Hello again, compat...
The use of the human mind’s universal model as a starting point
The aspects over which real people vary is a function of real lived histories, not random numbers in a neural template that you can just randomize independently. In order to dredge up a uniformly sampled mind who really could have lived, so that total measure of humans who really lived would be preserved over conditions of multiversal cooperation-, you would need to simulate randomly chosen actual human histories. That would be expensive. Most lived histories were pretty miserable, too, so there's an additional eudaimonia cost being imposed there.
If your generator abandons the brittle logic of real histories, you're going to end up with a majority (there are many more ways to miss reality than hit it) of specimens who, on inspection, could not have lived. Who have inconsistent memories, personalities that don't accord with their experiences, skills they never practiced.
I suppose, if you could command, say, several billion stars for a billion years, then stop once we've satisfied our obligations of the acausal treaty, grant the resurrected visitors from other timelines citizenship and the associated immortality, then get on with living out our eschaton, I suppose humans' grief might really be so deep as to motivate a project like that.
To know that our lost ones are alive out there in some other timeline, with the same measure they left off with, maybe that's really worth the cost for some people.
Hopefully not too many people. Some of us would rather just let the dead lie and make new, better people instead for a fraction of the cost.
This is a reasonable objection which may need Patch 4 for the whole method in order to escape "billion stars for billion years" (which is still small cost for universe-wide superintelligent AI, which will control billions of billions stars for tens of billions of years).
Fortunately, the Patch 4 is simple: we model just one mind history, which complies known initial conditions, and use random variables for unknown initial historical facts. In that case we get correct distribution of random minds, but we spend computational resources to simulate just one person. Some additional patches may be needed to escape intensive sufferings inside the simulation, like the use of only one playing character and turning off its subjective experiences if the pain is above unbearable threshold.
To resurrect all the dead, we don't need to run many past histories, but we need just one simulation of all human past, in which case all characters will be "playing characters". Running one simulation may be computationally intensive, but not billion stars for billion years.
The next step of such game would be resurrect "all possible people", which again could be done for a small cost via multiverse-wide cooperation. In that case, creation of new people, resurrection of past people and giving live to all possible minds will be approximately the same action or running different simulation with different initial parameters.
Moreover, we may be morally obliged to resurrect all possible minds to save these minds from very improbable timelines, where evil AI creates s-risks. I will address this type of multiverse-wide cooperation in the next post.
Fortunately, the Patch 4 is simple: we model just one mind history, which complies known initial conditions, and use random variables for unknown initial historical facts.
Shameless plug: you may enjoy my short fiction piece on a similar idea.
Just regarding
which is still small cost for universe-wide superintelligent AI, which will control billions of billions stars for tens of billions of years
All of the stars will be dead in 100 trillion years (although it's likely a good org will aestivate and continue most of its activities beyond that, which supposedly will get them a much higher operating efficiency than anything that's imaginable now). There are only 50 Bn stars in the local cluster, and afaik it's not physically possible to spread beyond the local cluster. All that stuff's just a bunch of fading images that we'll never touch. (I tried to substantiate this and the only simple account I could find was a youtube video. Such is our internet https://www.youtube.com/watch?v=ZL4yYHdDSWs best I could do)
(And it doesn't seem sound, to me, to guess that we'll ever find a way around the laws of relativity just because we really want to.)
It still seems profoundly hard to tell how much of the distribution of a history generator is going to be fictional, and it wouldn't surprise me if the methods you have in mind generate mostly cosmically unlikely life-histories. You essentially have to get the measure of your results to match the measure of people who really lived and died. We have access to a huge measure multiplier, but it's finite, and the error rate might just as huge.
How many lives-worth of energy are you trading away for every resurrection?
Personally, I think that it would not be computationally intense for an AI capable to create past simulations (and also it will create them anyway for some instrumental reasons), so it will be more likely to be less than 1000 years and a small fraction of one star energy. It is based on some ideas about limits of computations and power of human brain, and I think Bostrom had calculations in hist article about simulations.
However, I think that we are morally obliged to resurrect all the dead, as most of the people of past dreamed about some form of life after death. They lived and died for us and for our capability to create advance technology. We will pay the price back.
If the required kind of multiverse exists, this leads to all kinds of contradictions.
For example, in some universes, Personal Identity X may have given consent to digital resurrection, while in others, the same identity may have explicitly forbidden it. In some universes, their relatives and relationships may have positive prefrences regarding X's resurrection, in others, they may have negative preferences.
Given your assumed model of personal identity and the multiverse, you will always find that shared identities have contradicting preferences. They may also have made contradicting decisions in their respecting pasts, which makes multiverse-spanning acausal reciprocity highly questionable. For every conceivable identity, there are instances that have made decisions in favor of your values, but also instances who did the exact opposite.
These problems go away if you define personal identity differently, e.g. by requiring biographical or causal continuity rather than just internal state identity. But then your approach no longer works.
