Hacker News new | comments | show | ask | jobs | submit login
A Type of Simulation Which Some Experimental Evidence Suggests We Don't Live In [pdf] (philpapers.org)
35 points by bshanks 7 days ago | hide | past | web | favorite | 54 comments





> Suppose we live in a computer simulation with the following x-xˆ property: for each memory-bit x in any computer in our world, there is a memory-bit ˆx in the simulating computer such that ˆx is used to store which value is stored in x. Then any such x is subject to two different types of soft errors:

One thing I've never understood about the simulation theory is that if we live in a simulation, we have almost know basis for making any assumptions at all about the nature of the reality in which we're being simulated. If you assume it's like ours in every way, then it's also equally likely to be a simulation, and on to infinity.


The suggestion that we live in a simulation is really not a "theory" in any meaningful way at this time. It doesn't explain anything that we can't explain otherwise and it doesn't predict any behavior we haven't yet observed.

It is, at present, effectively unfalsifiable and not much different from a claim that "God made all of this." It's just a different notion of God.

This is not an argument that we shouldn't pursue this line of thinking, but we shouldn't call it a theory until it makes new predictions that we can confirm or disprove.


It's really an umbrella term for a class of theories, some of which do make testable predictions (like anisotropy of space or, in this case, a minimum bound on memory error rate with zero radiation.) It's like the term "string theory".

One thing that all simulation theories predict is that our universe is interesting. It's an alternative to anthropic theories to explain why our universe is one with life and not one that re-collapsed to a singularity or expanded out into a thin hydrogen plasma. See https://en.wikipedia.org/wiki/Fine-structure_constant#Anthro... for one reason why it's only a small range of possible universes that are interesting.


I wrote a paper arguing that 'ancestor' simulations are effectively impossible. Part of that argument was that, in order for the simulations to be meaningful, the world had to work largely the same way as our own. (Otherwise, it's not an 'ancestor simulation'. Other types of simulations could work.) You could, for instance, just project a 'video' of the sky every night to save some energy costs. Until the simulated people gain the ability to, say, practice photometry. Then your have to spend a bit more energy simulating the actual physics. So, as time progresses in your simulation, energy costs rise.

Eventually, your simulated beings develop the ability to compute. There are some things you'll be better at optimizing, but over time the simulated beings will get closer and closer to the optimal way of doing an operation; at which point any operation they do is an operation you are doing.

Until they generate a simulation of their own. That's when it all falls apart. At the point they make their own simulations, now you're paying in energy for your simulations, plus all their simulations and their simulations' simulations in an infinite chain. Voila. Infinite energy requirement. (There's some caveats here... some simulations may result in a dead world, but presumably every successful simulation would run more than one successful simulation of its own.)

So, your options are:

1. End the simulation when the simulation becomes capable of running its own simulation.

Useful, perhaps, but not very for a very old civilization. It's not clear that humanity 1 million years will be that interested in simulating the first X00,000 years of humanity.

2. Dumb down the simulation.

However, this would have to be done "just so" so that the energy use limit converges as the depth of simulations approaches infinity. This also provides a mechanism for a simulation to detect that it is a simulation (I forget why, it's been awhile), making the entire exercise not useful as an ancestor simulation.

3. Prevent simulations from running their own simulations.

Like (2), this potentially allows the simulation to discover that it is a simulation. It also may make the results of that simulation less useful. If you, for instance, wanted to know what the world might be like if event X happened differently, you could change that event and let the simulation run, but as soon as you got to the point where they would have run their own simulations, you can't be sure of your results.


You don't add any energy from a simulation that is running a simulation. The higher-level simulation has to devote some portion of its own resources to running the lower-level simulation, and that cannot grow larger than all the resources available in the higher-sim universe. If the lower-level sim has more simulated resources, that just means a resource in the higher-level sim has to be multiplexed somehow.

In terms of simulations like our own universe, if we simulate a universe twice as large as our own, using half the resources in our universe for the sim, and reserving the other half for us to continue to live in, then barring any shortcuts or optimizations and assuming multiplexing by time, the sim will advance one simulated second for every four real seconds. All that simulated energy is already tied up in the simulation hardware we built in our universe.

