I don't find this observation of Djikstra's to be one of his best. If there is a gadget that does a thing that no other gadget does, what does it even mean for the interface to be "terrible?" How can you even know if the interface is terrible, given that a better one has yet to be invented? Maybe the interface is as good as it can be for the tool in question.

I also don't love your mapping of this observation onto modern LLMs. The interface of an LLM is natural language text, along with some files written in plain text or markdown. Can it be improved? Undoubtedly! But as a baseline, it doesn't seem half bad to me. If it is so terrible, it should not be hard to propose an interface that will be significantly more productive. Can you?