There is currently no knowledge nor progress for what a cure for cancer, and nothing a LLM can draw upon.

You could generate "pregnant Elon Musk with four arms and three eyes doing yoga poses" because the image models have enough visual concepts of each of those individual things, but that specific image is (likely) not in any training dataset.

What I'm saying is if this thing can generate random things (noise) couldn't it make that or new tech like negative mass. Anyway I get it too if we don't know then something we made wouldn't know.