4 ms·
The perspective of the article seems to be of a person who has not worked in cookie-cutter software engineering environments, which is what 99% of software engi
by Maro 2y ago
The perspective of the article seems to be of a person who has not worked in cookie-cutter software engineering environments, which is what 99% of software engineering is. Here formal methods and verification are irrelevant, nobody has heard of it, and nobody cares. It's about churning out some crappy internal webapp or mobile app or data eng/science pipeline at the lowest cost by some soulless bigco (or startup). LLMs are already super useful for this.
Also, the argument about LLMs being black-box is a miss, because LLMs are writing software, they are not the software; programmers writing software are also black boxes. Also, there's nothing in the way of running formal methods on the software produced by an LLM, it will fail the same as it fails on software written by humans, since most formal verification doesn't even make sense for 99% of software.
Also, anybody who has used LLMs as aids for writing simple/smaller chunks of code (and other documents) knows that they're super useful (sometimes magic). It's like Steve Ballmer saying the iphone is a joke in 2007.
- EerkeBoiten 2y agoAuthor 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 2y agoThanks 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 2y agoThanks. 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.