither is particularly strong on either strategy or elegance. If you guide them, they can refactor, but their own ideas of refactoring are seldom sound. Yet, both are remarkably strong on syntax questions; I’ve yet to encounter a syntax error they cannot explain. They’re both terrific at analyzing crash logs, though they remain too confident that their proposed solutions will suffice.