I personally am not motivated to be created in other Everett branches, nor do I extend my reciprocity to acausal variants.
I think that most of your objections are addressed in the patch 2 in the post. As we use all biographical data about the person to create his model (before filling gaps with random noise) we will know if he wanted to be resurrected or not. Or we will not resurrect all those copies which do not want to be resurrected.
There are elements of biographical and causal continuity: We use all known biographical data to create the best possible model, and such information is received via causal lines from the original person, which creates some form of causal connection between original and resurrected copy.
TL;DR: Measure decline in random mind creation may be prevented if we take into account very large number of random mids created in other universes.
Summary: P.Almond suggested the idea of the resurrection of the dead via a quantum random generator which creates a random mind, but such an approach has several problems: non-human beings in our world, non-necessary suffering of non-perfect copies, and measure decline.
Here I suggest three patches, which prevent most of the undesired effects:
1. Human mind matrix to prevent pure random minds appearing.
2. Digital immortality data to create a person which satisfies all known external expectations, and the use of randomness only to fill unknown information.
3. Multiverse-wide cooperation for the “cross-resurrection” of the dead between multiple worlds via quantum random minds, so the total measure of all resurrected people will not decline.
1. Introduction
Almond in “Many-Worlds Assisted Mind Uploading: A Thought Experiment” suggested the following idea about the resurrection of the dead by the use of a quantum random generator, which would create a random mind within a computer (Almond, 2006):
[A technician who lost someone’s brain scan file] writes a computer program which takes input from a physical system. The physical system, known as a quantum event generator, generates "1"s and "0"s randomly as a result of quantum events. The program will use the physical system to tell it what sequence of "1"s and "0"s will be used to try to recreate the lost scan file. The program starts with an empty scan file which will be filled with "1"s and "0"s.
If the many-worlds interpretation of quantum mechanics is correct, all possible minds will appear in separate timelines starting from the moment of random mind creation, which would mean the resurrection of everyone from his own point of view. However, this approach will a) not help an outside observer, who wants to resurrect a relative, for instance, as the observer would see only a random mind, and b) the quantum "measure" of existence of each mind will be infinitely small.
2. Problems of Almond’s approach
To illustrate the problems with quantum mind uploading, I will explore a simplified thought experiment where only names will be restored using quantum mind uploading. First, here is what Almond suggested:
Thought experiment “Not-patched quantum mind uploading”:
Bob had a friend John Smith. John has died and Bob wants to resurrect him. Bob remembers only first letter of John’s name: S.
Bob and John are interested only in the uniqueness of name preservation, and no other identity considerations are important. Bob wants to observe his friend to be alive, and for his friend to be named “John S….” (I would call it immortality from the point of view of the external observer). John wants his own immortality, and will be satisfied only if “John Smith” is created.
Bob creates random quantum mind A using a quantum generator to choose each new letter in the names.
It turns out that A is “jYY2№@11”. Only less than 10-30 share of all such copies in the multiverse are named John Smith. Both Bob and John are unhappy.
This thought experiment leaves both John and Bob unsatisfied, and we see three reasons for that below:
2.1. Problem 1: Measure decline
Problem 1 is a problem for John.
Measure could be defined as a share of an observer of a given type between all possible observers. If the typical size of the simulated mind is, say, 10^15 bites, the chances that a randomly generated mind will be exactly the needed person is 2^(10^15). In other words, a quantum mind generator results in a measure decline of 2^-(10^15) which is an extremely large number. Even in our thought experiment 1 measure decline is 1030 times.
Many authors claim that large measure decline should be treated as death or as an infinitely small chance of survival. Such discussions appeared in the context of so-called quantum immortality, that is, the counterfactual possibility to survive death via existing in quantum multiverse timelines where a person will not die.
Even if the measure decline is not bad per se, it leads to a world where very small probability outcomes will dominate possible futures of an observer, and such parasitic outcomes may be full of suffering. For example, the quantum immortality improbable survival landscape may be dominated by people who are very old and dying but can’t die (it could be patched by signing up for cryonics).
If we use some expected utility calculations, and measure decline results in declining utility of any useful outcome associated with it, we could just ignore my copies with infinitely small measures.
2.2. Problem 2: Non-human and not welcomed minds
Problem 2 is mostly for Bob.
Another problem is that most random minds will be non-human, and will not be adapted to our world, so they will suffer or cause suffering to people living here. In our thought experiment “jYY2№@11” is an example of a non-human random mind.
Such random minds are also extremely bad for any outside observer, like Bob, as he will be very unlikely to meet anyone resembling his friend John Smith.
2.3. Problem 3: Damaged minds
Problem 3 is a problem for both Bob and for John.
Most randomly-created minds will be not minds at all, but some garbage code, or at “best case,” damaged minds. For example, if Bob wants to resurrect John Smith, there will be much more copies where his name (as well as his other properties) is a parody of the name Smith, for example Smthi, Smiht, Misth, Smitt, etc. For n bits long name, there are n individual names which have 1 bit difference.
