> bad engineers were always a liability This part of the article hits home for me. With AI, "bad" engineers can now amplify their "bad" engineering x10 across the organization. The most egregious of these cases for me is often long tenured engineers who have lost interest in the craft, creating a dangerous combination of having enough merit to ship but not enough interest to make what they ship _good_. I am still a f…
Absolutely on point. I just completed the port of a industrial application written in Python using a sophisticated console-based UI to C# and Avalonia UI. This is my second major coding project using Codex.
The one salient element of the experience has been that, as I keep saying to anyone who will listen, AI still does not understand anything it is doing. It presents an amazing simulation of it, but, no, it does not.
Here's a simple example: The original application had a clean communications protocol implementation. A single file with a class that implemented every single command you could have the industrial controller issue to the hardware. Codex was explicitly told to replicate that in C#. It did, at first, but then it started to duplicate command processor code in the individual functional blocks throughout the application. Which means that, when a bug surfaced, you had to fix it in five different places.
Another example: The application has a highly customized table view. Once again, it was told to make that a component to reuse throughout the application. It did, at first, and then weird bugs started to surface that made it clear that it had reimplemented the table control in different areas of the application.
If you don't know what you are doing you will probably not pick-up or even care about some of these things. It is easier to whack-a-mole bugs with AI than to worry about code structure, efficiency, maintainability, future-proofing, scalability, etc.
I purposely decided not to look or touch a single line of code during this project to see what's possible and where the issues might be. I learned a lot and continue to learn. I think the next step is some sort of an agentic approach, maybe using OpenClaw (or whatever, I don't really know right now) to create a team with coding, supervisory and testing agents.
While the project got done significantly faster than it would have without AI (four weeks instead of probably 4 to 6 months), the process was just as intense as coding, just operating at a different level, micromanaging architecture and implementation.