Based on Nick Bostrom's simulation argument, I think it's fairly likely that we're in a simulation. But what kind is it? Is it a full-scale replication of the history of the basement universe? Or is it something more specific and unrealistic? In the past, I have, by habit, tended to imagine the former, but now I realize that it's most likely the latter, unless there's something I'm missing.

In particular, it would be really expensive to simulate physics. If the simulators wanted to make the world accurate for the observations that people make, they'd have to go down to the quantum level, because people do make quantum-level observations. But this would be prohibitively costly. Think of how much computing power we'd need to simulate a single quantum particle accurately. Simulating a quantity X of physics probably requires many times X amount of computer hardware. So unless the simulation is running in a basement universe that's much bigger or allows for much easier computation, the simulation is unlikely to have quantum-level fidelity.

But then how do we explain the consistency of our observations of quantum physics? Here's Bostrom's suggestion from his original paper:

Simulating the entire universe down to the quantum level is obviously infeasible, unless radically new physics is discovered. But in order to get a realistic simulation of human experience, much less is needed – only whatever is required to ensure that the simulated humans, interacting in normal human ways with their simulated environment, don’t notice any irregularities. The microscopic structure of the inside of the Earth can be safely omitted. Distant astronomical objects can have highly compressed representations: verisimilitude need extend to the narrow band of properties that we can observe from our planet or solar system spacecraft. On the surface of Earth, macroscopic objects in inhabited areas may need to be continuously simulated, but microscopic phenomena could likely be filled in ad hoc. [...]

Moreover, a posthuman simulator would have enough computing power to keep track of the detailed belief-states in all human brains at all times. Therefore, when it saw that a human was about to make an observation of the microscopic world, it could fill in sufficient detail in the simulation in the appropriate domain on an as-needed basis. Should any error occur, the director could easily edit the states of any brains that have become aware of an anomaly before it spoils the simulation. Alternatively, the director could skip back a few seconds and rerun the simulation in a way that avoids the problem.

The skipping back a few seconds would only work before people became aware of the contradiction, because once a simulated experience is run, it "happened," even if you undo the memories later. Some fraction of observer-moments would still consist of observing the wrong thing.

Keeping track of the belief states of all minds seems like a lot of work. Maybe it wouldn't require vast computing power, but it would be a challenging software problem, because you'd need a classifier to map from brain-state data to a semantic label that "Harry believes X," which is a different level of abstraction.

In any event, the implication of all this is that if we're in a simulation, the physics that we observe is just made up. Maybe it's approximately correct to a degree that's cheap to compute, but it's probably not correct down to every quantum particle, or even to every cell. The cheapest experiences to create would be just brains with very shady input from the environment.

And, by consequence, it seems unlikely that our simulation is a near-exact replica of the way things were in the real basement universe, because it's too expensive to compute how things really were. Of course, people kept history books and took video recordings, and those could help stitch together a rough picture. But our simulated experiences would then be akin to a movie reenactment of the fall of Julius Caesar: The costumes and events and environment might be about right, but the simulation would be making up the gaps that were lost to history.

Timeless decision theory suggests that even if we are in a simulation, our choices can still matter a lot if they correspond to choices that our same brains made in the basement universe. But if we're not an exact simulation of the basement universe, then it seems our ability to influence the basement is less than we might have thought. Still, if there are lots of simulations of minds close to what was in the basement (e.g., people narcissistically creating vast copies of their former selves with approximately realistic environments), the correspondence may still hold enough to matter, but I don't know how much it's weakened. Perhaps the simulation is only being held on the right track to correspond with historical reality through artificial revisions, because the environment and choices of the simulated mind really aren't close enough to what happened in the basement due to butterfly effects resulting from the insufficient level of detail?

In view of this, my estimation of the power of our minds if we're in a sim is slightly lower than before, which means that it's slightly better than I previously thought to act as though we're not in a sim, because if we're not, then we potentially have much greater power. That said, the mind experiences generated in sims may be replicated astronomically many times, and to the extent that actions still have physically reasonable consequences in the sims, then our choices still do still matter both to ourselves and to our plentiful copies -- even if we don't control anyone in the basement.

Solipsism?