If we are living in a simulation, then the parent universe won't ever have to spend more on our simulation than the maximums they set when starting it. Unless we break out of the simulation and hack our own simulator to allocate more resources, of course. We can't up the requirements from within, because the parent could just start multiplexing some resource without any simulated person ever noticing.

Also we have no way to know whether the resources of a simulator-parent are infinite or not, or the aleph numbers of any infinite resources they may have.

We think we might have an infinite universe, but the portion of it we can observe is finite (but very large, encompassing many galaxies), bounded by our light cone and the expansion of spacetime. If quantum effects are really just lazy evaluation in the simulator, the simulator should be able to prune everything we cannot observe.


That’s really no different than (2). I’m specifically talking about ancestor simulations in which the world of the sim is largely identical to the real world. Changing from infinite resources to finite is a big change. And having the simulation run more slowly than real time probably defeats the point. Imagine waiting for 600+ years for the simulation to complete to find out what would have happened if Abraham Lincoln hadn’t been assasssinated.

I argue that the distinction is effectively pointless, because human imagination can already simulate our reality. Have you ever had a particularly vivid dream? Ever daydreamed about punching that smug customer in the face, or launching a rocket at that ass in front of you in traffic?

What would happen if you had a dream about having a particularly vivid dream? If you dream in that dream that you're having another dream, does your real brain cook itself from using too much processing power? If you dream that someone else is having a dream, does that matter at all to your simulation? It's entirely possible that you never have to simulate that at all, because the dreamer forgets all about their dream when they wake up, and that branch just gets pruned.

By that standard, the (1) simulation would halt the instant some random animal ancestor fell asleep. The (3) simulation could not model anyone that made a decision based on a portentious dream--such as the dream about seven fat kine and seven lean kine found in a millennia-old book with some cultural significance. Only a (2) simulation has any chance of producing meaningful results.

And given the predictability of events, you would need to run that Lincoln simulation hundreds of times to learn anything from it, and many times more if you want any reasonable degree of precision. Nobody wants to wait 1000 years just to find out with 95% confidence what Lincoln might have eaten for breakfast a month after not being assassinated. You could prune more extensively if you had a specific question you wanted answered. Even if you just want to wander around in a simulated Lincoln-unassassinated Earth, the sim could still prune based on what you decide to look at. The sim would probably have to simulate Lincoln's impact on journalism regardless of whether you read a newspaper or not, but if you never stop to smell the flowers, it could probably prune Lincoln's impact on gardening.

An ancestor simulation has a great advantage over an alien simulation, because it can play back recordings of real observations rather than simulating or randomizing results from scratch.


I think we are talking about apples and oranges. The dream metaphor is inapt, because your brain is consciously experiencing only 'level' of that dream. There is only one "you". In the case of the simulations, each 'level' is independently conscious.

The idea behind ancestor simulations is that the simulated people are conscious. If we aren't talking about that, then of course there is no chance we are in such a simulation.

So, you can't really skip what Lincoln had for breakfast, because Lincoln has to experience it and because it can affect him. Maybe the chef poisons him. Maybe it gives him the runs and he cuts a speech short. Maybe he has an allergic reaction and misses an important meeting.

In any case, if you could prune it and we are in a simulation, then none of us would ever have the experience of having breakfast.


The solipsistic argument says that I can't ever really tell if you are really a conscious being or just a simulation of a conscious being. I can't tell if the real world actually exists, or if I am a brain in a jar being fed simulated sensory data.

And we don't need the experience of having breakfast if we have the memory of having breakfast. If you copied one memory of me having breakfast, randomized a few variables, and replayed it to me later, it would be unlikely to impact any of the impactful decisions I make later.

Since you're running a simulation, you can choose whether or not my breakfast needs to be simulated in detail or elided over after identifying a decision node that might be significantly impacted by my breakfast. I honestly probably wouldn't do anything differently if you swapped gravlax and bagels for eggs and toast. So you can usually replay a partially memoized breakfast sequence as it happens, and revise the simulation, overwriting my memory and that of all witnesses, if it turns out later that my breakfast might have an impact on something else.

