I wonder which one is more unfixable: Hallucinations since they're the result of the fundamental way LLMs work; or prompt injections because every input — no matter if it's an instruction by you, or the system prompt, or a data source — is literally just text without any semantic difference to the model so it can't "know" that a data source isn't supposed to tell it instructions.
Holographic screens in science fiction not only need to be able to project light anywhere in space, they also have to block light from behind the "screen" or you won't see shit if your environment isn't all around dark.
Which, now that I think about it, also explains who rooms are barely lit in a lot of science fiction works.