AI Got Better While I Was Away. Software Didn't.

It has been a while since I last wrote, not because I ran out of opinions, but because the endless stream of AI-related debates about developers being replaced or not became tiresome. However, AI is very much active and has improved significantly. The quality of coding agents and generated code has vastly increased compared to just a year ago. Yet, despite this progress, there are still issues like overengineered backends, unnecessary abstractions, and features nobody asked for.

It seems that code generation was never the bottleneck, but rather we may have wanted it to be so. Developers often focus on coding as if it's where all the value lies, when in reality, the hardest part is deciding what should exist, if it's the right solution, and whether we need another layer or could delete something instead. While AI can generate 600 lines of respectable-looking code in a short time, that doesn't mean those lines should exist.

AI can help us overengineer, which sounds great, but friction sometimes prevents us from making terrible decisions. With AI, it's easier to construct complicated systems without suffering. It's very good at helping us overengineer, often producing more reasonable decisions than we can regret. As AI gets better, the more I prefer boring software - databases, APIs, frameworks, and deployment that are obvious and easy to understand. I don't need my stack to be interesting; I just need the product to be interesting.

Code generation is becoming cheap, and while it doesn't make developers worthless, the valuable part of development is shifting towards judgment, architecture, debugging, product thinking, and taste. This is a skill that's hard to measure but becomes crucial in an AI-heavy world. Those who excel in this area will do well.

We don't have a code shortage; we have a shortage of good decisions. AI provides more output, implementations, and options, which is great, but more isn't always better. Sometimes, more just means more. Developers need to get better at filtering what gets generated to avoid creating the most mediocre software in history efficiently.

I use AI for various tasks like coding, debugging, research, writing, idea exploration, and challenging my solutions. I don't want to stop using it, but I'm less impressed by AI writing entire features now. It's cool, but was it the right feature? Is the code maintainable? Did you understand what it generated? Did it add unnecessary complexity?

Those questions matter more now. The job might not have been writing code all along; it could have been making decisions under uncertainty, understanding systems, making tradeoffs, and finding the simplest solution that works. Occasionally, it even meant telling everyone: No, we don't need Kubernetes for this. AI can generate more software than we can manually write, but the real question now is whether we're getting better at deciding what software deserves to exist. I'm not convinced we are, and that might be a bigger problem than AI replacing programmers.