The AI Development Conundrum

Published ....
Last modified ....

Share this post on BlueskySee discussion on Bluesky

While I spent a bit of time talking about using AI to help write code in my recent blog post on offloading complexity, I've been thinking more about the trade-offs of heavily using AI to aid with software development.

I was working on a side project the other day, and I decided to give Cursor's composer mode a go at building out a whole feature from start to finish. Normally what I opt to do is start writing the code myself and then rely on the autocomplete features to fill in some of the details. However this time I was diving in head first on letting the AI do all of the deep thinking and implementation as well.

While the feature seemed to mostly work, aside from a few edge cases that I caught when testing it out locally, I also realized that I had no idea what the code was doing.

I was blindly hitting Accept ✅ on it's suggestions, and just assuming that it would do the right thing along the way.

It hit me after a few minutes that I was just there to prompt the system and accept or decline it's suggestions.

I've been thinking about this for about two weeks at this point, and I still feel fairly uncertain about the whole thing.

I got into software development because of the thrill of coding up something that I could see and interact with. Solving difficult problems has been what I really enjoy about the job and the hobby as well!

To borrow from my earlier blog post, offloading the problem solving part of the work (or hobby) feels like it's against the point of doing the work in the first place.

And yet - not using AI at this point to build software feels like rejecting C when you have been writing assembly (like - why wouldn't you opt to become more efficient?)