Thus, for any real person, there will be much larger set of his-her damaged copies, which implies suffering for such a person as the most probable outcome of the quantum random resurrection and s-risks for all people.
3. Patches
Fortunately, quantum random mind uploading could be patched, so it will provide much more satisfaction for John and Bob.
Patch 1. The use of the human mind’s universal model as a starting point
The goal of this patch is to escape minds of “aliens” or of non-workable gibberish code, and thus prevent suffering of most created minds. For example, for a human mind model, his-her possible name will be generated not as random symbols but from the preset of typical human names.
Such a human mind model may look like an untrained neural network which has the general architecture of a human mind, with some other constraints, so any random set of parameters will create a more-or-less normal human mind. We assume that some future assistant AI will be able to find an appropriate model.
In that case, Bob uses a random mind generator for parameters of the universal human mind model. He gets “Maria Stuart”. This will increase the share of the worlds where real John Smith is resurrected to 10-10. Both John and Bob are a little bit more satisfied, as Bob gets a human friend, and John increases his measure.
Obviously, some minds may not want to be resurrected, but this could an important parameter in the model, and models, where “resurrection preference = false” will be ignored.
Patch 2. The use of the digital immortality data to create only minds which comply with our expectations
The problem of Bob’s satisfaction could be overcome by the use of Bob’s expectations as priors, if there are no other current of future sources of data about John.
In that case, Bob could use his memories about John S. to create a model of John S. He remembers that John was either John Smith or John Simpson. He uses a random quantum coin to choose between Smith or Simpson, and gets “John Simpson”.
In another branch of the quantum multiverse, where the coin fails tails, John Smith appears, but his measure declines to 0.5. Both John and Bob are partly satisfied. Bob got someone who looks like his friend, but Bob knows that it is not exactly his friend, and that his friend has now smaller measure of existence.
Digital immortality, or indirect mind uploading, is the collecting information about a person while he is alive with hope that future advanced AI may be able to resurrect the person, by creating an advanced model of the personality based on all available information. Such a model will, by definition, satisfy Bob and all other relatives, as all available information has already been taken into account, including all relatives’ expectations. However, large chunks of information will never be known, and thus have to be replaced with some random data. Even if quantum randomness is used to fill the gaps, John will have an infinitely small share of all possible worlds, and in most other worlds he will be replaced by someone else.
Patch 3. The use of multiverse-wide cooperation for the cross-resurrection
The next step is that Bob considers that not only his universe exists, but all possible other universes exist in the Multiverse.
Bob concludes that because all possible observers exist in the Multiverse, his John Simpson created via a quantum random generator is a resurrection of some John Simpson from another universe, while John Smith who lived in our universe, will be resurrected in some other universe where another copy of Bob will do the same experiment.
In other words, Bob and Bob’s copies in other universes cooperate to resurrect the exact John Smith.
As the second universe is exactly the same as ours except for John’s name, there is another exact copy of Bob in it, and this Воb’s copy is also wanting to resurrect his friend John S., so he uses another quantum random mind generator. Now the following happens:
So, the total measure of John Smith has not declined, if Bob takes into account that other copies of Bob in other universes will run the same experiment. By deciding to start the random mind generator (and to not turn off the resulting mind), Bob joins a large group of other minds, who think similarly, but who are located in causally disconnected parts of the Multiverse. Everyone expects that some other random generator recreates an exact copy of their loved one.
In a real case of large missing data, like gigabytes, this requires a simultaneous run of an extremely large number of quantum random mind generators, like 10^(10^9), which is only possible via multiverse-wide cooperation. The measure will not decline in such a case too, as for every dead person there will be one random person, and given the large numbers, any person will be randomly recreated, at least in approximately one world. (Some may go deeper and take into account standard deviation, but because we use quantum generators in the many worlds interpretation, each universe creates exactly its share of John, and there will be no fluctuations, which would result in non-existence of some Johns and two copies of another.)
Any of Воb’s copies can join such a multiverse-wide cooperation by creating just one quantum random mind (and treating the resulting mind well).
4. Remaining problems
Multiverse. What if the multiverse doesn’t actually exist? In that case, Bob and John get partly satisfying results, as Bob gets John’s copy, but John’s copy is not perfect from John’s point of view. If the quantum multiverse is not real, but some other form of the multiverse exists, like the one based on inflational cosmology, the resurrection method will still work.
Defection. Bob may not create any random mind generators at all but still expect that someone else will recreate his friend. In general, the rate of defections may be known and compensated by increasing the number of random minds by those who have more resources.
There are several other possible generic problems of multiverse-wide cooperation, including infinite ethics, the possibility of acausal blackmail, a method to measure similarity between agents, and problems with agents that have other values as described in EA post's comment.
Conclusion
I hope that this post may increase one's hope in the future personal resurrection by superintelligent AI.