LLMs do solve the discoverability problem which underpins every power user feature ever. It is fantastic to be able to describe what you want without the domain knowledge that was previously a prerequisite to achieving that goal. Hopefully as more systems create declarative ways to interact with them (MCP, clis), power users will be able to profit from this as a side effect as well.
They sidestep it (partially), but I don't think they solve it. Sure, you'll get your task done if the agent does it for you, but the agent won't show you how to do it without their help. So the next time you have to do the task, you have to ask them again.
I also think this is misunderstanding "discoverability" a bit. For me "discoverability" also means you get an understanding of what options are available at all in the system, which can be helpful to get a better idea of what you want to do in the first place.
Seach and chat-based UIs always assume you already have a perfectly thought-out plan what to do (conveniently in terms if the software you're working with) and just have to type it in.