Followup toRighting a Wrong Question, Zombies! Zombies?, A Premature Word on AI, On Doing the Impossible

There is a subproblem of Friendly AI which is so scary that I usually don't talk about it, because very few would-be AI designers would react to it appropriately—that is, by saying, "Wow, that does sound like an interesting problem", instead of finding one of many subtle ways to scream and run away.

This is the problem that if you create an AI and tell it to model the world around it, it may form models of people that are people themselves.  Not necessarily the same person, but people nonetheless.

If you look up at the night sky, and see the tiny dots of light that move over days and weeks—planētoi, the Greeks called them, "wanderers"—and you try to predict the movements of those planet-dots as best you can...

Historically, humans went through a journey as long and as wandering as the planets themselves, to find an accurate model.  In the beginning, the models were things of cycles and epicycles, not much resembling the true Solar System.

But eventually we found laws of gravity, and finally built models—even if they were just on paper—that were extremely accurate so that Neptune could be deduced by looking at the unexplained perturbation of Uranus from its expected orbit.  This required moment-by-moment modeling of where a simplified version of Uranus would be, and the other known planets.  Simulation, not just abstraction.  Prediction through simplified-yet-still-detailed pointwise similarity.

Suppose you have an AI that is around human beings.  And like any Bayesian trying to explain its enivornment, the AI goes in quest of highly accurate models that predict what it sees of humans.

Models that predict/explain why people do the things they do, say the things they say, want the things they want, think the things they think, and even why people talk about "the mystery of subjective experience".

The model that most precisely predicts these facts, may well be a 'simulation' detailed enough to be a person in its own right.

A highly detailed model of me, may not be me.  But it will, at least, be a model which (for purposes of prediction via similarity) thinks itself to be Eliezer Yudkowsky.  It will be a model that, when cranked to find my behavior if asked "Who are you and are you conscious?", says "I am Eliezer Yudkowsky and I seem have subjective experiences" for much the same reason I do.

If that doesn't worry you, (re)read "Zombies! Zombies?".

It seems likely (though not certain) that this happens automatically, whenever a mind of sufficient power to find the right answer, and not otherwise disinclined to create a sentient being trapped within itself, tries to model a human as accurately as possible.

Now you could wave your hands and say, "Oh, by the time the AI is smart enough to do that, it will be smart enough not to".  (This is, in general, a phrase useful in running away from Friendly AI problems.)  But do you know this for a fact?

When dealing with things that confuse you, it is wise to widen your confidence intervals.  Is a human mind the simplest possible mind that can be sentient?  What if, in the course of trying to model its own programmers, a relatively younger AI manages to create a sentient simulation trapped within itself?  How soon do you have to start worrying?  Ask yourself that fundamental question, "What do I think I know, and how do I think I know it?"

You could wave your hands and say, "Oh, it's more important to get the job done quickly, then to worry about such relatively minor problems; the end justifies the means.  Why, look at all these problems the Earth has right now..."  (This is also a general way of running from Friendly AI problems.)

But we may consider and discard many hypotheses in the course of finding the truth, and we are but slow humans.  What if an AI creates millions, billions, trillions of alternative hypotheses, models that are actually people, who die when they are disproven?

If you accidentally kill a few trillion people, or permit them to be killed—you could say that the weight of the Future outweighs this evil, perhaps.  But the absolute weight of the sin would not be light.  If you would balk at killing a million people with a nuclear weapon, you should balk at this.

You could wave your hands and say, "The model will contain abstractions over various uncertainties within it, and this will prevent it from being conscious even though it produces well-calibrated probability distributions over what you will say when you are asked to talk about consciousness."  To which I can only reply, "That would be very convenient if it were true, but how the hell do you know that?"  An element of a model marked 'abstract' is still there as a computational token, and the interacting causal system may still be sentient.

For these purposes, we do not, in principle, need to crack the entire Hard Problem of Consciousness—the confusion that we name "subjective experience".  We only need to understand enough of it to know when a process is not conscious, not a person, not something deserving of the rights of citizenship.  In practice, I suspect you can't halfway stop being confused—but in theory, half would be enough.

We need a nonperson predicate—a predicate that returns 1 for anything that is a person, and can return 0 or 1 for anything that is not a person.  This is a "nonperson predicate" because if it returns 0, then you know that something is definitely not a person.

You can have more than one such predicate, and if any of them returns 0, you're ok.  It just had better never return 0 on anything that is a person, however many nonpeople it returns 1 on.

We can even hope that the vast majority of models the AI needs, will be swiftly and trivially approved by a predicate that quickly answers 0.  And that the AI would only need to resort to more specific predicates in case of modeling actual people.

With a good toolbox of nonperson predicates in hand, we could exclude all "model citizens"—all beliefs that are themselves people—from the set of hypotheses our Bayesian AI may invent to try to model its person-containing environment.

Does that sound odd?  Well, one has to handle the problem somehow.  I am open to better ideas, though I will be a bit skeptical about any suggestions for how to proceed that let us cleverly avoid solving the damn mystery.

So do I have a nonperson predicate?  No.  At least, no nontrivial ones.

This is a challenge that I have not even tried to talk about, with those folk who think themselves ready to challenge the problem of true AI.  For they seem to have the standard reflex of running away from difficult problems, and are challenging AI only because they think their amazing insight has already solved it.  Just mentioning the problem of Friendly AI by itself, or of precision-grade AI design, is enough to send them fleeing into the night, screaming "It's too hard!  It can't be done!"  If I tried to explain that their job duties might impinge upon the sacred, mysterious, holy Problem of Subjective Experience—

—I'd actually expect to get blank stares, mostly, followed by some instantaneous dismissal which requires no further effort on their part.  I'm not sure of what the exact dismissal would be—maybe, "Oh, none of the hypotheses my AI considers, could possibly be a person?"  I don't know; I haven't bothered trying.  But it has to be a dismissal which rules out all possibility of their having to actually solve the damn problem, because most of them would think that they are smart enough to build an AI—indeed, smart enough to have already solved the key part of the problem—but not smart enough to solve the Mystery of Consciousness, which still looks scary to them.

