It's like the 90/10 problem, right? LLMs today do things that are pretty amazing things but they are missing something that seems incredibly simple to laymen bu

It's like the 90/10 problem, right? LLMs today do things that are pretty amazing things but they are missing something that seems incredibly simple to laymen but is and has always been the actual hard problem. I think the comparison to self driving cars in incredibly apt. A self driving car can do incredible things at superhuman speeds but the simple things it can't figure out are actually really hard problem. For instance, seeing a floating grocery bag and knowing that it's floating in the wind with negligible mass that doesn't need to be slowed down for or avoided. Something that actually requires an actual understanding of the world, rather than just being trained on the past. A similar situation for LLMs seems to be these "hallucinations". The laymen see the vast knowledge that they seem to hold and thinks, just tell it to stop lying and it's fixed. The problem is that telling the model not to lie is really only simple if it has some understanding which LLMs don't seem to possess. Their hallucinations, to them, are just as true as when they happen to tell the the true. They have no method of discerning them.