The macro characteristics of entity dynamics in the world is that of a sparse metastable graph. There is considerable compression possible at the computational level, since most entities in the world are fungible in every discernible aspect, but the data model is not similarly reducible because what can be compressed away is context dependent. Since the computation is less tractable part, this makes the modeling problem eminently viable albeit still extraordinarily difficult. To model the effects of the relatively small number of nodes and entities that can shift the equilibrium you need to compute over the true graph, not some aggregate and summarized roll-up of the graph. There is a real tension between anonymization/privacy (and reducing computational costs) and the ability to build models of human systems that remotely reflect real-world dynamics. Models based on anonymized aggregations are notoriously rubbish when ground-truthed.
It is also important to treat these models as sparse metastable graphs. I am always slightly horrified by the number of serious academic disciplines that often default to modeling human systems as if they were a computational fluid dynamics (CFD) problem. Global human systems don’t have CFD-like structure but CFD is attractive because it is easy to reason about computationally and widely understood. A lot of scientific research on models of human dynamics (e.g. epidemiology) that were ultimately debunked had a root error in the use of CFD-like assumptions about how humans interact and behave.
It is technically possible to build a far higher fidelity model of physical world dynamics than I think most people would expect or intuit. Unfortunately, that is not the hardest part of making it useful.
To take a semi-long term example. If you go carve your name in a tree, when you come back 50 years later it better show your name distorted over those 50 years. If you cut open the tree. It better show a record of every weather event over those 50 years. Every piece of the world you inspect will need the simulation run and not just for it but for all the things that affected it and all the things that affected the things that affected it etc etc recursively until you've simulated every in the universe.
Minecraft is an example of the latter: if you leave a chunk, mobs despawn and decay and growth stops.
But a simulation need not work that way. E.g. instead of continuing to run a simulation, it can simulate macro-indocators, and continue to simulate micro-details with a strong bias towards reverting to a prediction from the macro-indicators, and each element that gets there can be discarded.
To go back to the Minecraft example, you set a "decay rate" for mobs that reflects a believable rate of dispersal, or simulate changes to vegetation towards what your generating function will produce until they match.
Then when the player re-enters at t+a_lot, you just run your generating function where time is one parameter, and apply whatever reduced delta is still relevant.
Yes, you'd still need to simulate a lot, but you'd be able to prune most things.
So the question is meaningless without defining your requirements.