These models don't have a frame of reference to ground themselves in reality with, so they don't really have a base "truth". Everything is equally valid if it is likely.
A human in a hallucinogenic state could hallucinate a lot of things that are true. The hallucination can feature real characters and places, and could happen to follow the normal rules of physics, but they are not guaranteed to do so. And since the individual has essentially become detached from reality, they have no way of knowing which is which.
It's not a perfect analogy, but it helps with understanding that the model "writing things that aren't true" is not some statistical quirk or bug that can be solved with a bandaid, but rather is fundamental to the models themselves. In fact, it might be more truthful to say that the models are always making things up, but that often the things they are making up happen to be true and/or useful.
Precisely, the model is just regurgitating and pattern matching using a large enough training set where the outputs happen to look factual or logical. Meaning that we're just anthropomorphizing these concepts onto statical models, so it's not much different than Jesus Toast.
I think this is a great way to think about it. Hallucinations are the default and an LLM app is one that channels hallucinations rather than avoids them.
A human in a hallucinogenic state could hallucinate a lot of things that are true. The hallucination can feature real characters and places, and could happen to follow the normal rules of physics, but they are not guaranteed to do so. And since the individual has essentially become detached from reality, they have no way of knowing which is which.
It's not a perfect analogy, but it helps with understanding that the model "writing things that aren't true" is not some statistical quirk or bug that can be solved with a bandaid, but rather is fundamental to the models themselves. In fact, it might be more truthful to say that the models are always making things up, but that often the things they are making up happen to be true and/or useful.