Hacker News Reader

Comment by _bin_ | Hacker News Reader

Semi-controversial headline but I think the point is fair. It doesn't yet produce results that meet a spec well. It's fine for scripts because if it insists you do it the bot's way, that's fine. If it uses the wrong library, whatever. And so on.

I'm going to keep shilling this until people start doing it: I'd love to see everyone who says he's vibecoded something non-trivial and posts/blogs about it start dumping and attaching his full chats to it. This would really help people to understand what's hype/exaggeration and what's real, and would help others be better at the real part.

View on HN

The one time so far that I actually implemented a full feature using only vibe coding (and it wasn't trivial), I had to be very specific about what I wanted, and it was in a very widely used stack (React/NextJS), it was a very common feature (sorting, filtering, pagination on a table) and it worked beautifully.

But it felt magical the one time it happened. Every other time, I'm mostly using tab autocomplete and that works quite well too.

And when I code in Elixir, I have a 20% success rate even with small requests.

Your last sentence is a bit of a gut punch tbh - I was meaning to get more familiar with Elixir for side projects, and I’ve made a few small web apps with LiveViews. Absolutely love it, even though I don’t feel I‘ve grasped the language fully yet. But some research shows that most LLMs don’t support Elixir that well.

But since my free time is so extremely limited, it would be really nice to get better LLM and agent support. So I hope the lesser used languages and frameworks get more training love, soon. Lest we all only use Next for hobby projects in the future.

Unfortunately this is consistent for non-ultra-mainstream langs. They’re next to unusable for complex Haskell. If you want to vibe code, it’s Python/typescript/react or nothing.

I wonder how well they do for Go and Java. The first is a boring language where there's little to none chance to stray from the default. The second is verbose, but with strict rules on the syntax.

They are fine with Go. Not with whole codebases but libraries/functions around a thousand lines of code are written perfectly with tests (Although, I always say, never let the LLM write the tests AND the code, or at least write your own tests in addition to what it generates).

Elixir has been my main language since 2017 and it’s been amazing to work with all these years. I’ve built and delivered so much value and the quality of libraries in general feels much more stable than JS land with its ever changing upgrades. I’ve never had to worry about leaving a code base untouched for 6 months. Whereas in JS land I’d have to groan about half the stuff not working, some new transpiler breaking changes, blah blah and just give up.

But developer ergonomics rough edges mostly around LiveView, and partially language servers and type support (which recently had some developments) and bad LLM support (no fix yet) are finally causing me to question whether I should look elsewhere.

The underlying tech (scalability, echo, etc) has still been excellent.

Honestly, I'd strongly suggest not using an LLM to get started on a new language. It'll be harder at first, but you'll build intuition about how things work under the covers (especially important when the runtime environment is unique like the Erlang OTP), and you'll build your ability to code idiomatically.

Instead, read lots of great code in that language and incrementally work to understand and imitate its style. The standard library and any popular tools written in it would be good places to start.

Reply on news.ycombinator.com