If I am in a sim where the external world is only adumbrated rather than thoroughly computed, does this imply solipsism? Maybe it increases the chance somewhat. For example, maybe my future self became wealthy and powerful and decided to simulate gazillions of copies of himself, without simulating other people and animals along for the ride. That said, I hope this isn't the case, because I would prefer my future self to use his resources for reducing suffering rather than reproducing his past life over and over. My life is quite good, but there are much better lives that could be simulated, and in any event, I think it's more urgent to reduce suffering first. Of course, solipsistic simulations of me could also be run by other agents besides my future self, but in this case, it's less clear why they're singling out just one mind to focus on.

Also, the agents that I interact with (people and animals) seem pretty realistic, and in order to compute some of their reactions to what I do, you'd probably need a good amount of their minds being run as well. In general, it seems easier to just run their minds too, and let the different minds interact. The sim would then be like a MMOG, and instead of wasting resources trying to fake the external minds like in the solipsistic scenario, you'd actually have the other minds for real. Presumably this would let the simulators run more minds per unit of computation, memory, storage, etc.

Altruism still matters

So, maybe a good way to imagine our sim is as a MMOG. In this context, our actions do affect the welfare of others, just like they do if we're not in a sim. The MMOG has its own laws of physics, and what we do physically causes injury or benefit to our companions. It doesn't matter that quantum-level phenomena aren't being simulated. Altruism makes just as much of a difference in a world where Newtonian physics engines are used as in a world where the universe is actually evolving according to the Schrodinger equation.

Our observations of the reliability of cause and effect would still be valid in a sim, and evidence-based assessments of which types of interventions will do the most good would still be prudent. Studies on animal sentience would still make sense, because if we see animals acting in complex ways, it could become more parsimonious to suppose that their minds are being simulated along with ours rather than that they're an elaborate, stylized presentation to our senses.

One difference of life in a sim from a non-simulated world is that apparent violations of physics, causality, etc. wouldn't be as heavily penalized by Occam's razor, because these things wouldn't actually violate physical law at all. But such violations still require special logic in the software, giving it a small Occam penalty, and in any case, empirically, we don't see many such violations. But the probability of paranormal explanations for weird things does increase a little.

If we are in a sim, then maybe we'd expect our lives could be cut short, if the simulators decide to end things early. Like Last Thursdayism suggests, we might also have been created just recently, rather than having gone through our whole lives to this point. However, the set-up of our world is pretty elaborate -- especially our memories -- and it's possible that the easiest way to get our memories and world-state to be the way they are is actually just to run the whole thing from scratch, rather than trying to set up the conditions correctly. Programmers will tell you that when you're trying to debug your code, it can be easiest just to rerun the program from the beginning and set a breakpoint rather than trying to change the state variables to mimic the point of the program that you're interested in. In theory, I suppose the simulators could have run the world once and stored data dumps at different points in history, so that they could then re-load the world at a desired point without recomputing the whole thing.

In any event, we have to remember that we may in fact not be in a sim. I don't know what probability to give, but say there's at least a 30% chance we're in the "real world." Contributing factors to this probability are the possibility that post-humans don't acquire lots of computing power, the possibility that they don't do these kinds of sims much, the possibility that our reasoning about the simulation argument is flawed, etc. If we're not in a sim, then business goes on as usual. Even if being in a sim somehow negated the importance of altruism -- and it probably doesn't -- we should still be altruistic because we may not be in a sim, and the probability that we aren't will never become vanishingly small.

Finally, remember timeless decision theory: Even if we were a solipsistic simulation, we're still instantiating cognitive algorithms that are common to other minds, including minds in the basement. If we exhibit altruism, this contributes -- to some degree -- to their exhibiting altruism as well.

Motivations for simulating?

Of all the things post-humans could simulate, why would they choose minds like ours? Why not something happier, or at least more interesting and complicated? It seems our simulators are not utilitarians....

If we look around today at why people build virtual worlds, the reason is usually for entertainment, rather than for industry or science. That said, in video games, the agents are not autonomous minds the way you and I (presumably) are. It's less fun to play a video game where you just watch self-directed creatures do stuff, unless the purpose is to be more like a movie than a video game. But if it's just like a movie, then why not record it once and replay the same thing to everyone? You don't need large numbers of copies of the sim for that.

Video games do have autonomous agents, but they're most often the villains -- the goombas and sand worms and Skulltulas and Bowsers. They might also be the fellow denizens of the world who aren't the protagonist. Assuming we're not enemies, maybe we're in this latter category: We're co-inhabitants of the virtual world who interact with the post-human game players.