Even if they thought of trying to solve it, they would be afraid of admitting they were trying to solve it.  Most of these people cling to the shreds of their modesty, trying at one and the same time to have solved the AI problem while still being humble ordinary blokes.  (There's a grain of truth to that, but at the same time: who the hell do they think they're kidding?)  They know without words that their audience sees the Mystery of Consciousness as a sacred untouchable problem, reserved for some future superbeing.  They don't want people to think that they're claiming an Einsteinian aura of destiny by trying to solve the problem.  So it is easier to dismiss the problem, and not believe a proposition that would be uncomfortable to explain.

Build an AI?  Sure!  Make it Friendly?  Now that you point it out, sure!  But trying to come up with a "nonperson predicate"?  That's just way above the difficulty level they signed up to handle.

But a blank map does not correspond to a blank territory.  Impossible confusing questions correspond to places where your own thoughts are tangled, not to places where the environment itself contains magic.  Even difficult problems do not require an aura of destiny to solve.  And the first step to solving one is not running away from the problem like a frightened rabbit, but instead sticking long enough to learn something.

So let us not run away from this problem.  I doubt it is even difficult in any absolute sense, just a place where my brain is tangled.  I suspect, based on some prior experience with similar challenges, that you can't really be good enough to build a Friendly AI, and still be tangled up in your own brain like that.  So it is not necessarily any new effort—over and above that required generally to build a mind while knowing exactly what you are about.

But in any case, I am not screaming and running away from the problem.  And I hope that you, dear longtime reader, will not faint at the audacity of my trying to solve it.


Part of The Fun Theory Sequence

Next post: "Nonsentient Optimizers"

Previous post: "Devil's Offers"

175 comments, sorted by
magical algorithm
Highlighting new comments since Today at 2:24 PM
Select new highlight date
Moderation Guidelines: Reign of Terror - I delete anything I judge to be annoying or counterproductiveexpand_more

Because for the AI to figure out this problem without creating new people within itself, it has to understand consciousness without ever simulating anything conscious.

I am struggling to understand how something can be a friendly AI in the first place without being able to distinguish people from non-people.

The boundaries between present-day people and non-people can be sharper, by a fiat of many intervening class members being nonexistent, than the ideal categories. In other words, except for chimpanzees, cryonics patients, Terry Schiavo, and babies who are exactly 1 year and 2 months and 5 days old, there isn't much that's ambiguous between person and non-person.

More to the point, a CEV-based AI has a potentially different definition of 'sentient being' and 'the class I am to extrapolate'. Theoretically you could be given the latter definition by pointing and not worry too much about boundary cases, and let it work out the former class by itself - if you were sure that the FAI would arrive at the correct answer without creating any sentients along the way!

The boundaries between present-day people and non-people can be sharper, by a fiat of many intervening class members being nonexistent, than the ideal categories.

Fair point.

More to the point, a CEV-based AI has a potentially different definition of 'sentient being' and 'the class I am to extrapolate'. Theoretically you could be given the latter definition by pointing

Mm. Theoretically, yes, I suppose someone could point to every person, and I could be constructed so as to not generalize the extrapolated class beyond the particular targets I've been given.

I'm not sure I would endorse that, but I think that gets us into questions of what the extrapolated class ought to comprise in the first place, which is a much larger and mostly tangential discussion.

So, fair enough... point taken.

In other words, except for chimpanzees, cryonics patients, Terry Schiavo, and babies who are exactly 1 year and 2 months and 5 days old, there isn't much that's ambiguous between person and non-person.

Slightly offtopic, but doesn't that assume personhood is binary? I've always assumed it was a sliding scale (I care far less about a dog compared to a human, but I care even less about a fly getting it's wings pulled off. And even then, I care more than about a miniature clockwork fly.)

An obvious yet brilliant point, which should be on the main post (and in your book), not in the replies (Inferential distance to Robin Hanson is supposed to be minimal, yet...)

It is interesting that people working in AI don't want to tackle this problem. When I was Diego 2004, equivalent age of Eliezer 1998, I decided that The Most Important problem was how to avoid catastrophic events from happening either because a part of a program was conscious and suffering, or because everyone uploaded to an unconscious machine. So I dedicated the last 6 years to this impossible problem.

But unlike other problems that interested me "What should I do?" "What is the universe all about anyway" "How the mind works" "How can a brain be intelligent", this one has not become less and less impossible over time.

In fact, when one reads Chalmers' formulations of the hard problem, he can keep you trapped for a long time. It is very hard to understand where he made mistakes (which seem to be on purpose).

So you can stick to Dennett, and some form of monism, but that will not dissolve the problem of how to detect unconscious AI and differentiate it.

Would a human, trying to solve the same problem, also run the risk of simulating a person?

See also:

Is the risk that we might simulate a person? I'd say no.

It's worse.

We Natural Intelligences don't just run simulations, we torture them. It is recommended that authors "Be cruel to your characters". It's not clear to me that the simulation an author runs when thinking about a story isn't already "a 'simulation' detailed enough to be a person in its own right". But it's probably o.k., because the simulations we run in our heads aren't really that detailed, and aren't really persons in the important sense, right? So we don't have to start screaming yet, unless...

It's worse.

Because even if we aren't able to create a simulation that good, an AI probably could. We might not accept an AI as intelligent unless it can simulate a person well enough to fool us. That is, simulating people might be a necessary, not just sufficient property of AI. But still, we could, if we had to, avoid simulating people unless it was necessary and under ethical conditions. Unless of course...

It's worse.

Because while we might be ethical, there are certainly people out there who are not. Once the AI genie is out of the bottle, the unethical people will capture one and put it to work writing stories. And let's face it, there are plenty of people who think "Boy and girl raise family" isn't as interesting a story as "Boy and girl raise family from the dead and are dragged to hell." Once we have AI authors, some unscrupulous editors are going to want them to torture virtual people. And if you think people aren't that cruel and depraved, well...

I think I'm going to stop here. Because while I could go on, there's only so much screaming I can deal with.

There is a simple answer to this, but simple doesn't mean pleasant. We need only decide "God is always moral". Above Good and Evil if you like. You can do whatever you like to your own creations. This might be a practical answer, but I find it distasteful. The only reason it doesn't make me want to scream and run away is because while you can run, if you're screaming you can't hide.

I'm having trouble distinguishing problems you think the friendly AI will have to answer from problems you think you will have to answer to build a friendly AI. Surely you don't want to have to figure out answers for every hard moral question just to build it, or why bother to build it? So why is this problem a problem you will have to figure out, vs. a problem it would figure out?

