Well, the issue is that LLMs do not support real time learning at all. If they would be able real time learn and use the base data from training data, I suppose they can understand a physic book even better than normal human with reading it once.

A Human without pre training is not able to understand a physic book without help. He would even be able to read.

If someone finds a way to train LLM realtime and have it decide with what weight each new training data is to be interpreted, I see all that above possible.

And of course if humanity ever creates something that behaves like AGI, humanity would not be able to tell if it is emulated AGI or real AGI. There is no known method to differentiate those two.