This and parent are both approaching toward what I see as the main obstacle, that we as a species don't know how in its entirety a human mind thinks (and it varies among people), so trying to "model" it and reproduce it is reduced to a game of black-boxing. We black box the mind in terms of what situations its been seen to be in and how it has performed, the millions of correlative inputs/outputs are the training data. Yet, since we don't know the fullness of the interior we can only see its outputs it becomes somewhat of a Plato's cave situation. We believe it 'thinks' this way but again we cannot empirically say it performed a task a certain way, so unlike most other engineering problems, we are grasping at straws while trying to reconstruct it. This doesn't not mean that a human mind's inner-workings can't ever be %100 reproduced, but not until we know it further.

And there is another important difference: Our environments have oodles of details that inform us, while LLM training data is just “everything humans have ever written”. Those are completely different things. And LLMs have no concept of facts, only statements about facts in their training data that may or may not be true.