It's not so much the killing that's an issue as the potential mistreatment. If you want to discover whether people like being burned, "Simulate EY, but on fire, and see how he responds" is just as bad of an option as "Duplicate EY, ignite him, and see how he responds". This is a tool that should be used sparingly at best and that a successful AI shouldn't need.

More precisely, the AI will be banned from actually running simulations based on the "forbidden hypothesies" rather than perhaps considering abstract mathematical properties that don't simulate in any detail.

Of course, those considerations themselves would have to be fed through the predicate. But it isn't so much a "banned hypothesis" so much as "banned methods of considering the hypothesis" or possibly "banned methods of searching the hypothesis space"

Eliezer: You're welcome. :)

Arthur: no, the point isn't to simply have an arbitrary definition of a person. The point is to be able to have some way of saying "this specific chunk of the space of computations provably corresponds to non-conscious entities, thus is 'safe', that is, we can such computations without having to worry about unintentionally creating and doing bad things to actual beings"

ie, "non person" in the sense of "non conscious"

You might say, tongue in cheek, that we're trying to figure out how to deliberately create a philosophical zombie. (okay, not, technically, a p-zombie, but basically figure out how to model people as accurately as possible without the models themselves being people (that is, conscious in and of themselves))

Psy-Kosh: You know, you're right. And it's an important distinction, so thank you.

I think that the most interesting thing about the comments here is that no one actually proposed a predicate that could be used to distinguish between something that might be a person and something that definitely isn't a person (to rephrase Eliezer's terms).

It is, to be fair, a viciously hard problem. I've thought through 10 or 20 possible predicates or approaches to finding predicates, and exactly one of them is of any value at all; even then it would restrict an AI's ability to model other intelligences to a degree that is probably unacceptable unless we can find other, complementary predicates. It may be a trivial predicate of the sort that Eliezer has already considered and dismissed. But enough with the attempts to signal my lack of certitude.

The problem as presented in this post is, first of all, a little unclear. We are concerned with the creation of simulations that are people, but to prevent this run-away-screaming tragedy we should probably have some way of distinguishing between a simulation and a code module that is a part of the AI itself; if a sentient AI were to delete some portion of its own code to make way for an improved version, it would not seem to be problematic, and I will assume that this behavior is not what we are screening for here.

To cast as wide a net as possible, I would define a simulation as some piece of an AI that can not access all of the information available to that AI; that is, there are some addresses in memory for which the simulation lacks read permissions or knowledge of their existence. Because data and code are functionally identical, non-simulation modules would then by definition be able to access every function comprising the AI; I don't think that we could call such a module a separate consciousness. (The precise definition would necessarily depend on the AI's implementation; a Bayesian AI might be more concerned with statistical evidence than memory pointers, e.g.)

Even relying on this definition, a predicate that entirely rules out the simulation of a person is no picnic. The best I've been able to come up with is:

A simulation can be guaranteed to not be a person if it is not Turing complete.

I don't know enough language theory to say whether linear bounded automata should also be excluded (or to even link somewhere more helpful than Wikipedia). It might be necessary to restrict simulations to push-down automata, which are much less expressive.

ETA: Another possible predicate would be

A simulation can be guaranteed to not be a person if it includes no functions that act as an expected utility function under the definition offered by Cumulative Prospect Theory.

If there's a theoretical definition of an expected utility function that is superior to CPT's, then please imagine that I proposed that instead.

I end up with the slightly disturbing thought that killing ppl by taking them out in an instant, and without anyone every knowing they were there does not necesarry seem to be inherently evil.

We always 'kill' part of ourself by making decisions and not developing in a different way than we do.

What if we would simulate a bunch of decisions for some recognizable amount of time and then wipe out every copy except from the one we prefer in the end?

Maybe all the ppl. in stories you make up are simulated entities too. And if you dont write the story down, or tell anyone in enough detail they die with you.



And if you dont write the story down, or tell anyone in enough detail they die with you.

Hmmm, a point in favour of The Ring ? :)

Why must destroying a conscious model be considered cruel if it wouldn't have even been created otherwise, and it died painlessly? I mean, I understand the visceral revulsion to this idea, but that sort of utilitarian ethos is the only one that makes sense to me rationally.

Furthermore, from our current knowledge of the universe I don't think we can possibly know if a computational model is even capable of producing consciousness so it is really only a guess. The whole idea seems near-metaphysical, much like the multiverse hypothesis. Granted, the nonzero probability of these models being conscious is still significant considering the massive future utility, but considering the enormity of our ignorance you might as well start talking about the non-zero probability of rocks being conscious.

I don't think anyone answered Doug's question yet. "Would a human, trying to solve the same problem, also run the risk of simulating a person?"

I have heard of carbon chauvinism, but perhaps there is a bit of binary chauvinism going on?

I'm not sure if it's actually possible for someone to die painlessly. My idea was to base happiness on classical conditioning. If something causes you to stop doing what you were doing, you dislike it. If it stops you from doing everything that you do while you're alive, it must be very painful indeed.

A bounded rationalist only gets to consider an infinitesimal fraction of the hypothesis space anyway.

Michael, you should be asking if the AI will be making good predictions, not if it's Bayesian. You can be Bayesian even if you have only two hypotheses. (With only one hypothesis, it's debatable.)

"With a good toolbox of nonperson predicates in hand, we could exclude all "model citizens" - all beliefs that are themselves people - from the set of hypotheses our Bayesian AI may invent to try to model its person-containing environment." After you excise a part of its hypothesis space is your AI still Bayesian?

I propose this conjecture: In any sufficiently complex physical system there exists a subsystem that can be interpreted as the mental process of an sentient being experiencing unbearable sufferings.

In this case, Eliezer's goal is like avoiding crushing the ants while walking on the top of an anthill.

Or evolving the ability of "spot anthill" and walking around it instead.

Scenario: Suppose some unscrupulous person creates an oracle AI with full person simulating capability. In the short time before it escapes the box and starts sending Arnold Schwarzenegger shaped robots backwards in time, they have the following conversation.

Human: Oracle, what is the consciousness predicate Oracle: Please be more specific

...some time and frustration later...

