I would be fine with FAI removing existential risks and not doing any other thing until everybody(’s CEV) agrees on it. (I assume here that removing existential risks is one such thing.) And an FAI team that creditably precommitted to implementing CEV instead of CEV would probably get more resources and would finish first.
I would be fine with FAI removing existential risks and not doing any other thing until everybody(’s CEV) agrees on it.
So what makes you think everybody’s CEV would eventually agree on anything more?
A FAI that never does anything except prevent existential risk—which, in a narrow interpretation, means it doesn’t stop half of humanity from murdering the other half—isn’t a future worth fighting for IMO. We can do so much better. (At least, we can if we’re speculating about building a FAI to execute any well-defined plan we can come up with.)
(I assume here that removing existential risks is one such thing.)
I’m not even sure of that. There are people who believe religiously that End Times must come when everyone must die, and some of them want to hurry that along by actually killing people. And the meaning of “existential risk” is up for grabs anyway—does it preclude evolution into non-humans, leaving no members of original human species in existence? Does it preclude the death of everyone alive today, if some humans are always alive?
Sure, it’s unlikely or it might look like a contrived example to you. But are you really willing to precommit the future light cone, the single shot at creating an FAI (singleton), to whatever CEV might happen to be, without actually knowing what CEV produces and having an abort switch? That’s one of the defining points of CEV: that you can’t know it correctly in advance, or you would just program it directly as a set of goals instead of building a CEV-calculating machine.
And an FAI team that creditably precommitted to implementing CEV instead of CEV would probably get more resources and would finish first.
This seems wrong. A FAI team that precommitted to implementing CEV would definitely get the most funds. Even a team that precommitted to CEV might get more funds than CEV, because people like myself would reason that the team’s values are closer to my own than humanity’s average, plus they have a better chance of actually agreeing on more things.
A FAI that never does anything except prevent existential risk—which, in a narrow interpretation, means it doesn’t stop half of humanity from murdering the other half—isn’t a future worth fighting for IMO. We can do so much better.
No one said you have to stop with that first FAI. You can try building another. The first FAI won’t oppose it (non-interference). Or, better yet, you can try talking to the other half of the humans.
There are people who believe religiously that End Times must come
Yes, but we assume they are factually wrong, and so their CEV would fix this.
A FAI team that precommitted to implementing CEV would definitely get the most funds. Even a team that precommitted to CEV might get more funds than CEV, because people like myself would reason that the team’s values are closer to my own than humanity’s average, plus they have a better chance of actually agreeing on more things.
Not bloody likely. I’m going to oppose your team, discourage your funders, and bomb your headquarters—because we have different moral opinions, right here, and if the differences turn out to be fundamental, and you build your FAI, then parts of my value will be forever unfulfilled.
You, on the other hand, may safely support my team, because you can be sure to like whatever my FAI will do, and regarding the rest, it won’t interfere.
No one said you have to stop with that first FAI. You can try building another. The first FAI won’t oppose it (non-interference).
No. Any FAI (ETA: or other AGI) has to be a singleton to last for long. Otherwise I can build a uFAI that might replace it.
Suppose your AI only does a few things that everyone agrees on, but otherwise “doesn’t interfere”. Then I can build another AI, which implements values people don’t agree on. Your AI must either interfere, or be resigned to not being very relevant in determining the future.
Will it only interfere if a consensus of humanity allows it to do so? Will it not stop a majority from murdering a minority? Then it’s at best a nice-to-have, but most likely useless. After people successfully build one AGI, they will quickly reuse the knowledge to build more. The first AGI that does not favor inaction will become a singleton, destroying the other AIs and preventing future new AIs, to safeguard its utility function. This is unavoidable. With truly powerful AGI, preventing new AIs from gaining power is the only stable solution.
Or, better yet, you can try talking to the other half of the humans.
Yeah, that’s worked really well for all of human history so far.
Yes, but we assume they are factually wrong, and so their CEV would fix this.