So every day, I have memoized breakfast routine. At 3 PM one day, there is a potential for a hangry conversation at work that might get me fired. Now you can backtrack, copy all my variables from this morning into the breakfast simulator engine, and this time everything is simulated in detail. Then you patch the detail results over the rough results, and I can't tell the difference, other than the fact that I can now remember significant details of my breakfast, rather than a blur of things I usually eat in the morning.

You might not be able to skip Lincoln's breakfast, but you could skip a few million other breakfasts.

So that leads to the existential horror. Do I actually still exist if I'm not doing something meaningful to an external observer right now? Maybe most of me is just a slipshod model of a generic person, that exists solely to produce some initialization variables for a detailed simulation of some-guy-on-the-Internet replying to you, and then after I'm done, I'll just go back to being an NPC with no personality of my own.


Any shortcut you take you are no longer running an ancestor simulation. The reason you don't memoize breakfast is because the entire point of the simulation is the simulation of the conscious beings you are simulating. If you memoize away their decisions and actions, you take that away. And I strongly suspect retrofitting reality to match countless decisions is going to be more expensive than simply running it in the first place.

> I honestly probably wouldn't do anything differently if you swapped gravlax and bagels for eggs and toast.

That is unknowable. Maybe in making the eggs and toast, you leave the house a second later, and a driver that ran a stop sign that would have hit you passes in front of you instead.

You're not wrong that you can skip breakfast and take all these short cuts. You can do whatever you want in the simulation. But if you do, you're not running the type of simulation that is called an 'ancestor simulation'.


> Maybe in making the eggs and toast, you leave the house a second later, and a driver that ran a stop sign that would have hit you passes in front of you instead.

This is likely unimportant to answer the questions that the simulator is supposed to answer, which means that no one would ever bother running one ancestor simulation--as you define it--when they could instead run multiple trials of a different kind of simulation for the express purpose of eliminating such confounding "want of a nail" moments. "We ran this ancestor simulation for 10 million years to find out what logfromblammo would eat for lunch on 13 August, 2018, and the results were ruined when the simulated logfromblammo was killed in traffic that very morning, just before the entire simulator planet was demolished to make way for a hyperspace bypass. We'll check the budget to see if we can reset for another run."

Maybe the other driver also had a breakfast that delayed their departure by one second, and now they hit me again. In one ancestor simulation, you won't know all the various circumstances that contribute to the causation of an event. If you do cut corners, you can say that there is an X% chance on any given day that I get involved in a traffic collision.

So an ancestor simulation--as you define it--is functionally identical to walling off a piece of your own universe, rearranging matter inside it to conform to initial conditions, and then watching it to see what happens.

If your universe is itself a simulation, how does the parent universe detect the difference between someone running a simulation and Magrathean-scale engineering projects? If I commission a planet that is a perfect copy of Earth on the day before Lincoln was assassinated, and black-bag Booth an hour before he shows up at the theater, is that an ancestor simulation? Do I need to wrap it in a perfectly black shell and project sunlight and starlight through it from time-appropriate locations? That only requires that the star you use for energy is a bit more energetic than the sun. If you simulate a planet that predates astronomy, perhaps to see what would happen if dinosaur-like creatures hadn't been wiped out, you wouldn't even need to do that. Your total energy requirements for the lifetime of the simulation are the delta-v required to build an Earth-like planet with a Moon-like moon and move it into an Earth-like orbit around a Sol-like star.

We wouldn't be able to detect if our planet is a simulation by dinosaurs, to see what would have happened if their planet had suffered an extinction-level event before they launched themselves into space. How do you tell the difference between a naturally-formed and artificially-constructed planet? Likewise, how do you distinguish between someone trying to compute an ancestor simulation and someone just running an immersive MMORPG? If you did create a single ancestor simulation larger than a few interacting atoms, it rapidly loses its scientific power, due to all the variables you intentionally did not constrain. How is that not just an amusement park?


Worm Ouroboros! Tail-recursive simulation FTW.

Another thing no one considers. If the simulator or maintainers of the simulation identify that the individuals being simulated are aware of or close to determining they are in a simulation, they may alter the observations in the minds of the individuals being simulated to indicate there is no simulation.

