As long as not all the memory is actually used, you can pretend that there is 20GB though. Once you actually need to map more than there is, you would need to s

As long as not all the memory is actually used, you can pretend that there is 20GB though. Once you actually need to map more than there is, you would need to stop the virtual machine. If you read e.g. the Bostrom 2003 paper, you will see that this is more or less what they claim: If your simulation technique is good enough, you can just fake most of the universe and only actually perform fine-grained simulations where it matters. This would e.g. assume that you have special simulation rules for brains, large scale chemical reactions etc., which allows you to just skip most of the details and make them up once you need them. You could even automatically detect whether your simulation is fine grained enough by running the same simulation at two different granularities at the same time, readjusting whenever the outcome between the two differs too much. I think the argument in its original form does not make too much sense (future humans simulating todays humans for research and fun). What you would probably simulate much more, are human-like entities in a much simpler universe (e.g. no weird quantum-mechanics that are really hard to simulate). So in my opinion, you could take away from the simulation argument that there is a non-zero chance that we are living in a simulation. But the computer that performs the simulation most probably lives in a universe that does not follow the same rules our universe does. It might be a universe with two or more time-like dimensions. It might be one where quantum mechanics are just super-easy to simulate for some reason (maybe in the same way it would be really easy to e.g. build Conways game of life in transistor hardware for us).