As suggested earlier, it's also possible that we're the product of narcissistic minds who wanted to create lost of copies of their past lives.

Presumably we're not being run as a form of utilitronium -- the simulators could do a lot better than this -- but maybe we're being run by people who value "life as it used to be" or "historical preservation" or some other unpleasant thing.

Some virtual agents today are run for scientific purposes -- to refine AI models of organisms. I suspect these are dwarfed by game AIs, but maybe this won't always be the case, especially if the AIs that survive the best highly value science. But is there really that much scientific value in simulating me? Maybe for psychology experiments? Or economics, game theory, sociology, historical science (which could become more scientific with the use of sims), etc.

Are there industrial applications of virtual-world sims? I can't think of many offhand. It's not as though we're performing a useful data-processing task or solving a computational problem -- except just to discover the computational result of a world with our initial conditions. It seems like most computationally useful operations wouldn't need to be sentient or, if they were, they wouldn't need to be simulations of a real world; they could just be minds conscious about doing their data-processing jobs.

Brian Tomasik wrote:Video games do have autonomous agents, but they're most often the villains -- the goombas and sand worms and Skulltulas and Bowsers. They might also be the fellow denizens of the world who aren't the protagonist. Assuming we're not enemies, maybe we're in this latter category: We're co-inhabitants of the virtual world who interact with the post-human game players.

No way! We're the heroes!:D!

I just read "I don't know, Timmy, being God is a big responsibility", and it's pretty relevant.

Thanks, Peter! Another friend shared that essay with me back in ~2008. Fun stuff.

Among many puzzles of our potential simulation is why the simulators have made physics look so complicated. I'm guessing it's a lot simpler to apply Newtonian mechanics to macroscopic bodies than to have small particles exhibit quantum phenomena. Why have our simulators included quantum physics, complicated particle physics, and all sorts of tough questions that physicists are still puzzling over?

There are plenty of other hard problems about which one could ask the same questions -- in biology, economics, psychology, etc. But at least there, it's plausible that the complexity arises for real from the interactions of high-fidelity sims of minds. If there aren't high-fidelity sims of physics down to the quantum level, why pretend that it's so complicated when we examine it in the lab?

If we're a simulation, I suspect we aren't interestingly different from a basement universe in the sense serving some function for intentional simulators. Presumably, we could compare the probabilities of possible parent universes by multiplying each universe's prior probability of existence or measure by the number of our-universes we'd expect them to run. I don't know how to think meaningfully about the former (algorithmic complexity?) but universes with our own laws of physics are clearly limited in the number of universes they can simulate, even if universes with their own laws of physics are particularly interesting to them. Presumably most of the probability mass is taken up by universes where simulation is trivially easy, and in that case I don't see why they shouldn't just be spitting out universes with just about any laws of physics, one of which is us. Or to mirror Bostrom's argument as closely as possible, either:

a) there are no such easy-computation universes,
b) they don't want to simulate anything like us, or
c) one of them is our parent.

On the other hand, I don't have anything close to the intelligence to actually say why (c) is more likely other than the intuition that (a) and (b) are stronger claims while being just as weird.

On the other hand, the idea that the world is governed by deceptive angelic powers is just delightfully weird. In principle, we should be able to outsmart them, no?

Interesting ideas, Oligopsony!

Oligopsony wrote:If we're a simulation, I suspect we aren't interestingly different from a basement universe in the sense serving some function for intentional simulators.

Maybe, although in the basements where simulation is trivially easy, people may simulate willy nilly without much reason for doing so. After a while, they might run out of good reasons.

Oligopsony wrote:Presumably most of the probability mass is taken up by universes where simulation is trivially easy, and in that case I don't see why they shouldn't just be spitting out universes with just about any laws of physics, one of which is us.

Well, it does depend on the tradeoff between how easy simulation is vs. how likely our universe is. For example, say that when beings in our own universe simulate a universe, they choose a universe like ours with probability 1, but when beings in an easy-simulation universe simulate, they choose universes like ours with probability p, perhaps related to the complexity of our universe. The easy-simulation universes can run N times as many simulations as we can. Depending on whether Np is bigger or less than 1, there would be more or fewer of us in the easy-simulation universe. Then, as you pointed out, we also have to consider the intrinsic probability of easy- vs. hard-simulation universes.

Oligopsony wrote:In principle, we should be able to outsmart them, no?