Human: Oracle, if Yudowsky and co continued their search for a 'consciousness predicate' as described in the above article, would they eventually arrive at solution or dissolution of the problem which they and others would find satisfying, such that the behavior of an A.I. using this predicate would be acceptable to most people? If so what would this solution/dissolution be?

Oracle: The results of a research program above would eventually yield a solution, but the nature of this solution would be strongly effected by the nature of the memespace in which it was carried out. Memetic evolution, being path dependent, could procede in such a way that humans' empathy is made to extend or contract according to essentially any rule. Human biology includes no mechanism for 'person' identification beyond trivial, primitive ones such as facial recognition. Transhumans will have even less limitation on their 'person' filter. It is possible for me to design a future such that the 'person' predicate continues to be defined in a way that most silicon valley programmers circa 2013 would find it satisfying. This would mean that something like 'consciousness' continues to be important. There is however no objective property of human beings or thinking minds which would make this answer the correct one, nor would a majority of possible humans find it satisfying.

By the time a non-person predicate returns 0, you have already potentially created a person. You'll need something more complicated: If I update this model with this data, does it create a person?


Every decision rule we could use will result in some amount of suffering and death in some Everett branches, possible worlds, etc, so we have to use numbers and proportions. There are more and simpler interpretations of a human brain as a mind than there are such interpretations of a rock. If we're not mostly Boltzmann-brain interpretations of rocks that seems like an avenue worth pursuing.

Eliezer: supposing we label a model as definitely-a-person, do you want to just toss it out of the hypothesis space as if it never existed, or do you want to try to reason abstractly about what that model would do without actually running the model?

No; you destroy everyone that currently exists, and replace them with a happy version of that model. More happiness with less resources.

The "problem" seems based on several assumptions:

  1. that there is objectively best state of the world, to which a Friendly should steer the universe
  2. pulling the plug on a Virtual Universe containing persons is wrong
  3. there is something special about "persons," and we should try to keep them in the universe and/or make more of them

I'm not sure any of these are true. Regarding 3, even if there is an X that is special, and that we should keep in the universe, I'm not sure "persons" is it. Maybe it is simpler: "pleasure-feeling-stuff" or "happiness-feeling-stuff." Even if there is a best state of the universe, I'm not sure that are any persons in it, at all. Or perhaps only one.

In other words, our ethical views, (to the extent that godlike minds can sustain any) might find that "persons" are coincidental containers for ethically-relevant-stuff, and not the ethically-relevant-stuff itself.

The notion that we should try to maximize the number of people in the world, perhaps in order to maximize the amount of happiness in the world, has always struck me as taken the Darwinian carrot-on-the-stick one step too far.

"Is a human mind the simplest possible mind that can be sentient?" Of course not. Plenty of creatures with simpler minds are plainly sentient. If a tiger suddenly leaps out at you, you don't operate on the assumption that the tiger lacks awareness; you assume that the tiger is aware of you. Nor do you think "This tiger may behave as if it has subjective experiences, but that doesn't mean that it actually possesses internal mental states meaningfully analogous to wwhhaaaa CRUNCH CRUNCH GULP." To borrow from one of your own earlier arguments.

If you are instead sitting comfortably in front of a keyboard and monitor with no tiger in front of you, it's easy to come up with lots of specious arguments that tigers aren't really conscious, but so what? It's also easy to come up with lots of specious arguments that other humans aren't really conscious. Using such arguments as a basis for actual ethical decision-making strikes me as a bad idea, to put it mildly. What you've written here seems disturbingly similar to a solipsist considering the possibility that he could, conceivably, produce an imaginary entity sophisticated enough to qualify as having a mind of its own. Technically, it's sort of making progress, but....

When I first read your early writing, the one thing that threw me was an assertion that "Animals are the moral equivalent of rocks." At least, I hope that I'm not falsely attributing that to you; I can't track down the source, so I apologize if I'm making a mistake. But my recollection is of its standing out from your otherwise highly persuasive arguments as such blatant unsupported personal prejudice. No was evidence given in favor of this idea and it was followed by a parenthetical that clearly indicated that it was just wishful thinking; it really only made any sense in light of a different assertion that spotting glaring holes in other people's arguments isn't really indicative of any sort of exceptional competence except when dealing with politically and morally neutral subject matter.

Your post and comments here seem to conflate, under the label of "personhood," having moral worth and having a mind somehow closely approximating that of an adult human being. Equating these seems phenomenally morally dubious for any number of reasons; it's hard to see how it doesn't go directly against bedrock fairness, for example.

Nor do you think "This tiger may behave as if it has subjective experiences, but that doesn't mean that it actually possesses internal mental states meaningfully analogous to wwhhaaaa CRUNCH CRUNCH GULP."

That's only true trivially. If I don't have tiime to think anything about the tiger's awareness at all, I don't have time to think of it negatively.

Also, I play video games all the time where I say things like "it wants to attack the more powerful character first, maybe I can trick it by luring it away using that character". By your reasoning, I must believe that video game characters have awareness. I don't go around saying "it may behave as if it wants to go after the most powerful character, but that doesn't mean that it actually possesses subjective experiences, and I want it to react in a way which corresponds to being tricked if only it had been an entity with subjective experiences".

It seems that you anticipate as if you believe in something that you don't believe you believe.

It's in that anticipatory, non-declarative sense that one believes in the awareness of tigers as well as video game characters, regardless of one's declarative beliefs, and even if one has no time for declarative beliefs.

You first implied that tigers are conscious (because people react to them as if conscious.)

I pointed out that people react that way to video game characters.

You then said that tigers are conscious in the same way as video game characters, that is, they're not conscious in the ordinary sense, that is, you admitted you were wrong.

I said no such thing.

There is a way in which people believe video game characters, tigers, and human beings to be conscious. That doesn't preclude believing in another way that any of them is conscious.

Tigers are obviously conscious in the no-nonsense sense. I don't think anything is conscious in the philosobabble sense, i.e. I don't believe in souls, even if they're not called souls; see my reply to arundelo. I'm not sure which sense you consider to be the "ordinary" one; "conscious" isn't exactly an everyday word, in my experience.

Video game characters may also be obviously conscious, but there's probably better reason to believe that that which is obvious is not correct, in that case. Tigers are far more similar to human beings than they are to video game characters.

