Retired dev here, I’m curious about the nature of “the mess”. Is it buggy AI-generated code that got into production? I know an active dev who uses ChatGTP every day, says it saves him a hell of a lot of work. What he does sounds like “vibe coding”. If you’re using AI for grunt work and keep a human is in the workflow to verify the code, I don’t see how it would differ from junior devs working under a senior. Have some companies been using poorly managed all-AI tools or what? Sorry for the long question.
An example from work a few weeks ago. I fixed some vibe coded UI code that had made it to prod. The layout of the UI was basically just meant to be an easy overview of information relevant to an item. The LLM had done everything right except it assumed a weird mix of tailwind and bootstrap, mixing and matching css classes from both. After I implemented the classes myself it went from a single column view to grids and nested grids grouping the data intuitively.
I talked with the dev who implemented it, and basically it was just something quickly cobbled together with AI until it was passable. The AI had added a lot of extra that served no function and that didn’t conform to a single css framework, but looked like it could. For months noone questioned it despite talk about that part of the UI needing a facelift.
I don’t know how representative it is, but about half the time I’m thoroughly confused about a piece of code and why it was written the way it was, the answer has turned out to be AI. And unlike when a developer wrote it, there rarely is any reason to have written it the weird way.
TBH that sounds like a lot of code I’ve seen from outsourcing companies in India. Their typical approach is to copy an existing program, module, web page or whatever and modify it as quickly as possible to turn it into what’s needed. The result is often a mishmash of irrelevant code, giant data queries that happen to retrieve some field that’s needed along with a ton of unnecessary crap, mixing frameworks, etc.
Retired dev here, I’m curious about the nature of “the mess”. Is it buggy AI-generated code that got into production? I know an active dev who uses ChatGTP every day, says it saves him a hell of a lot of work. What he does sounds like “vibe coding”. If you’re using AI for grunt work and keep a human is in the workflow to verify the code, I don’t see how it would differ from junior devs working under a senior. Have some companies been using poorly managed all-AI tools or what? Sorry for the long question.
An example from work a few weeks ago. I fixed some vibe coded UI code that had made it to prod. The layout of the UI was basically just meant to be an easy overview of information relevant to an item. The LLM had done everything right except it assumed a weird mix of tailwind and bootstrap, mixing and matching css classes from both. After I implemented the classes myself it went from a single column view to grids and nested grids grouping the data intuitively. I talked with the dev who implemented it, and basically it was just something quickly cobbled together with AI until it was passable. The AI had added a lot of extra that served no function and that didn’t conform to a single css framework, but looked like it could. For months noone questioned it despite talk about that part of the UI needing a facelift.
I don’t know how representative it is, but about half the time I’m thoroughly confused about a piece of code and why it was written the way it was, the answer has turned out to be AI. And unlike when a developer wrote it, there rarely is any reason to have written it the weird way.
TBH that sounds like a lot of code I’ve seen from outsourcing companies in India. Their typical approach is to copy an existing program, module, web page or whatever and modify it as quickly as possible to turn it into what’s needed. The result is often a mishmash of irrelevant code, giant data queries that happen to retrieve some field that’s needed along with a ton of unnecessary crap, mixing frameworks, etc.
Think of AI as a hard working, arrogant, knowledgeable, unimaginative junior intern.
The vibe coding is great for small, self contained tasks. It doesn’t scale to a codebase (yet?).