First, they may not factually wrong about the events they predict in the real world—like everyone dying—just wrong about the supernatural parts. (Especially if they’re themselves working to bring these events to pass.) IOW, this may not be a factual belief to be corrected, but a desired-by-them future that others like me and you would wish to prevent.
Second, you agreed the CEV of groups of people may contain very few things that they really agree on, so you can’t even assume they’ll have a nontrivial CEV at all, let alone that it will “fix” values you happen to disagree with.
Not bloody likely. I’m going to oppose your team, discourage your funders, and bomb your headquarters—because we have different moral opinions, right here, and if the differences turn out to be fundamental, and you build your FAI, then parts of my value will be forever unfulfilled.
You, on the other hand, may safely support my team, because you can be sure to like whatever my FAI will do, and regarding the rest, it won’t interfere.
I have no idea what your FAI will do, because even if you make no mistakes in building it, you yourself don’t know ahead of time what the CEV will work out to. If you did, you’d just plug those values into the AI directly instead of calculating the CEV. So I’ll want to bomb you anyway, if that increases my chances of being the first to build a FAI. Our morals are indeed different, and since there are no objectively distinguished morals, the difference goes both ways.
Of course I will dedicate my resources to first bombing people who are building even more inimical AIs. But if I somehow knew you and I were the only ones in the race, I’d politely ask you to join me or desist or be stopped by force.
As long as we’re discussing bombing, consider that the SIAI isn’t and won’t be in a position to bomb anyone. OTOH, if and when nation-states and militaries realize AGI is a real-world threat, they will go to war with each trying to prevent anyone else from building an AGI first. It’s the ultimate winner-take-all arms race.
This is going to happen, it might be happening already if enough politicians and generals had the beliefs of Eliezer about AGI, and it will happen (or not) regardless of anyone’s attempts to build any kind of Friendliness theory. Furthermore, a state military planning to build AGI singleton won’t stop to think for long before wiping your civilian, unprotected FAI theory research center off the map. Either you go underground or you cooperate with a powerful player (the state on whose territory you live, presumably). Or maybe states and militaries won’t wise up in time, and some private concern really will build the first AGI. Which may be better or worse depending on what they build.
Eventually, unless the whole world is bombed back into pre-computer-age tech, someone very probably will build an AGI of some kind. The SIAI idea is (possibly) to invent Friendliness theory and publish it widely, so that whoever builds that AGI, if they want it to be Friendly (at least to themselves!), they will have a relatively cheap and safe implementation to use. But for someone actually trying to build an AGI, two obvious rules are:
Absolute secrecy, or you get bombed right away.
Do absolutely whatever it takes to successfully launch as early as possible, and make your AI a singleton controlled by yourself or by nobody—regardless of your and the AI’s values.
Will it only interfere if a consensus of humanity allows it to do so? Will it not stop a majority from murdering a minority?
If the majority and the minority are so fundamentally different that their killing each other is not forbidden by the universal human CEV, then no. On what moral grounds would it do the prevention?
The first AGI that does not favor inaction will become a singleton, destroying the other AIs and preventing future new AIs
Until everybody agree that this new AGI is not good after all. Then the original AGI will interfere and dismantle the new one (the original is still the first and the strongest).
you can’t even assume they’ll have a nontrivial CEV at all, let alone that it will “fix” values you happen to disagree with.
But I can be sure that CEV fixes values that are based on false factual beliefs—this is a part of the definition of CEV.
I have no idea what your FAI will do
But you can be sure that it is something about which you (and everybody) would agree, either directly or if you were more intelligent and knew more.
there are no objectively distinguished morals
But there may be a partial ordering between morals, such that X<Y iff all “interfering” actions (whatever this means) that are allowed by X are also allowed by Y. Then if A1 and A2 are two agents, we may easily have:
[assuming Endorses(A, X) implies FAI does not perform any non-interfering action disagreeable for A]
if and when nation-states and militaries realize AGI is a real-world threat, they will go to war with each trying to prevent anyone else from building an AGI first. It’s the ultimate winner-take-all arms race. This is going to happen, it might be happening already if enough politicians and generals had the beliefs of Eliezer about AGI, and it will happen (or not) regardless of anyone’s attempts to build any kind of Friendliness theory.
