I've been thinking there's some parallels between how AIs hallucinate and how human toddlers do. If you ask a toddler/young child a question about a fact they don't know, they will usually say, "iuno" (even when they should), but depending on the child and the circumstances, they will sometimes just make up a story on the spot and sound as if they believe it. "Who invented ice cream?" "Santa Claus! Mommy left him milk and cookies and he turned it into ice cream." It doesn't make any real sense but it seems facially plausible in their universe.
But somewhere between first learning to speak and around 7ish, kids become markedly more accurate how they model the world, and their responses become correspondingly less fanciful. And they continue to improve beyond that point.
So how are kids doing what LLMs are currently incapable of? How do we teach ourselves not to hallucinate? Or do we, really? I mean, if I tell myself I'm going to make it through the intersection before the light turns red, but I end up running the red light, was I just mistaken, or was that a self-delusion, i.e., a mini-hallucination of sorts? Probably a self-driving car would be less likely to make that category of mistake, so maybe I shouldn't be so smug about being grounded in reality.
Obviously it's worth it to try and eliminate the incorrect information, but what grand-op is saying is we don't want to do that if it takes away some valuable emergent properties.