A chat interface won’t replace UIs but give LLMs access to a component library, and it can piece them together to form an interface.
Since ChatGPT plugins can call apis with no issues, this capability should already exists.
What percentage of the time does an interface failing make it unusable? Or even worse silently failing? In most cases, I suspect that number is pretty close to 0. Given LLM’s hallucinate a non trivial amount of the time, we seem quite far away from this. Or at-least until we have AGI.
It doesn't have to write code for each button in a UI, it just has to know that it's showing a UI and you click on a button titled "increase quantity" or whatever and the changes can be handled by the model itself.