That’s possible. If we are simulated, it’s also possible that we emerged in this simulated universe and the maintainers of the simulation don’t understand the human mind well enough, similar to how we don’t. It’s also possible “they” don’t know we’re there in some random corner of their simulation.

On the other hand it’s possible that the entity who would run this is aware and can manipulate our minds, or rewind time.

The problem is that if we just assert that we live in a simulation, anything imaginable or, if you allow for the outside universe to have completely different laws of nature, anything unimaginable as well.

I’m not a proponent of any simulation theories at all (they’re fun to think about, though), but I’d imagine that just like in “ordinary” science, people are interested in only testing falsifiable theories about the kind of simulation that we could understand and retain knowledge about.

Assuming for example that time can be restarted to prevent an insightful experiment to happen won’t ever give any meaningful result to begin with. As a corollary it’s never possible to disprove “simulation”, as a fuzzy term, in general. Scientists can however assume a much more constrained simulation that is maybe closer to how we would do it, and construct experiments that at least allow us to refute whether the alleged simulation seems to have those simpler properties or not.

In this case, it seems their simple model does not hold, and the simulation might be different. There can still also be no simulation, but we can never be certain.


I've had a theory for a while that we might be in a simulation run for historical purposes. That in a distant future, we or other civilizations evolved to a point where we were capable of creating a simulation of our own universe for the purpose of discovering more about the history of our universe and ourselves as a scientific endeavor. That or this is a virtual prison complex in which we are reevaluated for release back into society in the external reality. Just a couple of interesting thought experiments I've had.

You might enjoy this presentation on the Simulation Hypothesis as it touches on your idea of historical simulations. https://www.youtube.com/watch?v=nXIpR_agyl4

Perhaps the maintenance has been outsourced to h̷̞́̓̚o̴̲͉̳̝̖͆͛̈́̈́̔̉̕͘͠r̷̡̘̞͇͖̔̆͆̎̎͂̈͒͐̌ͅz̷̹̺͎̼̗̬̥̜͌̏ḧ̶̯͈́͌͋̑͛̈́̀̄̎̈́̆͋͊̚͝l̵̗̜͎͈̄̂ and the business impact of the simulated becoming aware isn't too bad to mandate the extra Y̸̞̰͆̌͜ͅK̷̨̛̲̣͕̙̪̮͋̃͆̊͂̾̏̇̏^̷̛͖̻͔̞̠̰̩̳̳̀͊͐͝ͅͅ}̴͈͓͚̲̳͚̳͔̮́̄̄̈́̏͂̓͌g̵͚͓̭͓͎͊͑̄͑͐͛̄̇͠ḥ̵̹͎̜́̐̊̇*̴̙̥͚̠̔̆̂͒́̅̂̐͂͒̽͜4̵̧̭̦̣̱͚̲͖̻̳̏̃v̷̪̄͛͠ţ̷̛̛͍̣̦̤̠̜̤̲̹̰̘̔͒̅͋͆̔̈͋̏̕͜͝ͅ>̴̛̛̻͖͔̘̱̠̙͍͇̻̣̪̼̍̌̍͗̒͌̊͠ͅF̶̛̰̣̝̞͔̠̦̘̜̺̎̏̅͜/̴̖̜͇̋̓̄͆̍̊̃̾̓̚͠?̵̢̢̡͈̦͚̞̝̘̫̺͕̺̰̇̑̃̿͆̃̋̍̍̇̌̕͝R̶̥͓̓̌ required to pay for the alterations.

Could you please not do that here? HN is a text-only site.

Your zalgo convinced my chrome that half the sentences on this page needed translated from English to Vietnamese. Impressive.

These sort of cheap jokes really do harm the conversation. They're net negative.

this seems eerily similar to how comic book writer Grant Morrison described the universe of super hero comics.

I want to dig for a better video but Wisecrack has a good upload on the topic: https://www.youtube.com/watch?v=EFtbwP4UuFM

of course Morrison takes us a much more...theological outcome, but yeah.


I have a theory for you - that the maintainers monitor for these events from a teapot somewhere between Earth and Mars...

This paper is assuming the simulating universe has things like "particles" and "cosmic rays". The properties of the simulating universe, including foundational things like rules of logic and mathematics, could be literally unimaginable (by design or accident) within the simulation.