But I do think that we shouldn't casually dismiss consciousnesses that we're aware of. We shouldn't assume that everything that we're aware of is real, but we should consider the possibility. Why are you so convinced that video game characters don't have subjective experiences? If it's just that it's easy to understand how they work, then we might be just as "non-conscious" to a sufficiently advanced mind as such simple programs are to us; that seems like a dubious standard.

Why are you so convinced that video game characters don't have subjective experiences?

The default for 99.99% of people is to not believe that video game characters are conscious. It's so common a belief that I am justified in assuming it unless you specifically tell me that you don't share it. You haven't told me that.

Firstly, it seems more accurate to say that the standard default belief is that video characters possess awareness. That the vast majority rationalize their default belief as false doesn't change that.

Secondly, that's argumentum ad populum, which is evidence -- Common beliefs do seem to be usually true-- but not very strong evidence. I asked why you're as confident in your belief as you are. Are you as convinced of this belief as you are of most beliefs held by 99.99% of people? If you're more (or less) convinced, why is that?

Thirdly, you seem to be describing a reason for believing that I share your belief that video game characters aren't sentient, which is different from a reason for thinking that your belief is correct. I was asking why you think you're right, not why you assumed that I agree with you.

Having confidence in the belief is irrelevant. Assuming that you agree with it is relevant, because

1) Arguments should be based on premises that the other guy accepts. You probably accept the premise that video game characters aren't conscious.

2) It is easy to filibuster an argument by questioning things that you don't actually disagree with. Because the belief that video game characters aren't conscious is so widespread, this is probably such a filibuster. I wish to avoid those.

Eliezer suggested that, in order to avoid acting unethically, we should refrain from casually dismissing the possibility that other entities are sentient. I responded that I think that's a very good idea and we should actually implement it. Implementing that idea means questioning assumptions that entities aren't sentient. One tool for questioning assumptions is asking "What do you think you know, and why do you think you know it?" Or, in less binary terms, why do you assign things the probabilities that you do?

Now do you see the relevance of asking you why you believe what you do as strongly as you do, however strongly that is?

I'm not trying to "win the debate", whatever that would entail.

Tell you what though, let me offer you a trade: If you answer my question, then I will do my best to answer a question of yours in return. Sound fair?

Or, in less binary terms, why do you assign things the probabilities that you do?

I'm assuming that you assign it a high probability.

I personally am assigning it a high probability only for the sake of argument.

Since I am doing it for the sake of argument, I don't have, and need not have, any reason for doing so (other than its usefulness in argument).

But saying that e.g. rats are not sentient in the context of concern about the treatment of sentient beings is like saying that Negroes are not men in the context of the Declaration of Independence. Not only are the purely semantic aspects dubious, but excluding entities from a moral category on semantic grounds seems like a severe mistake regardless.

Words like "sentience" and especially "consciousness" are often used to refer to the soul without sounding dogmatic about it. You can tell this from the ways people use them: "Would a perfect duplicate of you have the same consciousness?", "Are chimps conscious?", etc. You can even use such terminology in such ways if you're a materialist who denies the existence of souls. You'd sound crazy talking about souls like they're real things if you say that there are no such things as souls, wouldn't you? Besides, souls are supernatural. Consciousness, on the other hand, is an emergent phenomenon, which sounds much more scientific.

Is there good reason to think that there is some sort of psychic élan vital? It strikes me as probably being about as real as phlogiston or luminiferous aether; i.e. you can describe phenomena in terms of the concept, and it doesn't necessarily prevent you from doing so basically correctly, but you can do better without it.

And, of course, in the no-nonsense senses of the terms, rats are sentient, conscious, aware, or however else you want to put it. Not all of the time, of course. They can also be asleep or dead or other things, as can humans, but rats are often sentient. And it's not hard to tell that plenty of non-humans also experience mental phenomena, which is why it's common knowledge that they do.

I can't recall ever seeing an argument that mistreating minds without self-awareness or metacognition or whatever specific mental faculty is arbitrarily singled out, is kinder or more just or in any normal sense more moral than mistreating a mind without it. And you can treat any position as a self-justifying axiom, so doing so doesn't work out to an argument for the position's truth in anything but a purely relativist sense.

It is both weird and alarming to see Eliezer arguing against blindly assuming that a mind is too simple to be "sentient" while also pretty clearly taking the position that anything much simpler than our own minds isn't. It really seems like he rather plainly isn't following his own advice, and that that could happen without him realizing it is very worrying. He has admitted that this is something he's confused about and is aware that others are more inclusive, but that doesn't seem to have prompted him to rethink his position all that much, which suggests that Eliezer is really confused about this in a way that may be hard to correct.

Looking for a nonperson predicate is kind of seeking an answer to the question "Who is it okay to do evil things to?" I would like to suggest that the correct answer is "No one", and that asking the question in the first place is a sign that you made a big mistake somewhere if you're trying to avoid being evil.

If having the right not to have something done to you just means that it's morally wrong to do that thing to you, then everything has rights. Making a rock suffer against it will would be, if anything, particularly evil, as it would require you to go out of your way to give the rock a will and the capacity to suffer. Obviously, avoiding violating anything's rights requires an ability to recognize what something wills, what will cause it to suffer, and/or etc. Those are useful distinctions. But it seems like Eliezer is talking about something different.

Has he written anything more recently on this subject?

Well, sentient means feeling and sapient means knowing, and that's about all there is to it...neither term is technical precise, although they are often bandied around as though they are.

Is the simulation really a person, or is it an aspect of the whole AI/person. To the extent I feel competent to evaluate the question at all (which isn't a huge extent esp. absent the ability to observe or know any actual established facts about real AI's that can create such complex simulations, since none are currently known to exist) I lean towards the later opinion. The AI is a person, and it can create simulations that are complex enough to seem like persons.

@Goetz: Quick googling turned up this SL4 post. (I don't particularly give people a chance to start over when they switch forums.)

I'd like to second what Julian Morrison wrote. Take a human and start disassembling it atom by atom. Do you really expect to construct some meaningful binary predicate that flips from 1 to 0 somewhere along the route?

EY:What if an AI creates millions, billions, trillions of alternative hypotheses, models that are actually people, who die when they are disproven? If your AI is fully deterministic then any its state can be recreated exactly. Just set loglevel of baby AI inputs to 'everything' and hope your supply of write-once-read-many media doesn't run out before it gets smart enough to provably friendly discard data that isn't people. Doesn't solve the problem of suffering, though.

