AI and Software Engineering, the Conflict Within

I wrote a post a while ago about my current workflow when working with AI, my initial skepticism and some brief mentions of the concept of recreational programming. This post is going to be a continuation of the train of thought presented in that previous post, but it will be a stand alone piece still. In other words there is no need to have read the previous one to follow along in this post.

So the past months I have been using AI quite a bit when writing code and I think I can finally articulate why I was a “skeptic” or “AI resistant” the past year and a half. The TLDR is pretty much that in my opinion or rather for me personally the reason is that reaching the end state of a task as fast as possible has never been my primary motivation. I code to explore ideas and problem spaces, it sounds almost pretentious even to me, but I don’t really know how else to put it.

This could be a bit of a spicy take but I think there are two categories of software engineers in the world right now, and most people fall into either of these roughly speaking. The first one is those that got into programming because they are interested in the results that is the end product of writing code, and if they could avoid having to write code to get to those results they would be super happy and not write any code ever again in their lives. The second category is those that love programming and see it as an activity worth doing in and of itself, they love learning how to solve a problem in the best way possible and want to push the state of the art slightly further even.

One could argue, and I would probably agree, that I’m being a bit too reductive here, but I am trying to make a point. I actually think that most people writing code today, based on my day to day experience, would actually be happier doing something else. For most people that write code today it is a means to an end and therefore they are super happy that they can outsource even some part of it to a LLM to get their task done a bit faster etc. For me, I would put myself in the second category in case that was not clear, using a LLM to code is making me feel like I’m being robbed of the experience that I want to have.

Let me expand that last statement a bit more. Since the goal, motivation and cause for dopamine release when programming for me is the learning and exploration of domain / problem space, having the AI do any of that is literally robbing me of the experience that I want to have. So I have now learned to not use AI at all for something like that.

Furthermore using AI when building something new or setting up scaffolding for the first time makes it super hard to internalize what is going on in the codebase for me, and I have found it actually makes me less productive. What I do love to use AI for is generating solutions to problems I already know how to solve, because nothing makes me lose focus than having to solve or write code for something I already have done or know how to do.

In the end there is a bunch of stuff that I do still use AI for when it comes to programming and writing code, but I have identified for me when I should not use it so that I still get the experiences that I want to have. This does take a fair bit of self discipline in a professional setting though because there is almost always a, sometimes self imposed, pressure to deliver results as fast as possible. In a hobby setting I essentially only use AI as a compressed search engine and nothing more.

To reach this “insight” I did have to lean super hard into using LLM’s for almost everything though to really learn how to use them and when. Without committing to going all in I don’t think I would have “grokked” it, because the pace, hype and amount of content around this technology is a bit insane at the moment. Slowing down and going deep (Cal Newport reference) has always been my way of figuring out things for myself, and usually it pays off in the long run also. But I guess only time will tell.

As a side note, I do think there is a third category being created as we speak and that is programmers / people that enjoy getting the LLM to output what they have in mind. Just browsing some of the subreddits and other forums of people that talk about how they use swarms of AI agents, massive prompts or many custom prompts for every possible scenario makes it evident that a new way of working with software is definitely being shaped. I find it fascinating honestly but I have not personally gone that deep into it yet, because it’s unclear to me currently how much of the claimed improvements of all that work is going to just be built into the models in the future vs not.

Ultimately, I think the key insight here is that different programmers get satisfaction from different aspects of the development process. Understanding what drives you personally can help you make better decisions about when and how to use AI tools, rather than feeling pressured to adopt them wholesale or reject them entirely. The technology isn’t going anywhere, so finding a sustainable and fulfilling way to work with it seems, to me, like the only sensible approach.