I agree. But there's still a valid use here for LLMs: If I have 1 Million bits of information and I want to communicate a synthesis of 500 bits to you, LLMs can be a viable helper. It is rarely done that way - I agree, but the way I use LLMs: Drop in all the relevant context information (PDFs, HTML, Markdown, Pictures etc. - up to usually 200-300k tokens), then compile a synthesis/summary prompt (usually 1-2 A4 pages of handwritten text. Then copy the output (1-2 A4 Pages), manually edit and send off (often, including an archive of the full original conversation, so the human on the other side can consult the unfiltered original prompt + inferencing, if wanted).

In other words: If I didn't reduce the 500 pages of text for you, you wouldn't know what I mean or what is relevant, or how to filter it yourself.