There's an AI-written PR open in my editor, waiting for a review for 4 days. It's only 3k lines. I have struggled reviewing it. Instead, I seem to prefer playing Manor Lords or writing a blog post like this to vent. This is unprecedented. There were better times when I managed to write and review larger PRs in one sitting.
I started using AI for coding a couple of months back, and since then my eagerness to code has plummeted. There used to be mornings when I couldn't wait to sit at the computer to pick up where I'd left off, or sometimes to implement those wild late night ideas. Lately, I have to force myself to code, often with less-than-ideal results.
I thought to myself, why isn't AI coding satisfying? I used to think through the problem, solve it, code it, and tweak it until finally it worked. I got that rush most of us call the "dopamine hit". Now, I still solve the problem, but tell AI what to do, and it does it. When it works, I feel nothing. Why? Was it never the problem solving that drove me but the coding itself? I coudn't be, I generally like solving problems and puzzles.
Then I realized: half of the "problems" I work with aren't actually problems. I already know how to implement a CRUD for a model or push a notification over WebSockets. There's nothing to solve there. But, coding was always a bunch of mini-puzzles to solve: naming variables, finding the right level of abstraction, all those small decisions. With AI, I lost the satisfaction of coding these kinds of problems.
The other half of the problems are actual new problems, the kind where I might need to read an RFC or find a good dependency. Before AI, these gave me two satisfactions: solving the problem and writing the code. I still solve the problem myself before handing the implementation to the AI. So, I should get 50% of the satisfaction, right? I don't. Why?
I think it's this: Even though I thought I solved the problem before handing it over to AI, I hadn't. I realized, it is almost impossible to solve a hard problem without coding. For new problems, code isn't just how I implement the solution. Coding is how I think. Take picking a dependency, for example. Before AI, I'd pick a dependency, learn about it, and run into its weird quirks. I'd discover that it was actually a fork of an old project whose original author abandoned it after a heated argument on a forum. Finally, I'd write a minimal implementation to try it out, only to find out it couldn't support my extremely niche use case. So I'd settle on another one. It was fun. I learned a lot and found a good solution.
With AI, I have two options: let AI choose the dependency, or pick it myself. The first offers me no learning and no satisfaction, and there is no guarantee that AI chose the correct one. The second is almost impossible without actually writing code. The problem requires coding to best solve it.
So, what's the solution? I don't know. I am going to try traditional coding in the coming weeks to see if going back is possible.
Comments