Well, don’t you think this is just ridiculous? Does it look like the most rational behavior? Wouldn’t it be better for everybody to cooperate in this Prisoner’s Dilemma, and do it with a creditable precommitment?
If the majority and the minority are so fundamentally different that their killing each other is not forbidden by the universal human CEV, then no.
I don’t understand what you mean by “fundamentally different”. You said the AI would not do anything not backed by an all-human consensus. If a majority of humanity wishes to kill a minority, obviously there won’t be a consensus to stop the killing, and AI will not interfere. I prefer to live in a universe whose living AI does interfere in such a case.
On what moral grounds would it do the prevention?
Libertarianism is one moral principle that would argue for prevention. So would most varieties of utilitarianism (ignoring utility monsters and such). Again, I would prefer living with an AI hard-coded to one of those moral ideologies (though it’s not ideal) over your view of CEV.
Until everybody agree that this new AGI is not good after all. Then the original AGI will interfere and dismantle the new one (the original is still the first and the strongest).
Forever keeping this capability in reserve is most of what being a singleton means. But think of the practical implications: it has to be omnipresent, omniscient, and prevent other AIs from ever being as powerful as it is—which restricts those other AIs’ abilities in many endeavors. All the while it does little good itself. So from my point of view, the main effect of successfully implementing your view of CEV may be to drastically limit the opportunities for future AIs to do good.
And yet it doesn’t limit the opportunity to do evil, at least evil of the mundane death & torture kind. Unless you can explain why it would prevent even a very straightforward case like 80% of humanity voting to kill the other 20%.
But I can be sure that CEV fixes values that are based on false factual beliefs—this is a part of the definition of CEV.
But you said it would only do things that are approved by a strong human consensus. And I assure you that, to take an example, the large majority of the world’s population who today believe in the supernatural will not consent to having that belief “fixed”. Nor have you demonstrated that their extrapolated volition would want for them to be forcibly modified. Maybe their extrapolated volition simply doesn’t value objective truth highly (because they today don’t believe in the concept of objective truth, or believe that it contradicts everyday experience).
I have no idea what your FAI will do
But you can be sure that it is something about which you (and everybody) would agree, either directly or if you were more intelligent and knew more.
Yes, but I don’t know what I would approve of if I were “more intelligent” (a very ill defined term). And if you calculate that something, according to your definition of intelligence, and present me with the result, I might well reject that result even if I believe in your extrapolation process. I might well say: the future isn’t predetermined. You can’t calculate what I necessarily will become. You just extrapolated a creature I might become, which also happens to be more intelligent. But there’s nothing in my moral system that says I should adopt the values of someone else because they are more intelligent. If I don’t like the values I might say, thank-you for warning me, now I shall be doubly careful not to evolve into that kind of creature! I might even choose to forego the kind of increased intelligence that causes such an undsired change in my values.
Short version: “what I would want if I were more intelligent (according to some definition)” isn’t the same as “what I will likely want in the future”, because there’s no reason for me to grow in intelligence (by that definition) if I suspect it would twist my values. So you can’t apply the heuristic of “if I know what I’m going to think tomorrow, I might as well think it today”.
I think you may be missing a symbol there? If not, I can’t parse it… Can you spell out for me what it means to just write the last three Endorses(...) clauses one after the other?
Does it look like the most rational behavior?
It may be quite rational for everyone individually, depending on projected payoffs. Unlike a PD, starting positions aren’t symmetrical and players’ progress/payoffs are not visible to other players. So saying “just cooperate” doesn’t immediately apply.
Wouldn’t it be better for everybody to cooperate in this Prisoner’s Dilemma, and do it with a creditable precommitment?
How can a state or military precommit to not having a supersecret project to develop a private AGI?
