My entirely unasked for reflections on AI

david@mohundro.com (David Mohundro) — Mon, 18 May 2026 00:00:00 GMT

I apologize in advance. Everyone is talking about "AI" right now. Well, actually everyone is still talking about the various flavors of generative AI and LLMs... but WHAT DO I KNOW.

Anyway, I'm adding to the noise. Sorry.

I've been using various LLM tools to assist with coding since GitHub Copilot first came out. Back then, it really was just a (sometimes) smarter intellisense. Since then, I've used Windsurf and Cursor, Gemini and OpenAI, and more Copilot. Today, my preferred tool is Claude Code.^∗My preferred tool is actually a combination of Ghostty with multiple tabs and panes open, multiple Claude Code sessions, Neovim, Jetbrains IDEs and VSCode...

Almost a year ago, we were really pushing the boundaries of what the coding tools could do. We had a multiple month long project that was on a very outdated and unsupported technology... even without LLMs, we would be looking at a significant amount of "busy work" just to get it to the point where we could work effectively in it. Makes perfect sense to use some sort of automation, right? First, we wrote a paragraph or two of explanation and were relatively impressed with the conversion results... maybe it could handle more? Realistically, though, the tool could only convert 2 or 3 at a time... it wouldn't have been too difficult to write a tool to orchestrate multiple LLM calls with a prompt... the problem space was well known to us, we just wanted to automate the pain away.

Four to six months in, we were still finding issues with it, but it definitely helped us move faster than had we done it by hand. Over the next few months, we continued to use LLMs in a similar fashion, slowly increasing our usage from conversion efforts to small features and then on to even larger features.

Earlier this year, once we had decided on a launch date for the new project with our client, everyone got more serious about testing and validation and... of course, many issues show up. Putting a date on the calendar tends to make people more serious about ensuring a system works! In that time, we pushed more pull requests through the system than we had probably pushed out in the prior 2 months. In addition, we started hitting our GitHub Action budgets... multiple times.

(heads up, I'm about to shock industry veterans with something they've known forever...)

THE BOTTLENECK WASN'T CODE.

The bottleneck was a well defined backlog. Clear requirements. And so on.

When I speak with students, they're frequently surprised when I say something like 50% of my job is not coding. And that was true before LLMs. It is still true. Most of my job is still ironing out requirements, clarifying ambiguities, and telling the computer what these conflicting requirements actually mean. The difference is that, today, I tend to use "English-like specs" to do it instead of translating it to code. I still have to understand the code and the architecture at a deep level.

In other words, it is the same issues that Fred Brooks talked about in Mythical Man Month decades ago. Now, LLMs have dramatically reduced the time it takes to code specifically... and they're also shortening the time it takes to debug issues. They're even helping with exploration, diagramming, and explaining (assuming the recipient reads the contents, too... is anyone reading anything anymore? Or just AI summaries?).

In the past month or two, I actually have turned a corner in my usage... for some projects, I especially enjoy using LLMs. If I'm given a nightmare, legacy codebase, it can help me find hotspots and understand the thing in a much faster time.

But I still enjoy the art of coding. I spent a year teaching myself VIM for crying out loud. I still love using keyboard shortcuts in Jetbrains IDEs to navigate through a solution quickly. I love to refactor code.

I'm torn between delivering things quickly and appreciating the craft, coming up with an elegant system that is easy to understand for both the users and the future maintainers.

I honestly have no idea what the future holds for this craft, but I do have an honest idea about where we are right now. I was talking with a coworker recently and I realized that our "AI" today has far more in common with computers on Star Trek than with your typical sci-fi AIs. We're not dealing with Terminator or the Matrix at all... we're also not dealing with Data from Star Trek. We're at the "Scotty talking to a computer" stage (see the YouTube clip below...):

David Mohundro - Developer, Architect, Learner, Dad and more - ai tag

My entirely unasked for reflections on AI