carado.moe

tammy's blog about
AI alignment, utopia, anthropics, and more;

also on twitter, lesswrong, rss, and at

finding earth in the universal program

this post expands on step one of the Peerless: creating virtual people. brain scans-and-simulations are apparently still quite far off, so i'll be focusing on the second approach: resimulating earth and plucking out persons.

(one great side-advantage of this method is that, if we can relocate earth to pluck out persons for the simulation of alignment researchers, then we can later also relocate earth in order to restore it once we've solved alignment. so resimulating and locating earth, regardless of having early enough mind-plucking-out tech, is something we might need to do anyways.)

if compute is infinite and we don't mind being inefficient, then we can use exponential or even infinite compute to locate earth. one approach is the following: create a big informational beacon — perhaps a copy of a huge portion of the internet, along with MRI scans of as many people as we can afford. then, we use some type of (non-intelligent) deterministic error-bound statistical location procedure to locate patterns that look like that beacon inside the universal program. we can afford the statistical detection to be imperfect — if it misses on one encoding of earth, there will be different ones in the universal program.

because of the time penalty of the universal program, however, we may find just compressed copies of the beacon (instead of a full simulation of earth leading to the time at which we build that beacon), and because of the deterministic bound, we want need to stop on the first match; if this first match is just the beacon, without earth, then we fail; perhaps superintelligence can notice that it's not finding any nearby minds to pluck out, or perhaps it plucks out garbage. so we can start the universal program with not one step per program, but rather a very large number of steps — i hear stephen wolfram has estimates on the number of computation steps it takes to get to the current state of the universe. this will favor programs that takes every long to lead to the beacon, but are themselves shorter program.

(what if the first program to contain earth is itself a universal program without that huge constant, such that it finds the beacon before it finds earth? i am not sure how to address this. perhaps we can explore programs in an order that favors worlds that look like our physics instead of looking like discrete iterations of all computations?)

there's also the concern that the universal program, just like the universal distribution, is malign. i'd think plain top-level earth, maybe especially as detectable by a simple enough beacon locator, would tend to occur before malign aliens emitting our beacon to trick us; but that's a risk to keep in mind.

if we do care about computational efficiency, then there are two main factors we need to account for:

• can our universe can be ran in polynomial time on whatever computers the superintelligence can build? for example, can it be ran in polynomial time on quantum computers, and can quantum computers be built? note that if this is the case we might need to step through quantum steps of quantum programs to run the search in the expected time. this doesn't need we need to build quantum computers outselves, mind you — superintelligence can just notice that a quantum computer would run the computations we describe efficiently, and build and use those.
• is the "seed" program for the universe small? intuitively i believe it is, and i find wolfram's efforts to reproduce the behavior of particles from the standard model using simple graph rewriting, to be evidence in that direction. that said, if it is large, then finding that program is an exponential search again — and so, again, we might need to build a search that "favors" our physics to save on exponential search time.

finally, we might want to put a hard bound on the number of tries the superintelligence will run to locate earth. the reason for that is that, if for some reason we messed up something in the beacon locator and it never, ever finds earth, then it will instantiate all computations, which appears to me to be a potential S-risk. in fact, even if we do find earth, it may not be worth it if we have to simulate exponentially much potential suffering before running our utopia — what if, after solving alignment, we have a great time, but then decide to eventually fade away after only polynomial time? then we will might have created exponentially much suffering in total.

intermediary simulation

in case isolating minds from this simulation is hard, we could build an intermediary step between the location of earth in simulation-space, and booting the peerless simulation proper — superintelligence could, once it has located our beacon, get in touch with our organization inside the simulation of earth, and give it extraordinary computational (and maybe physical?) ability within the simulation to either take over everything, or figure out brain plucking-out and then let us press a big "ok, start now" button.

note, however, that we might not want to remain in this intermediary simulation for too long — it is still vulnerable to inner unaligned superintelligences, just like our top level reality is. we want to get to a safe, sandboxed, computationally weak environment as early as possible.

this is also a great argument for readying ourselves to build the beacon and utilize this contact-from-superintelligence as early as we can; indeed, to make that the first step of implementing the peerless plan. the reason for that is that the earlier we are able to take advantage of it, the earlier the time step of the simulation superintelligence can help us start bootstrapping towards the proper simulation of the peerless, and the less likely we are to be doomed by other superintelligences, if we need some intermediary "pre-peerless" simulation time.