Umm... I'm doubtful.

Does the simulation argument ignore quantum uncertainty? I agree with Bostrom that we do not have to simulate every atomic interaction for an ancestor simulation, because most atoms are uninteresting, like rock under the Earth's surface. However, if quantum uncertainty does affect consciousness, then you would very quickly get many different scenarios, so it seems like a long-term simulation would be untenable. See <http://holtz.org/Library/Philosophy/Scientific%20American%20Parallel%20Universes%20-%20Tegmark%202003.htm> where they postulate an alternate universe for every quantum choice.

Interesting, Recumbent. I think the easiest solution is just to not actually simulate quantum splitting and just pretend like you did if the need arises.

Why is it so important that the ancestor simulations have the same physics? They could do them so much easier with Newtonion physics. Even if they want to do the former, unless they do it to a comparable amount, we would expect simple physics. Instead we get something where it's nigh impossible to calculate the waveforms of electrons in a single atom if it has more than a couple of them. For those of you not versed in quantum physics, entangled particles act as a system with three times as many dimensions as there are particles. Thus, an atom with 10 electrons will have 30 dimensions, and if you tried to calculate it numerically with only a width of ten pixels for the atoms (which would be nowhere near enough), that would be 10^30 pixels.

DanielLC wrote:Instead we get something where it's nigh impossible to calculate the waveforms of electrons in a single atom if it has more than a couple of them.

Could the physics of the basement universe be such that these computations are actually simple? For example, maybe they can compute complicated things like this with a single CPU cycle.

You can simulate quantum physics okay with quantum computers. Even so, I don't think they'd bother unless they had to. If they were simulating a different universe, they'd stick with simpler classical computing. If they're simulating their own, then they're in a universe where quantum computers are extremely difficult to deal with.

Brian Tomasik wrote:Timeless decision theory suggests that even if we are in a simulation, our choices can still matter a lot if they correspond to choices that our same brains made in the basement universe.

Even in the scenario of a direct replica of your mind, there remains a question about how to partition credit. For example, say you're in one sim, and you have an exact copy in the exact same environment in another sim. You rescue a bug from drowning in a bucket of water in your world, and this logically implies that your copy has rescued the same bug in his world. By making the choice to save the bug, do you get credit for saving 2 bugs, or just 1? I don't know. A friend of mine has the intuition that you should divide the amount of good done by the number of copies, but this isn't obvious to me.

This could matter because it affects how to trade off various choices. For example, suppose your sim is going to last 10 minutes, but the sim that contains an identical copy and environment as yours will be shut off after 5 minutes. There are 2 bugs drowning in a bucket near you, but if and only if you don't waste time to save them, you can venture to another location and save 3 bugs in a bucket there. It will take 9.5 minutes to reach that other location. By that time, your copy's sim would have been shut off. If, when you save the 2 bugs in the bucket near you, you only get credit for saving 2, then it seems you should venture off to help the farther bugs, because you can save 3 of them. But when you choose not to help the 2, you imply(*) that your copy doesn't help them either, so only 3 bugs are saved instead of 4.

I added (*) because in fact, if your copy knows he'll be shut off after 5 minutes, he's not an identical copy after all. He can realize that you should go save the 3, but he should continue saving the 2 nearby. If you're uncertain whether you'll be shut off after 5 minutes or 10, then symmetry can be restored, but now the expected value of going after the 3 far away is only 3/2 < 2, so again, there's no problem: You both save the two nearby and help 4 bugs.

If it's true that credit should be divided by the number of copies, then it's no longer the case that influencing the basement through being an identical copy is vastly more important than being a different kind of sim. Suppose that in one scenario, there's a basement and 99 perfect copies of it. If you're a copy and do something good that affects lots more sims that aren't copies of you (say preventing 10^100 suffering insect sims run by the basement), then you did affect the basement, but your impact is divided by 100 copies of you, making the total benefit only 10^98. Similarly, suppose the basement didn't run any identical copies but instead ran 99 different kinds of ancestor-like sims. In this case, you don't have timeless influence over the basement. Still, you don't know if you're in the basement or in one of the 99 copies. There's a 1/100 chance you're in the basement, and if you are, you alone control the 10^100 suffering insects. Again, your expected influence is 10^98. Thus, if we divide credit by number of copies, it doesn't seem to matter whether you're an identical copy.