I think this idea of "generative UI" stems from a genuine confusion about what Spotlight, Alfred, or Chrome Omnibox are: they are a collection of utilities, not an app of themselves.<p>They work because a designer has put some thought into anticipating what kinds of things users search for – formulae, definitions, maps, flights, timers – and designed tiny, polished experiences for each – important in their own right, but not enough to be standalone apps. The NLP component only acts as a dispatcher, calling the right intent with the right arguments.<p>If these micro-apps are untested unpolished, and most likely broken, having been created here and there without a human touch, they will likely prove inadequate for the user. And if the user is now forced to debug or extend them with natural language and no understanding/control of the underlying code, allowing the power users to write community plugins will have become preferable to this. Not to mention the latency and unpredictability which defeat the standard quick search model.