Hacker News new | ask | show | jobs
by jmull 701 days ago
> But you know what AI will be great at? > ...Turning wireframes into frontend code (limited logic) > ...Wielding (and extrapolating) your design system > ...Creating beautiful visuals from screenshots and mood boards

I don't think the argument works. It imagines AI will be able to do things that our current AI do not.

Our current AI work in terms of what they've been trained on. Natural language to code works as well as it does because coders have been asking and answering questions about code a lot on the internet, that's been collected into data sets, and those data sets have been used to train LLMs.

For AIs that work like our current ones to be great at the things the author envisions, they will need a lot of examples of those things, and of a high quality too.

The author wants AI to turn natural language into wire frames, let the designer adjust and improve the wireframe, and then turn the improved wireframe into production code.

But I don't think the training data for those two transformations exist. What even is the intermediate wireframe format? AU will produce and consume it, but it also needs to be in a form human designers can iterate on.

Since the training data doesn't exist, it would need to be created. Is that feasible? If you can't harvest the internet for it for almost nothing, it would cost a great deal to create, and I wonder about the quality of the result.

1 comments

Frankly, I am not even sure what "frontend code" means in this context. What stack? How is state managed? How is this code integrated into the larger production system? If by "frontend code" they mean fairly clean React code - that's just a small subset of what is actually needed for production.
Exactly. It's a dreamland where frontend styling is completely decoupled from logic and behaviour. Sounds good for a 2005 web designer.