And while it’s beneficial for some players to join in a cooperative effort, it may well be that a situation of several competing leagues (or really big players working alone) develops and is also stable. It’s all laid over the background of existing political, religious and personal enmities and rivalries—even before we come to actual disagreements over what the AI should value.
If a majority of humanity wishes to kill a minority, obviously there won’t be a consensus to stop the killing, and AI will not interfere.
This assumes that CEV uses something along the lines of a simulated vote as an aggregation mechanism. Currently the method of aggregation is undefined so we can’t say this with confidence—certainly not as something obvious.
I agree. However, if the CEV doesn’t privilege any value separately from how many people value it how much (in EV), and if the EV of a large majority values killing a small minority (whose EV is of course opposed), and if you have protection against both positive and negative utility monsters (so it’s at least not obvious and automatic that the negative value of the minority would outweigh the positive value of the majority) - then my scenario seems to me to be plausible, and an explanation is necessary as to how it might be prevented.
Of course you could say that until CEV is really formally specified, and we know how the aggregation works, this explanation cannot be produced.
If a majority of humanity wishes to kill a minority, obviously there won’t be a consensus to stop the killing, and AI will not interfere
The majority may wish to kill the minority for wrong reasons—based on false beliefs or insufficient intelligence. In which case their CEV-s won’t endorse it, and the FAI will interfere. “Fundamentally different” means their killing each other is endorsed by someone’s CEV, not just by themselves.
But you said it would only do things that are approved by a strong human consensus.
Strong consensus of their CEV-s.
Maybe their extrapolated volition simply doesn’t value objective truth highly (because they today don’t believe in the concept of objective truth, or believe that it contradicts everyday experience)
Extrapolated volition is based on objective truth, by definition.
If I don’t like the values I might say, thank-you for warning me, now I shall be doubly careful not to evolve into that kind of creature!
The process of extrapolation takes this into account.
I think you may be missing a symbol there? If not, I can’t parse it...
Sorry, bad formatting. I meant four independent clauses: each of the agents does not endorse CEV, but endorses CEV.
How can a state or military precommit to not having a supersecret project to develop a private AGI?
That’s a separate problem. I think it is easier to solve than extrapolating volition or building AI.
The majority may wish to kill the minority for wrong reasons—based on false beliefs or insufficient intelligence. In which case their CEV-s won’t endorse it, and the FAI will interfere
So you’re OK with the FAI not interfering if they want to kill them for the “right” reasons? Such as “if we kill them, we will benefit by dividing their resources among ourselves”?
But you said it would only do things that are approved by a strong human consensus.
Strong consensus of their CEV-s.
So you’re saying your version of CEV will forcibly update everyone’s beliefs and values to be “factual” and disallow people to believe in anything not supported by appropriate Bayesian evidence? Even if it has to modify those people by force, the result is unlike the original in many respects that they and many other people value and see as identity-forming, etc.? And it will do this not because it’s backed by a strong consensus of actual desires, but because post-modification there will be a strong consensus of people happy that the modification was made?
If your answer is “yes, it will do that”, then I would not call your AI a Friendly one at all.
Extrapolated volition is based on objective truth, by definition.
My understanding of the CEV doc differs from yours. It’s not a precise or complete spec, and it looks like both readings can be justified.
The doc doesn’t (on my reading) say that the extrapolated volition can totally conform to objective truth. The EV is based on an extrapolation of our existing volition, not of objective truth itself. One of the ways it extrapolates is by adding facts the original person was not aware of. But that doesn’t mean it removes all non-truth or all beliefs that “aren’t even wrong” from the original volition. If the original person effectively assigns 0 or 1 “non-updateable probability” to some belief, or honestly doesn’t believe in objective reality, or believes in “subjective truth” of some kind, CEV is not necessarily going to “cure” them of it—especially not by force.
But as long as we’re discussing your vision of CEV, I can only repeat what I said above—if it’s going to modify people by force like this, I think it’s unFriendly and if it were up to me, would not launch such an AI.
