Jared Norman (@jardo.dev)

interesting to see some effort to actually quantify the effects of LLM use in software development. this mostly fits what I've seen, though I've also seen some real standout bad uses where developers were able to amplify their terrible decisions www.linkedin.com/posts/dave-...

loading . . .

#softwareengineering #ai #devops #continuousdelivery #codequality | Dave Farley | 17 comments Is AI creating a maintenance nightmare? We actually measured it. We hear constantly that AI coding tools (Copilot, ChatGPT, etc.) make us "faster" and "more efficient." But most of these claims rely on naivity. They measure typing speed or feature count short-term metrics that ignore the economic reality of software development. The dominant cost in software is MAINTAINING code. Maintenance represents 50-80% of the total cost of ownership. If AI helps you type faster but creates "slop" that is impossible to change later, we are optimizing for the wrong thing. We decided to stop guessing. I contributed to a controlled experiment involving 150 participants. We split the study into two phases: 1. Creation: Developers wrote code (some with AI, some without). 2. Maintenance: A different group of developers had to evolve that code, without knowing if it was AI-generated or human-written. The results were not what many of us expected. Here is what the data showed: • The "Slop" Myth: The https://www.linkedin.com/posts/dave-farley-a67927_softwareengineering-ai-devops-activity-7426940209507840000-BJSL