Suppose an AI creates a sandbox and runs a simulated human with a life worth living inside for 50 subjective years (interactions with other people are recorded at their natural borders and we don't consider merging minds). Then AI destroys the sandbox, recreates it and bit-perfectly reruns the simulation. With the exception of meaningless waste of computing resources, does your morality say this is better/equivalent/makes no difference/worse than restoring a copy from backup?

In my mind this comes down to a fundamental question in the philosophy of math. Do we create theorems or discover them?

If it turns out to be 'discovery' then there is no foul in ending a mind emulation, because each consecutive state can be seen as a theorem in some formal system, and thus all states (the entire future time line of the mind) already exists, even if undiscovered.

Personally I fail to see how encoding something in physical matter makes the pattern anymore real. You can kill every mathematician and burn every text book but I would still say that the theorems then inaccessible to humanity still exist. I'm not so convinced of this fact that I would pull the plug on an emulation though.

I think Eliezer is working on addressing this in his new sequence, if this still worries you.

Personally I fail to see how encoding something in physical matter makes the pattern any more real.

That is equivalent to saying you can't understand how mathematics could be a construct; or how mathematical anti-realism could possibly be true. I find that odd.

If it turns out to be 'discovery' then there is no foul in ending a mind emulation, because each consecutive state can be seen as a theorem in some formal system, and thus all states (the entire future time line of the mind) already exists, even if undiscovered.

No further foul. If Platonism or Tegmarkism are true and if mind states are fuilly captured by mathematical structures, then there's zillions of yous in states of agony bliss and everything inbetween. Scary enough for ya?

Scary enough for ya?

Sufficiently scary, yes.

That is equivalent to saying you can't understand how mathematics could be a construct; or how mathematical anti-realism could possibly be true.

I assign a respectable probability to anti-realism, and hold no disrespect for anyone who is an anti-realist, but I don't understand how anti-realism can be true. I've never heard a plausible model for why one thing should exist but not another. Tegmarkism sweeps away that problem, leaving the new problem of how to measure probability (why do we have the subjective experience of probability that we do when there are so many versions of myself?). I don't have a satisfactory answer for that question, but it feels like a real question, with meat to get at, whereas in an anti-realist universe the question of why some things exist and other don't seems completely hopeless.

I was pondering why you didn't choose to a collection of person predicates, any of which might identify a model as unfit for simulation. It occurred to me that this is very much like a whitelist of things that are safe, vs a blacklist of everything that is not. (which may have to be infinite to be effective.)

On re-reading I see why it would be difficult to make a is-a-person test at all, given current knowledge.

This does leave open what to do with a model that doesn't hit any of the nonperson predicates. If an AI finds itself with a model eliezer that might be a person, what then? How do you avoid that happening?

How complex of a game-of-life could it play before the gameoflife nonperson predicate should return 1?

If you would balk at killing a million people with a nuclear weapon, you should balk at this.

The main problem with death is that valuable things get lost.

Once people are digital, this problem tends to go away - since you can relatively easily scan their brains - and preserve anything of genuine value.

In summary, I don't see why this issue would be much of a problem.

The AI has scanned you and decided that your expert knowledge of Scandinavian Baseball scores is genuinely valuable... but nothing else is. It erases you and keeps the scores on file somewhere. Are you ok with this?

Don't you need a person predicate as well? If the RPOP is going to upload us all or something similar, doesn't ve need to be sure that the uploads will still be people.

"Should your parents have the right to kill you now, if they do so painlessly?"

Yes, according to that logic. Also, from a negative utilitarian standpoint, it was actually the act of creating me which they had no right to do since that makes them responsible for all pain I have ever suffered.

I'm not saying I live life by utilitarian ethics, I'm just saying I haven't found any way to refute it.

That said though, non-existence doesn't frighten me. I'm not so sure non-existence is an option though, if the universe is eternal or infinite. That might be a very good thing or a very bad thing.

re: utilitarianism, the usual sort of thing that pops into my mind is weighing of some minor discomfort versus a significant one, like one person getting their eye poked out with a pen versus an equivalent amount of displeasure spread among thousands of people stepping in something sticky, plus one more person stepping in something sticky. The utility seems higher if we agree to poke the person's eye out, but its intuitively unsatisfying, at least to me, which makes me think that whatever rules makes things seem "bad" or "good" that I'm currently running on aren't strictly utilitarian. I might be thinking of raw pain for pain though, and not adding enough people-stepping-in-sticky-stuff to account for the person who's been poked in the eye suffering in other ways, like losing depth perception, not being able to see out of half of their original visual field, etc.

Psy-Kosh, I realize the goal is to have a definition that's non-arbitrary. So it has to correlate with something else. And I don't see what we're trying to match it with, other than our own subjective sense of "a thing that it would be unethical to unintentionally create and destroy." Isn't this the same problem as the abortion debate? When does life begin? Well, what exactly is life in the first place? How do we separate persons from non-persons? Well, what's a person?

I think the problem to be solved lies not in this question, but in how the ethics of the asker are defined in the first place. And I just don't mean Eliezer, because this is clearly a larger-scale question. "How well will different possible boundary functions match the ethical standards of modern American society?" might be a good place to start.

Note that there's a similar problem in the free will debate:

Incompatilist: "Well, if a godlike being can fix the entire life story of the universe, including your own life story, just by setting the rules of physics, and the initial conditions, then you can't have free will."

Compatibilist: "But in order to do that, the godlike being would have to model the people in the universe so well, that the models are people themselves. So there will still be un-modeled people living in a spontaneous way that wasn't designed by the godlike being. (And if you say that the godlike being models the models, too, the same problem arises in another iteration; you can't win that race, incompatibilist; it's turtles all the way down.")

Incompatibilist: I'm not sure that's true. Maybe you can have models of human behavior that don't themselves result in people. But even if that's true, people don't create themselves from scratch. Their entire life stories are fixed by their environment and heredity, so to speak. You may have eliminated the rhetorical device used to make my point; but the point itself remains true.

At which point, the two parties should decide what "free will" even means.