I meant four independent clauses: each of the agents does not endorse CEV, but endorses CEV.
Understood. But I don’t see how this partial ordering changes what I had described.
Let’s say I’m A1 and you’re A2. We would both prefer a mutual CEV than a CEV of the other only. But each of us would prefer even more a CEV of himself only. So each of us might try to bomb the other first if he expected to get away without retaliation. That there exists a possible compromise that is better than total defeat doesn’t mean total victory wouldn’t be much better than any compromise.
How can a state or military precommit to not having a supersecret project to develop a private AGI?
That’s a separate problem. I think it is easier to solve than extrapolating volition or building AI.
If you think so you must have evidence relating to how to actually solve this problem. Otherwise they’d both look equally mysterious. So, what’s your idea?
So you’re OK with the FAI not interfering if they want to kill them for the “right” reasons?
I wouldn’t like it. But if the alternative is, for example, to have FAI directly enforce the values of the minority on the majority (or vice versa) - the values that would make them kill in order to satisfy/prevent—then I prefer FAI not interfering.
“if we kill them, we will benefit by dividing their resources among ourselves”
If the resources are so scarce that dividing them is so important that even CEV-s agree on the necessity of killing, then again, I prefer humans to decide who gets them.
So you’re saying your version of CEV will forcibly update everyone’s beliefs
No. CEV does not updates anyone’s beliefs. It is calculated by extrapolating values in the presence of full knowledge and sufficient intelligence.
If the original person effectively assigns 0 or 1 “non-updateable probability” to some belief, or honestly doesn’t believe in objective reality, or believes in “subjective truth” of some kind, CEV is not necessarily going to “cure” them of it—especially not by force.
As I said elsewhere, if a person’s beliefs are THAT incompatible with truth, I’m ok with ignoring their volition. Note, that their CEV is undefined in this case. But I don’t believe there exist such people (excluding totally insane).
That there exists a possible compromise that is better than total defeat doesn’t mean total victory wouldn’t be much better than any compromise.
But the total loss would be correspondingly worse. PD reasoning says you should cooperate (assuming cooperation is precommittable).
If you think so you must have evidence relating to how to actually solve this problem. Otherwise they’d both look equally mysterious. So, what’s your idea?
Off the top of my head, adoption of total transparency for everybody of all governmental and military matters.
If the resources are so scarce that dividing them is so important that even CEV-s agree on the necessity of killing, then again, I prefer humans to decide who gets them.
The resources are not scarce at all. But, there’s no consensus of CEVs. The CEVs of 80% want to kill the rest. The CEVs of 20% obviously don’t want to be killed. Because there’s no consensus, your version of CEV would not interfere, and the 80% would be free to kill the 20%.
No. CEV does not updates anyone’s beliefs. It is calculated by extrapolating values in the presence of full knowledge and sufficient intelligence.
I meant that the AI that implements your version of CEV would forcibly update people’s actual beliefs to match what it CEV-extrapolated for them. Sorry for the confusion.
As I said elsewhere, if a person’s beliefs are THAT incompatible with truth, I’m ok with ignoring their volition. Note, that their CEV is undefined in this case. But I don’t believe there exist such people (excluding totally insane).
A case could be made that many millions of religious “true believers” have un-updatable 0⁄1 probabilities. And so on.
Your solution is to not give them a voice in the CEV at all. Which is great for the rest of us—our CEV will include some presumably reduced term for their welfare, but they don’t get to vote on things. This is something I would certainly support in a FAI (regardless of CEV), just as I would support using CEV or even CEV to CEV.
The only difference between us then is that I estimate there to be many such people. If you believed there were many such people, would you modify your solution, or is ignoring them however many they are fine by you?
PD reasoning says you should cooperate (assuming cooperation is precommittable).
As I said before, this reasoning is inapplicable, because this situation is nothing like a PD.
The PD reasoning to cooperate only applies in case of iterated PD, whereas creating a singleton AI is a single game.
