Ask HN: Is vibe coding overhyped or am I missing something? | Hacker News Reader

Ask HN: Is vibe coding overhyped or am I missing something?

3 pointstom13376 comments

In the recent months / weeks I'll ready so many headlines about "AI will be able to create whole applications easily", "No coders needed, everything will be AI", "Vibe Coding is the future" so I thought I'd try it out. I've been building React Apps and TypeScript Backends for a few years so I am not a total newbie. I wanted to create a simple React Native App which has a Bottom Sheet and display a map. Already at the first step - creating the map - ChatGPT o3 gave me completely invalid code. It used a pretty old version of the library which wasn't compatible with the latest react release. It also got few things with expo vs. "pure" react native wrong. Sometimes the code it threw at me included unused styles or variables.

Am I missing something? How are people seemingly building complex applications if the models struggle with basic things like keeping up to date with releases? What are your experiences?

You aren't missing a thing. LLMs produce code that is somewhat correct, but wrong enough that it needs correction. The people promoting "vibe coding" are saying that you can just tell it the concerns, and eventually it will land on something that works. Everyone else codes the corrections themselves. But AI simply does not produce working code without human interventions.

It also does not keep up to date with releases. That is not a valid expectation, and if you are working with leading-edge releases and libraries, LLMs will not help. On the same note, it cannot innovate - it will give you code that fits common conventions. Most of the time this works, but if you are doing truly novel work, its results will frustrate you more often than not.

I have been using AI assistants to help with a project that uses MUI for React. We're using a version of MUI that is a few versions old. It is hugely helpful at figuring out what kind of sx to put on an element or what to put in a theme to get things to look right. It makes mistakes, but if you tell it what you see in the F12 tools, read the manual and think a bit, it does pretty well.

It loves to use deprecated APIs, for instance there are a lot of props that you aren't supposed to use anymore and you're supposed to use the "slots" prop instead. It's not hard to change the code and you could probably ask them to do it, but it's the kind of problem I run into.

With SQLAlchemy I find assistants cannot make up their mind what version of SQLAlchemy they are generating code for. Between talking to my assistants and reading the manual I understand a lot about what changed in different versions, but what I really want is to have it look at my POM/packages.json/pyproject.toml and always give me advice for the versions I am using.

This isn't just an LLM problem, one of the "product ideas I don't have time to work on" in my file is a search engine for programmers that indexes the documentation for the actual versions of all your dependencies, so you are not stuck with Google's brain damaged behavior of returning results from random JDK versions. For instance, "String Javadoc" returns results for JDK 8 and JDK 11 (end of life!),

https://www.google.com/search?client=firefox-b-1-e&channel=e...

whereas a lot of us are developing to JDK 21.

With AI coding it makes projects even faster to get started. Like anything else, total cost will be dominated by long tail problems. Vibe coding can't solve the longest tail problem of "am I building the right thing?" These interstitial problems of code quality you mention will be worked out I think, but it won't matter. To put it the opposite way: if I already know in advance what I need to build to be "successful", the cost of getting code to production was already a minor concern.

That being said, the speed at which I can explore solutions with it is greatly improved. This has a big effect on my motivation to start ideas. In some domains, simply getting more times at bat is the critical path.

CNBC had a bit about "vibe coding" this morning so its officially lamestream. lol

I use roocode and auto-approve most tasks and then I do a code review at the end. It saves a ton of time but costs a ton of money too

You're right, but missing some ideas. Sure, it will use an older library sometimes, leave unused variables or make mistakes. But if you're really vibe coding, the question is "if I respond that this doesn't work, will it be fixed?"

If it does get fixed, who cares about the variables or mixed patterns? Vibe coding at this stage is expected to give running code, not clean and maintainable one. And it often does, but it's not there for complex applications yet.

Vibe coding isn't over-hyped, it is a myth. :)

Reply on news.ycombinator.com