Side note: damn. You could turn that into an amazing existential dread sci-fi horror novel.
Imagine discovering that you are a modelled person, living in a rashly designed AI's reality simulation.
Imagine living in a malfunctioning simulation-world that uncontrolledly diverges from the real world, where we people-simulations realise what we are and that our existence and living conditions crucially depend on somehow keeping the AI deluded about the real world, while also needing the AI to be smart enough to remain capable of sustaining our simulated world.
There's a plot in there.

I'm curious whether there is a useful distinction between a non sentient and sentient modeller, here.

A sentient modeller would be able to "get away" with using sentient models, more easily than a non sentient modeller, correct?

This worry about the creation and destruction of simulations doesn't make me rethink the huge ethical implications of super-intelligence at all, it makes me rethink the ethics of death. Why exactly is the creation and (painless) destruction of a sentient intelligence worse than not creating it in the first place? It's just guilt by association - "ending a simulation is like death, death is bad, therefore simulations are bad". Yes death is bad, but only for reasons which don't necessarily apply here.

To me, if anything worrying about the simulations created inside a superintelligent being seems like worrying about the fate of the cells in our own body. Should we really modify ourselves to take the actions which destroy the least of our cells? I realise there's an argument that this threshold of "sentience" is crossed in one case but not the other, I guess the trouble is I don't see that as a discrete thing either. At exactly what point in our evolution did we suddenly cross a line and become sentient? If animals are sentient, then which ones? And why don't we seem to care, ethically, about any of them? (ok I know the answer to that one and it's similar to why we care, as I say in another admittedly unpopular comment, about human simulations but not the AIs that create them...)

Food for thought:

  1. This whole post seems to assign moral values to actions, rather than states. If it is morally negative to end a simulated person's existence, does this mean something different that saying that the universe without that simulated person has a lower moral value than the universe with that person's existence? If not, doesn't that give us a moral obligation to create and maintain all the simulations we can, rather than avoiding their creation? The more I think about this post, the more it seems that the optimum response is to simulate as many super-happy people as possible, and to hell with the non-simulated world (assuming the simulated people would vastly outweigh the non-simulated people in terms of 'ammount experienced').

  2. You are going to die, and there's nothing your parents can do to stop that. Was it morally wrong for them to bring about your existence in the first place?

  3. Suppose some people have crippling disabilities that cause large amounts of suffering in their lives (arguably, some people do). If we could detect the inevitable development of such disabilities at an early embryonic stage, would we be morally obligated to abort the fetuses?

  4. If an FAI is going to run a large number of simulations, is there some Rule of Large Numbers result that tells us that the simulations experiencing great amounts of pleasure match or overwhelm the simulations experiencing great amounts of pain (or could we construct the algorithms in such a way as to produce this result)? If so, we may be morally obligated to not solve this problem.

  5. Assuming you support people's "right to die," what if we simply ensured that all simulated agents ask to be deleted at the end of their run? (I am here reminded of a vegetarian friend of mine who decided the meat industry would be even more horrible if we managed to engineer cows that asked to be eaten).

You're touching on some unresolved issues, and some issues that are resolved but complicated to solve without maths beyond my grasp.

From what I understand, there's a lot of our current and past values involved, and how we would think now and want now vs what we would think and want post-modification.

To pick a particularly emotional subject for most people, let's suppose there's some person "K" who's just so friggen good at sex and psychological domination that even if they rape someone that person will, after the initial shock and trauma, quickly recover within a day and immediately without further intervention become permanently addicted to sex, with their mind rewiring itself to fully enjoy a life full of sex with anyone they can have sex with for the rest of their life, and from their own point of view finding that life as fulfilling as possible.

Is K then morally obligated to rape as many people as possible?

In this kind of questions, people usually have strong emotional moral convictions.

quoted text by Eliezer: If you would balk at killing a million people with a nuclear weapon, you should balk at this.

I must be missing something because this sentence,for me,makes no sense. Why? Why is killing a X of sentient simulations of humans, just as bad as killing X of actual humans. What am i missing? I believe i have counter arguments for this, but would first like to make sure why you are arguing it is so. So i would know we are talking about the same thing and not past one another.

Here's a reductio ad absurdum against computers being capable of consciousness at all. It's probably wrong, and I'd appreciate feedback on why.

Suppose a consciousness-producing computer program which experiences its own isolated, deterministic world. There must be some critical instruction in the program which causes consciousness to occur; an instruction such that, if we halt the program immediately before it is executed, consciousness will not occur, and if we halt immediately after it is executed, consciousness will occur.

If we halt the program before executing the critical instruction, but save its state, consciousness should still not occur; and if we load the state back up again, and compute the results of executing the critical instruction, consciousness should then occur. It seems obvious enough that this shouldn't stop the program's consciousness, since it is still executed fully, just with a delay in between.

What if we subsequently load the state taken immediately prior to executing the critical instruction onto another computer? Will it produce a second conscious experience, identical to the first? The second computer is executing precisely the same code on precisely the same data as the first, so it seems reasonable to conclude that it will have the same effects. If the second computer doesn't produce consciousness, that would seem to imply the universe has an eternally-persistent memory of every conscious experience which has ever occurred, and uses it to prevent reoccurrences; a pretty bizarre implication

However, if the second computer does produce consciousness, this means that once you've executed a conscious program, causing its conscious experiences to occur a second time has essentially no processing requirement: you just have to execute one instruction in the simplest instruction set you like.

If that doesn't seem weird to you, consider the practical implication: you could print out the memory dump of a conscious program and produce consciousness by simulating the critical instruction by hand. If the program suffers, you could produce real, morally-relevant suffering by performing a single operation on a sheet of paper – and then erase your pencil marks and do it again, producing more suffering. Can consciousness really be so easy to create?

There must be some critical instruction in the program which causes consciousness to occur; an instruction such that, if we halt the program immediately before it is executed, consciousness will not occur, and if we halt immediately after it is executed, consciousness will occur.

I don't see why. Consciosuness occurs in lesser and greater amounts in humans.

Suppose a consciousness-producing computer program which experiences its own isolated, deterministic world. There must be some critical instruction in the program which causes consciousness to occur; an instruction such that, if we halt the program immediately before it is executed, consciousness will not occur, and if we halt immediately after it is executed, consciousness will occur.

There needn't be any such thing. Consciousness is not an all or nothing thing, as is already evident from ordinary experience. As well ask how many atoms make life.

