Generative UI is unsolved because current models do not have taste and end up generating the same kind of slop.
I'm amazed that this blog doesn't even have a single screenshot/photo of the kind of UI they can generate.
Focusing on benchmarks in this domain feels very wrong.