[EerkeBoiten]

Author here. Fair enough on my industry experience. But I hope components, unit testing, regression testing, etc aren't as easily dismissed in real SE environments - no trouble believing formal methods and verification are off the radar. The article is not about using AI to write code (which may work to some level of satisfaction for some people) but about using AI as code.

[Maro]

Thanks for your reply! Hope my comment wasn't offensive.

I definitely think components, unit testing, regression testing are good things and are done at good software houses. In my experience however, most of these things are mostly cargo culted at best in a many other environments.

When I wrote my comment I was wondering about the "AI to write code" vs "AI as code" point. In my vocabulary, "AI as code" would be "Data Science models", like a ranking engine for ads in a newsfeed? I certainly understand the idea of having an AI "emulate" an application like Word.exe or Doom.exe, and there's been research into this direction, but as far as I can tell that is not the general direction the industry is headed in --- rather it's the "AI to write code" direction.

[EerkeBoiten]

Thanks. In a broader sense, with "AI as code" I mean any situation where we ask an AI model for answers or decisions where we otherwise might have written a program to solve it. See also "LLM functionalism" in the article. Particularly where we need to rely on the outcome - so not "predictions", "suggestions", or "recommendations" all of which we expect to have limited reliability which we mitigate through modifying or ignoring.