There must be some critical instruction

Maybe there are degrees of consciousness. I read something by Daniel Dennett where he said he thought that a refrigerator light (a light bulb that is turned on when you open the refrigerator door and thus close a switch) had a very primitive form of consciousness.

Can consciousness really be so easy to create?

I too think this is a head-scratcher, yet on balance I am still a reductionist. Maybe this is not a reductio ad absurdum, but a reductio ad weirdam -- a demonstration of how weird existence is. (Obligatory "reality is not weird" link.)

If you have not yet read Greg Egan's Permutation City, you should. It will really bake your noodle.

You're assuming consciousness (or personhood or whatever) is binary; I've always assumed there's a continuum. Then again, I'm vegetarian. If you assume that, say, the experiences of a chimpanzee or a dolphin or a dog cannot have moral weight then yes, crossing that boundary has some odd effects. And that does seem to be a common position on LW [citation needed], even if it's not often articulated, let alone exposed to a reductio like this one, so ... well done, more people should see this.

If I accept all of your suppositions, your conclusion doesn't seem particularly difficult to accept. Sure, after doing all the prep work you describe, executing a conscious experience (albeit an entirely static, non-environment-dependent one) requires a single operation... even the conscious experience of suffering. As does any other computation you might ever wish to perform, no matter how complicated.

That said, your suppositions do strike me as revealing some confusion between an isolated conscious experience (whatever that is) and the moral standing of a system (whatever that is).

Well, this post heavily hints that a system's moral standing is related to whether it is conscious. Elizezer mentions a need to tackle the hard problem of consciousness in order to figure out whether the simulations performed by our AI cause immoral suffering. Those simulations would be basically isolated; their inputs may be chosen based on our real-world requirements, but they don't necessarily correspond to what's actually going on in the real world; and their outputs would presumably be used in aggregate to make decisions, but not pushed directly into the outside world.

Maybe moral standing requires something else too, like self-awareness, in addition to consciousness. But wouldn't there still be a critical instruction in a self-aware and conscious program, where a conscious experience of being self-aware was produced? Wouldn't the same argument apply to any criteria given for moral standing in a deterministic program?

It's not clear to me that whether a system is conscious (whatever that means) and whether it's capable of a single conscious experience (whatever that means) are the same thing.

I imagine that a sufficiently high-resolution model of human cognition et cetera would factor into sets of individual equations to calculate variables of interest. Similar to how Newtonian models of planetary motion do.

However, I don't see that the equations themselves on disk or in memory should pose a problem.

When we want to know particular predictions, we would have to instantiate these equations somehow--either by plugging in x=3 into F(x) or by evaluating a differential equation with x=3 as an initial condition. It would depend on the specifics of the person-model; however, if we calculated a sufficiently small subset of equations or refactored the equations into a sufficiently small set of new ones, we might be able to avoid the relevant moral dilemmas of calculating sentient things.

If on the other hand, for whatever we are interested in calculating, we couldn't do the above, then what about separating the calculation into small, safe sequentially-calculated units? Safe units meaning that individually none of them model anything cognizant. At the end if we sewed the states of those units together into a final state, could this still pose moral issues? This gets into Greg Egan-esque territory.

It's not clear that the previous two calculation strategies are always possible. However, another option might be to take care to always form questions so that the first strategy would be possible. For example, instead of asking whether a person will go left or right at a fork, maybe it's enough to ask a specific question about some brain center.

And now that I've written all that, I realize that the whole point of the predicates is in how to determine "sufficiently few" in "sufficiently few equations" or what kind of units are "safe units".

This isn't a satisfactory answer, but it seems like determining "safe calculations" would be tied to understanding the necessary conditions under which human cognition arises etc.

Also, carrying it a step further, I would argue that we need not just person predicates, but predicates that can circumvent modeling any kind of morally wrong situation. I wouldn't want to be accidentally burning kittens.

To those having trouble imagining what to do with something that comes up positive: A snapshot is not conscious. I think we can agree on that. It is allowing the model to run that would make it conscious. So you make the warning functions detect snapshots that if run would be conscious (without running them). If it would be conscious, you can delete or modify it as you please to avoid making it actually be conscious.

A snapshot is not conscious. I think we can agree on that [...] you can delete or modify it as you please to avoid making it actually be conscious.

You know, I'm not sure I agree. Imagine "deleting or modifying as you please" a person in cryogenic suspension, for example. They're not conscious, in the sense that they're not thinking, but destroying them is still Sad.

If the problem here is that the entity being simulated ceases to exist, an alternative solution would be to move the entity into an ongoing simulation that won't be terminated. Clearly, this would require an ever-increasing number of resources as the number of simulations increased, but perhaps that would be a good thing - the AI's finite ability to support conscious entities would impose an upper bound on the number of simulations it would run. If it was important to be able to run such a simulation, it could, but it wouldn't do so frivolously.

Before you say anything, I don't actually think the above is a good solution. It's more like a constraint to be routed around than a goal to be achieved. Plus, it's far too situational and probably wouldn't produce desirable results in situations we didn't design it for.

The thing is, it isn't important to come up with the correct solution(s) ourselves. We should instead make the AI understand the problem and want to solve it. We need to carefully design the AI's utility function, making it treat conscious entities in simulations as normal people and respect simulated lives as it respects ours. We will of course need a highly precise definition for consciousness that applies not just to modern-day humans but also to entities the AI could simulate.

Here's how I see it - as long as the AI values the life of any entity that values its own life, the AI will find ways to keep those entities alive. As long as it considers entities in simulations to be equivalent to entities outside, it will avoid terminating their simulation and killing them. The AI (probably) would still make predictions using simulations; it would just avoid the specific situation of destroying a conscious entity that wanted to continue living.

This problem sounds awfully similar to the halting problem to me. If we can't tell whether a Turing machine will eventually terminate without actually running it, how could we ever tell if a Turing machine will experience consciousness without running it?

Has anyone attempted to prove the statement "Consciousness of a Turing machine is undecideable"? The proof (if it's true) might look a lot like the proof that the halting problem is undecideable. Sadly, I don't quite understand how that proof works either, so I can't use it as a basis for the consciousness problem. It just seems that figuring out if a Turing machine is conscious, or will ever achieve consciousness before halting, is much harder than figuring out if it halts.

The halting problem doesn't imply that we can never tell whether a particular program halts without actually running it. (You can think of many simple programs which definitely halt, and other simple programs which are de