Hacker News Reader

Comment by hajile | Hacker News Reader

The fact that they could be "vibe coded" means that they (or something similar) probably already existed, but you just didn't know they did.

View on HN

No, there's a very very long tail of tasks that are too niche for a proper tool to exist for but it would be annoying effort to figure out yourself.

My most recent example of this: write a script to delete all git branches that point to HEAD.

Go ahead see how long it takes you to figure out the command for that, because it took me under 60 seconds with AI.

This is the learning to fish problem. If you understand git well, a solution isn't very hard (eg, rev-parse each branch and compare to the rev-parse of the current branch and deleting if they match). If you don't understand git well, you shouldn't be turning into a script-kiddy pasting in whatever the AI spits out and hoping that it works.

Your method still requires a script though

You are using "script-kiddy" as an ad hominem rather than making an argument with proper justification. "script-kiddy" refers to website vandalism and cracking. That has nothing to do with using AI to learn what rev-parse does.

Learning about `git rev-parse` through documentation and learning about `git rev-parse` through AI fundamentally have the same outcome at the end of the day: you have learned how to use `git rev-parse`.

> Learning about `git rev-parse` through documentation and learning about `git rev-parse` through AI fundamentally have the same outcome at the end of the day: you have learned how to use `git rev-parse`.

But with a non-zero chance of hallucination.

If you learn many things that way the chance grows to 100%.

I’m using it as a pejorative (not an ad hominem) for people who blindly trust unknown code because they don’t have any real understanding. It’s a shortcut for a whole, well-known and obvious argument about the danger of playing with things being your understanding.

http://www.catb.org/jargon/html/S/script-kiddies.html

If you are using AI to learn, understand, and verify what it spit out, by definition, you aren’t a script kiddy. My argument was about how you use AI rather than a commentary on if you should.

It is an ad hominem. "you shouldn't be turning into a script-kiddy" is guilt by association, see Wikipedia[1] or ask AI[2]. If you think that learning from AI is dangerous you should articulate why you think that. I find it neither "well-known" or "obvious".

[1]: https://en.wikipedia.org/wiki/Ad_hominem#Guilt_by_associatio... [2]: https://chatgpt.com/share/6817b741-8a80-800c-8d5d-1b315806dd...

> eg, rev-parse each branch and compare to the rev-parse of the current branch and deleting if they match

Yeah now go and do that in under 60 seconds.

You seem to be a bit confused. I understand Git very well. But AI can simply do this much faster than I could.

IDK, there are a lot of little chores that I find needing done that previously would have been ten minutes of research and cli fiddling that are a prompt and copy paste now.

It doesn't feel that different but it is a little faster

If you don't understand what the copy/paste is actually doing, you shouldn't be script-kiddying it in your project. If you do understand, then the AI is just quickly writing what you already intended.

If you are having to understand something you didn't understand, it's probably taking a bunch of time to read and verify what the script does. This can be a good learning experience and reveal unknown unknowns, but probably isn't a massive speedup.

This seems false to me. I vibe code shell scripts all the time. I understand what it's doing perfectly, but it would have taken me ages to look up all the bash syntax and get it all correct. Is it one square bracket pair around if statements, or two?

You are in the category I mentioned that already know the answer and are using AI to write the code you already understand faster. This is rather different from “vibe coding” in my opinion.

It's different by the canonical definition - you aren't supposed to review the code when "vibe coding", only the app behavior.

But the amount of time saved because one could vibe code it rather than search through Google’s terrible results is immense and a benefit in and of itself.

Maybe. But anytime I take on the task to automate something, it was a nice opportunity to check how others do it, what other things the program can accomplish.

The decision to automate is already done, no need to rush through it.

Parts of it could probably already exists in the LLM's corpus, but being able to join them together to build new things is what makes "vibe coding" so useful.

Even though, from personal experience, at scale it still falls apart

I think it is somewhat similar to the emergent translation capabilities, it doesn't need to have trained on a translation of a text to do one of it. But it does need good coverage of the area.

What the end product is depends on the specific parts chosen and the way they're put together

Just because the parts are available doesn't mean every possible combination of them has already been made

Not if you know the field really well, I talk to my users.

Reply on news.ycombinator.com