I was out for 2 weeks, and the resident meat proxies on my team managed to half-destroy the project by doing everything the way AI wanted to, instead of how the company systems work, introduced like half a dozen cross-connections across barriers that should never be crossed, introduced a parallel artifact repo, permissions system, and different build tooling to what the rest use (including what the current project was built with), and managed to create a lot of drama with external teams who refused to work with these new systems.
AI still needs lots of guidance, and might seem like a genius that does everything on its own, until you give it to people who can't guide it and things begin spiraling out of control.
The same could easily have happened pre LLMs (just slower). Its why things like ArchUnit existed, or style guides, or annoying architects who stopped engineers from using latest JS framework pre LLMs, to define high level requirements for software.
I try to avoid justifying my own existence, but its a reason why I think skilled engineers are way more productive using LLM's than person off the street. Because they can confine them. Though there is no technical reason that ability can't be codified.
I don't think it'd have been likely. First, pre-LLM, people read code 90% of the time and wrote it 10% of the time - so the flow would've been to first understand the exsiting code and figure out how to interact with it, and/or asking others when stuck, rather than just plowing ahead with the favorite set of tools that particular LLM has
loading story #50483477