Unlike PD, the payoffs are different between players, and players are not sure of each other’s payoffs in each scenario. (E.g., minor/weak players are more likely to cooperate than big ones that are more likely to succeed if they defect.)
The game is not instantaneous, so players can change their strategy based on how other players play. When they do so they can transfer value gained by themselves or by other players (e.g. join research alliance 1, learn its research secrets, then defect and sell the secrets to alliance 2).
It is possible to form alliances, which gain by “defecting” as a group. In PD, players cannot discuss alliances or trade other values to form them before choosing how to play.
There are other games going on between players, so they already have knowledge and opinions and prejudices about each other, and desires to cooperate with certain players and not others. Certain alliances will form naturally, others won’t.
adoption of total transparency for everybody of all governmental and military matters.
This counts as very weak evidence because it proves it’s at least possible to achieve this, yes. (If all players very intensively inspect all other players to make sure a secret project isn’t being hidden anywhere—they’d have to recruit a big chunk of the workforce just to watch over all the rest.)
But the probability of this happening in the real world, between all players, as they scramble to be the first to build an apocalyptic new weapon, is so small it’s not even worth discussion time. (Unless you disagree, of course.) I’m not convinced by this that it’s an easier problem to solve than that of building AGI or FAI or CEV.
The resources are not scarce at all. But, there’s no consensus of CEVs. The CEVs of 80% want to kill the rest.
The resources are not scarce, yet the CEV-s want to kill? Why?
I meant that the AI that implements your version of CEV would forcibly update people’s actual beliefs to match what it CEV-extrapolated for them.
It would do so only if everybody’s CEV-s agree that updating these people’s beliefs is a good thing.
If you believed there were many such people, would you modify your solution, or is ignoring them however many they are fine by you?
People that would still have false factual beliefs no matter how much evidence and how much intelligence they have? They would be incurably insane. Yes, I would agree to ignore their volition, no matter how many they are.
The PD reasoning to cooperate only applies in case of iterated PD
Err. What about arguments of Douglas Hofstadter and EY, and decision theories like TDT?
Unlike PD, the payoffs are different between players, and players are not sure of each other’s payoffs in each scenario
This doesn’t really matter for a broad range of possible payoff matrices.
join research alliance 1, learn its research secrets, then defect and sell the secrets to alliance 2
Cooperating in this game would mean there is exactly one global research alliance. A cooperating move is a precommitment to abide by its rules. Enforcing such precommitment is a separate problem. Let’s assume it’s solved.
I’m not convinced by this that it’s an easier problem to solve than that of building AGI or FAI or CEV.
Maybe you’re right. But IMHO it’s a less interesting problem :)
The resources are not scarce, yet the CEV-s want to kill? Why?
Sorry for the confusion. Let’s taboo “scarce” and start from scratch.
I’m talking about a scenario where—to simplify only slightly from the real world—there exist some finite (even if growing) resources such that almost everyone, no matter how much they already have, want more of. A coalition of 80% of the population forms, which would like to kill the other 20% in order to get their resources. Would the AI prevent this, althogh there is no consensus against the killing?
If you still want to ask whether the resource is “scarce”, please specify what that means exactly. Maybe any finite and highly desireable resource, with returns diminishing weakly or not at all, can be considered “scarce”.
It would do so only if everybody’s CEV-s agree that updating these people’s beliefs is a good thing.
People that would still have false factual beliefs no matter how much evidence and how much intelligence they have? They would be incurably insane. Yes, I would agree to ignore their volition, no matter how many they are.
As I said—this is fine by me insofar as I expect the CEV not to choose to ignore me. (Which means it’s not fine through the Rawlsian veil of ignorance, but I don’t care and presumably neither do you.)
The question of definition, who is to be included in the CEV? or—who is considered sane? becomes of paramount importance. Since it is not itself decided by the CEV, it is presumably hardcoded into the AI design (or evolves within that design as the AI self-modifies, but that’s very dangerous without formal proofs that it won’t evolve to include the “wrong” people.) The simplest way to hardcode it is to directly specify the people to be included, but you prefer testing on qualifications.
