Hacker News Reader

Comment by rcruzeiro | Hacker News Reader

Would you be comfortable sharing a bit about the kind of work you do? I’m asking because I mostly write iOS code in Swift, and I feel like AI hasn’t been all that helpful in that area. It tends to confidently spit out incorrect code that, even when it compiles, usually produces bad results and doesn’t really solve the problem I’m trying to fix.

That said, when I had to write a Terraform project for a backend earlier this year, that’s when generative AI really shined for me.

View on HN

For ios/swift the results reflect the quality of the information available to the LLM.

There is a lack of training data; Apple docs arent great or really thorough, much documentation is buried in WWDC videos and requires an understanding of how the APIs evolved over time to avoid confusion when following stackoverflow posts, which confused newcomers as well as code generators. Stackoverflow is also littered with incorrect or outdated solutions to iOS/Swift coding questions.

Cannot comment on swift but I presume training data for it might be less avaialble online. Whereas Python, what I use and in my anecdotal experience, it can produce quite decent code, and some sparks of brilliance here and there. But I use it for boilerplate code I find boring, not the core stuff. I would say as time progresses and these models get more data it may help with Swift too (though this issue may take a while cause I remember a convo with another person online who said the swift code GPT3.5 produced was bad, referencing libraries that did not exist.)

Which LLMs have you used? Everything from o3-mini has been very useful to me. Currently I use o3 and gemini-2.5 pro.

I do full stack projects, mostly Python, HTML, CSS, Javascript.

I have two decades of experience. Not just my work time during these two decades but also much of my free time. As coding is not just my work but also my passion.

So seeing my productivity double over the course of a few months is quite something.

My feeling is that it will continue to double every few months from now on. In a few years we can probably tell the AI to code full projects from scratch, no matter how complex they are.

I think LLMs are just better at Python and JS than other languages, probably because that's what they're more extensively trained on.

LLMs are better at languages that are forgiving, like those two, because if something is not exactly right the interpreter will often be able to just continue on

Depends on what one defines as a criteria for "better". Getting something to run and work, or actually writing good, readable, mostly self-explanatory, maintainable, easily testable, parallelizable, code.

better is relative, code that works but is kind of bad is still better than code that doesn't work and is also kind of bad. whether LLMs produce good code is another question

As long as you're just rebuilding what already exists, yes.

I’ve found it to be really helpful with golang.

With swift it was somewhat helpful but not nearly as much. Eventually stopped using it for swift.

Sometimes it’s PEBCAK. You have to push back on bad code and it will do better. Also not specifying the model used is a red flag.

Reply on news.ycombinator.com