I often use the (equivalent) argument "how could Mario escape, even if he somehow 'knew' he were in a simulation?" to help illustrate why it may not be possible

I often use the (equivalent) argument "how could Mario escape, even if he somehow 'knew' he were in a simulation?" to help illustrate why it may not be possible to escape even if we are in a simulation -- but I think this only covers a range of possible scenarios. I think there is at least one way that we could "escape" almost simulation, which is described early in the paper, with the mobile fish tank example -- if we managed to communicate with the outside world, and either solicit cooperation from the beings running the simulation, or managed to influence things in that world to be able to build our own mobile fish tank, we could in theory feed data from sensors in the outside world into the simulation, and data from the simulation back out to actuators in the real world. In this type of scenario, in some sense of course we would never truly "escape" from the simulation, as we would remain running on the same substrate, but we could experience and influence the outside world to a similar degree to any natural inhabitant of the outside world. The second type of scenario would be an information based transfer. Today we can train a neural network in a simulated world and then install that neural network in a physical robot in the real world. I'm not suggesting that today's neural network's are close to sufficient to capture the entirely of a human's consciousness, but assuming that technology keeps advancing at the pace it is, it's not hard to imagine that we will eventually have something that is, whether that's in 10 years, 100 years, 1m years, etc. At that point there's no reason that what we use to capture consciousness in our world could not be reproduced in some other medium in the outside world.