I think you underestimate what can be accomplished with human level agi. Human level agi could mean 1 million Von Neumann level intelligences cranking 24/7 on humanity's problems.
One million Von Neumanns working on that 'problem' is not something I'm looking forward to.
That is a hard limit on intelligence, but neural networks can't even reach that. What is the actual limit? No one knows. Maybe it's something relatively close to that, modulo physical constraints. Maybe it's right above the maximum human intelligence (and evolution managed to converge to a near optimal architecture). No one knows.
As far as we know, it's impossible to model any part of the universe perfectly, although it usually doesn't matter.