Yes, what if instead of the computer being an Internet Communications Device (as Steve Jobs called the iPhone), it would just pretend to allow us to communicate with other humans while actually trapping us in a false reality, as if we were all in the Truman Show?
It might work, as indicated by the results in your link ("Human raters are only slightly better than random chance at distinguishing short clips of the game from clips of the simulation."), but the result would be a horrific dystopian nightmare, so why would we do this to ourselves?
Anyway, there is one aspect where the STEPS work is similar to this idea, in that it tries to build a more concise model of the system. But it does this using domain-specific languages rather than lossy blobs of model weights, so the result is (ideally) the complete opposite of what you proposed: A less blobby, more transparent and more comprehensible expression of what the system does.
We already interact with false reality through our "old school" computers – internet is full of bots arguing with each other and with us. But my proposition doesn't have to distort the interpreted content.
Neural nets (RNNs) are Turing-complete, so they can simulate web browser 1:1. In theory, of course. Let say we find a way to train a neural net to identically simulate web browser. The weights of this blob might at first seem like an opaque non-sense, but in reality it would/could contain a more efficient implementation than whatever we have came up with.
Alan Kay believed computer science should take its cues from biology. Rather than constructing software like static buildings, we ought to cultivate it like living, evolving organisms.
It might work, as indicated by the results in your link ("Human raters are only slightly better than random chance at distinguishing short clips of the game from clips of the simulation."), but the result would be a horrific dystopian nightmare, so why would we do this to ourselves?
Anyway, there is one aspect where the STEPS work is similar to this idea, in that it tries to build a more concise model of the system. But it does this using domain-specific languages rather than lossy blobs of model weights, so the result is (ideally) the complete opposite of what you proposed: A less blobby, more transparent and more comprehensible expression of what the system does.