As I wrote in another comment, you can still come up with theories for certain types of simulation, and test whether their properties hold or not. I suspect no one in the “simulation community” ever expects to disprove a simulation completely, but as the authors did you can try to come up with a likely simulation model (maybe close to how we would do it) and try to refute that.

> The properties of the simulating universe ... could be literally unimaginable ... within the simulation.

Totally. This is an area that religion claims to address—believers in a religion [1] could frame God as The Simulator and a divine revelation as a message from the simulator, perhaps regarding the purpose of the simulation—that otherwise there is no way to know.

The author of the paper describes a limitation of this line of inquiry:

> The type of simulation ruled out is very specific. Perhaps that is the price one must pay to make any kind of Popperian progress.

It reminds me of this:

> For were I to praise [God] ... I would find that my praise of Thee can befit only such as are like unto me, who are themselves Thy creatures, and who have been generated through the power of Thy decree ...

It's an area that's inherently challenging to explore from inside the simulation!

[1] Disclaimer: Baha'i here, watch out

[2] Baha'u'llah — http://www.bahai.org/library/authoritative-texts/bahaullah/p...


The beauty of logic and mathematics is precisely independence of the universe in which they've been discovered. I don't see how logic could be any different in any simulating universe.

From the abstract: > The type of simulation ruled out is very specific.

Really, it seems to apply to external soft errors with any cause, even unimaginable ones. But yes, it is still pretty specific.


>The properties of the simulating universe, including foundational things like rules of logic and mathematics, could be literally unimaginable (by design or accident) within the simulation.

This could be the quadrillionth iteration by which point they'd be pretty confident though.


(a) Maybe the simulator uses ECC RAM (b) Maybe the simulator's world is as similar to our world as the Game of Life is to Earth.

Author here. The type of simulation ruled out is "external-soft-error-prone simulation with (two other properties)". If the simulator uses ECC RAM, then the simulation is not external-soft-error-prone.

You might say: "But surely the simulator does use ECC RAM!" Yes, but that's speculation. It's nice (and surprising) that a concrete experiment can empirically support it.


How exactly would the 1:1 bit correspondence work in practice? What is the binding between simulated and external bit? Wouldn't we need to operate at the edge of Landauer's principle to really rule out this possibility?

Imagine if you were to program a video game in which players could interact with in-game computers, which were implemented as genuine full-feature emulations of working computers. It seems plausible that a memory-bit in one of the in-game computers would have its value stored in a memory-bit in the real computer on which the game was running.

What do you precisely mean by genuine full-feature emulations of working computers? If the in-game computer is represented as some "physical model using fundamental particles" than each bit of simulated memory is backed by, I don't know, gigabytes of data about properties of those particles? Isn't what constitutes a bit strictly dependent on the process making the actual reading? How would the simulator know what is considered a bit of information within the simulated system, and intentionally map it 1:1 onto a real memory bank?

If we discard these problems, wouldn't that imply that if such mapping were possible, that the simulation would also have to be strictly Anthropic i.e. the only running "software" are minds being fed data as "sensory" input?


Hi, not sure if you're still reading this, but I thought some more about your question.

It makes me realize, computers-in-the-simulation aren't really important in my argument. What really matters is that the people in the simulation discover some way of making observations about microscale events. It could just as well happen that, say, they make many observations of sterile petri dishes and eventually notice "soft errors" in the petri dishes (e.g. spontaneous infection occurring due to real cosmic ray hitting simulating computer). So the inhabitants move their petri dishes into a vault, etc., and check whether that was enough to protect the petri dishes.

I think computers are probably the best way to quickly get the idea across, but it could definitely generalize.


Things that the players think are computers. Sorry for such a vague answer, but here's an example.

Suppose some people live in a Minecraft server, unaware there's anything else. Within the game, they build some computers (search for "working computer in Minecraft", it's been done many times).

They run their computers a long time and observe soft errors. The soft errors are caused by cosmic rays hitting the Minecraft server in the real world, but our heros don't know that. They attempt to protect their computers, moving them underground and setting guards to make sure no creepers get in, etc. But nothing they do stops the soft errors. Of course! What could they possibly do to stop soft errors caused by real cosmic rays hitting the Minecraft server?


