nighthawk, 6 months ago

@krzyzanowskim This is still pretty amazing though. The result makes more sense if you think of it as a two step approach on GPT’s part:

1. Extract an alt text of the image describing what’s in there.
2. Generate code from that alt text.

That’s a very lossy conversion.

jandamm, 6 months ago

@krzyzanowskim but on the other hand the GPT4 screen looks way better than the Amazon design 😅

rene, 6 months ago

@krzyzanowskim No offense, but why did you expect it to generate any kind of working code in the first place? It’s a LLM, it doesn’t know what the hell it’s generating there.

krzyzanowskim, 6 months ago

deleted_by_author

emiliopelaez, 6 months ago

@krzyzanowskim @rene I’m out of the loop, what happened?

krzyzanowskim, 6 months ago

deleted_by_author

emiliopelaez, 6 months ago

@krzyzanowskim @rene 😂

brindy, 6 months ago

@krzyzanowskim try again in 18 months. It’s still an infant.

krzyzanowskim, 6 months ago

deleted_by_author

brindy, 6 months ago

@krzyzanowskim April I think yeah, but I’m more talking about capability than knowledge. It would have done a better job of the Swift if it was able to describe that image better and that will come eventually.

Regarding Swift specifically I find it doesn’t do as well as other languages like JavaScript or python. I suspect that while the knowledge is available there’s more content weighted towards other languages so it just learns those better.

krzyzanowskim, 6 months ago

deleted_by_author

brindy, 6 months ago

@krzyzanowskim What happened when you fed that description in and simply asked it create a SwiftUI view from it? Was your screenshot the result?

I find that, like a child, you need to lead it on a journey to get to the output you want. It will definitely get better.

krzyzanowskim, 6 months ago

deleted_by_author

brindy, 6 months ago

@krzyzanowskim Looks like it has cascaded some assumptions, starting with 'this is an iPhone app'. As a human familar with app development I'd say that was a web page and would ask it to describe the UI in those terms instead. This is the kind of context that AI is missing just now: experience.

But still, what it created is not bad.

elryry, 6 months ago

@brindy @krzyzanowskim wouldn’t waste your time… looks like he works on Judo so this is a product defense thread

krzyzanowskim, 6 months ago

deleted_by_author

elryry, 6 months ago

@krzyzanowskim @brindy healthy response

krzyzanowskim, 6 months ago

then I copy and paste @judoapp generated #SwiftUI code and run it in Preview, and guess what, no flaws and pixel perfect. #iOSDev

ctietze, 6 months ago

@krzyzanowskim Wait, @judoapp can also reverse engineer a picture?

krzyzanowskim, 6 months ago

deleted_by_author

krzyzanowskim, 6 months ago

deleted_by_author

ctietze, 6 months ago

@krzyzanowskim so did you trace it then, more or less? Or did you make it in Judo first to get the ''''screenshot'''' to ChatGPT :)

krzyzanowskim, 6 months ago

deleted_by_author

ctietze, 6 months ago

@krzyzanowskim Oof, ok, I guess now I understand the joke behind this :D I took this for a genuine Amazon app screenshot

krzyzanowskim, 6 months ago

deleted_by_author

ctietze, 6 months ago

@krzyzanowskim yeah, exactly, it does!