Hacker News Reader

Comment by Gigachad | Hacker News Reader

The old way was not really that you would spend 2 hours googling for the exact procedure you need, but that you would spend time following getting started tutorials, following along with example projects, and generally learn the entire program. Then when you go to make something specific, you'd be able to come up with the steps required yourself because you know the program.

I'm not discounting the value of having ChatGPT just hand you the answer straight up. If you just want to get the task done as fast as possible that's a pretty cool option that didn't used to exist. But the old way wasn't really worse.

View on HN

The problem is that stacking up these solutions creates technical debt. This technical debt needs to be solved by understanding the code base and fixing the issues. Can an ai coding agent do that? Sure maybe. But what I find is that ai coding agents need your help to do pretty much everything. So you need to have a good understanding of the code base in order to be helpful. So eventually you sit and wonder if vibe coding a stack of tools was actually helpful. Don’t get me wrong I think Claude code is amazing, but it’s not “And therefore there no longer needs to be jobs” amazing.

> following along with example projects

What the LLM gives you is essentially an example project, and you can ask for the specific examples you need. You can compare and contrast alternative ways of doing it. You can give it an example in terms of what you already know, and ask it to show you how that translates into whatever you're trying to learn. You don't have to just blindly take what it produces and use it unread.

But then it wouldn't be vibe coding! /s

It's endlessly mind-boggling to me how there's so many people who can't grasp the idea of just using llms as a tool in your engineering toolkit, and that you should still be responsible, thoughtful, and do code review - as you would if you delegated to a junior dev (or anyone!)

They see complete fools just accepting the output wholesale, and somehow extrapolate that to thinking everyone works that way

This is why EXAMPLES in getting started pages are so important.

LLMs are making up for the lack of this.

It’s the Backus-Naur approach vs the Human approach.

Humans learn by example. IMHO this is why math education (and most software documentations) fails so hard - starting with axioms instead of examples.

Reply on news.ycombinator.com