It might be interesting to do exactly that - rerun the OP example in a Minecraft server, attempting to prove that it is a simulation.

Proving the simulation hypothesis true in Minecraft could give ideas on how to prove it in our universe.


It's equally plausible that our single bit is far too high-level of an object for them. Why not just simulate everything that makes up our bits? (germanium atoms, etc)

True. The paper doesn't rule out such a simulation. (Unless maybe the experiment could be run with memory-bits so hyper-sensitive that a single germanium atom could cause a macroscale bit-flip? Just kidding)

You joke, but aren't we're running into those limits? How many silicon atoms in a nanometer? Also... I misspoke. I was thinking gallium.

I don't really follow the argument. If the number of errors in the vault were nonzero, it would most likely indicate that the vault did not screen some ("real", "physical") effect that could cause errors. For instance, you could have errors due to radioisotope decay, low-probability interactions with neutrinos, etc.

They acknowledge that:

> Without additional technology, we are unable to tell which soft errors were external and which were internal.

...

> Perhaps this paper’s most interesting conclusion is just that a non-contrived simulation hypothesis is falsifiable in a concrete way.


I think compute power is what matters, not amount of simulations. Compute power available in simulated realities vs base reality will always be significantly in favor or root reality.

And parsimonious simulation won't work, or would be even easier to detect than what this paper was talking about.


> A paper by O’Gorman et al [1] describes (p. 46) the following experiment and its results. A total of 864 modules were first run on the second floor of a two-story building for 4,671 hours, during which time, 24 soft errors were detected. Then, the same 864 modules were run for 5,863 hours in a nearby vault shielded by about 20m of rock, during which time, zero soft errors were detected.

Interesting, but there's no guarantee that time moves at the same rate both inside the simulation and outside of it -- indeed there's a good chance that the simulation runs much faster than the external environment. A 25% increase in observation time seems way too low to draw any kind of conclusion.


There are also no guarantees that the environment outside of the simulation has anything in common with the environment inside of it. Perhaps that universe doesn't have particles at all. Maybe they are simulating this universe because they are curious about what a universe like ours would be like.

I doubt you can map the time at all. At best, you can probably make the case that the internal time of the simulation is bookended by the external start and end times of the simulation.

Assuming external-soft-error-proneness and x-x and uni-directionality, we can think of a "space" of possible simulations we might live in, thought of as quadrant 1 of an x-y plane where x represents "how many seconds pass in our simulation for each second in the real universe" and y represents "frequency with which soft errors occur in the simulating computer". Different experiments would rule out different parts of that plane with different probability.

The computer could have enough error correction where it doesn't matter.

I love this title and abstract; we need more of this kind of plain speaking and rigor when dealing with mysticism-adjacent areas.

I think we are not in a simulation. Though our world is pretty close in principle to a simulation -- there are certain (small number of) building blocks which contruct this complex world. That should be it.

The simulation conjecture is a funny way to get around to Intelligent Design.

I don't see how that follows.

If we're in a simulation, it must be running on a simulator. That simulator is presumably not just a natural feature of the universe - that's just "laws of nature deeper than we have discovered so far", which I take as something different from "simulator". In fact, the simulator (by my definitions) has to be an artificial construction - that is, created by an intelligence for a purpose.

That doesn't necessarily get you intelligent design of life, but it does get you intelligent design of the universe.

I, too, find it amusing that we've decided that God doesn't exist, that the natural universe is all that exists. And then some people decide that we're in a simulation, that the universe is an artificial construction created by someone outside of it (but who totally is not God). I'd find it even more amusing if I thought that the simulator people had a reasonable position. (To me, "you can't prove me wrong" does not imply "therefore I'm probably right".)


Curiously, parallel to this thread, there's another one on HN, which discusses (the lack of) exceptions in golang. And the reasoning starts out with "we don't need exceptions", so we throw them out the door, but then, after a couple of intermediate steps, they come back through the window. Same door-window scenario is playing out here :)

God wants you to set a breakpoint, read out some locals, and put a watch on one of them.



Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact

Search: