Comment by muzani | Hacker News Reader

Comment by muzani | Hacker News Reader

"But a month later, looking back, I’d have no real memory or context for that commit."

This is the exact same feeling I got when I was coding things before AI. There's a meme we had at the office where someone runs git blame on the shitty code and realizes they wrote it. We tried to have people do tech talks about the hardest things they had to build, but most people don't even remember it after 6 months.

I think when people are in flow, they act possessed. It's not even muscle memory forgetting. They make the same mistake. They add a line that fixes Bug A but causes Bug B. They then remove the line causing Bug B, and Bug A regresses. Then they'll scratch their heads wondering why Bug A is still there.

I have the opposite experience with vibe coding. I know where the models are. I know what's in the DB, the tables, every migration, even though I focus on FE. I can tell AI where the files are, before it loads the first prompt.

I know when it's creating memory leaks, whereas I usually miss my own leaks because I'm in the code and forget obvious things like destroying a thread or coroutines running in the wrong space.

Exactly.

Plus, these are the kinds of AI posts that always make me laugh. They are like saying that calculators are bad because if you only punch in calculations and don’t think about them, you didn’t learn how to do math.

If you don’t know what your code does when you use AI, it is a poor reflection on you as a developer who didn’t understand any of the code before committing it to the project.

What makes it more humorous is that I guarantee most of those same people don’t have a clue about the code in all the packages they pull in and use, but suddenly the AI is to blame for willfully having not understood your own code.

You make a fair point about git blame—we've always forgotten our old code.

But there's a crucial difference between forgetting the "what" (the specific lines of code) and forgetting the "why" (the architectural trade-offs and business reasons). My concern is that AI-driven development accelerates the forgetting of the "why."

You should have the why ironed out well before engaging the AI. Draw your UI, model your architecture, etc. If you must have the what before the why, that's perfectly normal and agile, but do it after the LLM is done type.

If you do want to use a LLM for brainstorming, design, etc, it should be done before the coding. Like humans, they're not great at doing both at the same time.

The models we have now are just very poor with setting up things from scratch. I think vibe code leans to frameworks like Expo because these are a little more stable without needing any architectural work, but they're rigid and limited.

In mobile, we always use a pattern like view-viewmodel-repository-service. There's little reason to use less for an app with lots of async (e.g. web apps). There's also patterns like design systems and putting all your themes and colors in one box. But the AI will almost never set up this pattern. It's well aware of it and it'll seamlessly when you tell it to do it. But you have to set it up first.

Reply on news.ycombinator.com