Alan Kay's answer to Can an “AI command line” replace the GU…

Some interesting thoughts, but I think "GUI" is too coarse a term to be useful. A typical GUI consists of a workspace (say, a document or an image) and a bunch of tools/affordances for altering it. Photoshop and MS Word have dozens of buttons and controls affecting what you can do.

LLM-based UIs make the giant toolbars mostly obsolete, because you can just say what you want. Of course you still want the domain object in the workspace to be displayed graphically if appropriate.

• The "mostly" is because you still want some way to convey to the user the list of possible actions. But there are better ways than toolbars and such.

There's a lot of griping about the importance of undo and its absence in most modern UIs. I agree with this. I think the reason is because most modern UIs are not about exploring a design space (where undo lets you try things out without risk); they are about social or other transactions. We have git branches, which is sort of serving the same exploratory goal at a coarser and less interactive level.

OK, this has convinced me that whar the world needs now is an explorable design space manager, for any kind of object. You can undo, try experiments, generate 10 random alternatves and see the consequences. LLMs can be used for control but the main purpose is to make the space of possibilities legible to the human user.