However this is realized, it would give people even more incentive to influence or stop your AI building process or to start their own to compete, since they would be afraid of not being included in the CEV used by your AI.
The PD reasoning to cooperate only applies in case of iterated PD
Err. What about arguments of Douglas Hofstadter and EY, and decision theories like TDT?
TDT applies where agents are “similar enough”. I doubt I am similar enough to e.g. the people you labelled insane.
Which arguments of Hofstadter and Yudkowsky do you mean?
Cooperating in this game would mean there is exactly one global research alliance.
Why? What prevents several competing alliances (or single players) from forming, competing for the cooperation of the smaller players?
A coalition of 80% of the population forms, which would like to kill the other 20% in order to get their resources
I have trouble thinking of a resource that would make even one person’s CEV, let alone 80%, want to kill people, in order to just have more of it.
The question of definition, who is to be included in the CEV? or—who is considered sane?
This is easy, and does not need any special hardcoding. If someone is so insane that their beliefs are totally closed and impossible to move by knowledge and intelligence, then their CEV is undefined. Thus, they are automatically excluded.
TDT applies where agents are “similar enough”. I doubt I am similar enough to e.g. the people you labelled insane.
We are talking about people building FAI-s. Surely they are intelligent enough to notice the symmetry between themselves. If you say that logic and rationality makes you decide to ‘defect’ (=try to build FAI on your own, bomb everyone else), then logic and rationality would make everyone decide to defect. So everybody bombs everybody else, no FAI gets built, everybody loses. Instead you can ‘cooperate’ (=precommit to build FAI<everybody’s CEV> and to bomb everyone that did not make the same precommitment). This gets us a single global alliance.
I have trouble thinking of a resource that would make even one person’s CEV, let alone 80%, want to kill people, in order to just have more of it.
shrug Space (land or whatever is being used). Mass and energy. Natural resources. Computing power. Finite-supply money and luxuries if such exist.
Or are you making an assumption that CEVs are automatically more altruistic or nice than non-extrapolated human volitions are?
This is easy, and does not need any special hardcoding. If someone is so insane that their beliefs are totally closed and impossible to move by knowledge and intelligence, then their CEV is undefined. Thus, they are automatically excluded.
Well it does need hardcoding: you need to tell the CEV to exclude people whose EVs are too similar to their current values despite learning contrary facts. Or even all those whose belief-updating process differs too much from perfect Bayesian (and how much is too much?) This is something you’d hardcode in, because you could also write (“hardcode”) a CEV that does include them, allowing them to keep the EVs close to their current values.
Not that I’m opposed to this decision (if you must have CEV at all).
We are talking about people building FAI-s. Surely they are intelligent enough to notice the symmetry between themselves.
There’s a symmetry, but “first person to complete AI wins, everyone ‘defects’” is also a symmetrical situation. Single-iteration PD is symmetrical, but everyone defects. Mere symmetry is not sufficient for TDT-style “decide for everyone”, you need similarity that includes similarly valuing the same outcomes. Here everyone values the outcome “have the AI obey ME!”, which is not the same.
If you say that logic and rationality makes you decide to ‘defect’ (=try to build FAI on your own, bomb everyone else), then logic and rationality would make everyone decide to defect. So everybody bombs everybody else, no FAI gets built, everybody loses.
Or someone is stronger than everyone else, wins the bombing contest, and builds the only AI. Or someone succeeds in building an AI in secret, avoiding being bombed. Or there’s a player or alliance that’s strong enough to deter bombing due to the threat of retaliation, and so completes their AI which doesn’t care about everyone else much. There are many possible and plausible outcomes besides “everybody loses”.
Instead you can ‘cooperate’ (=precommit to build FAI<everybody’s CEV> and to bomb everyone that did not make the same precommitment). This gets us a single global alliance.
Or while the alliance is still being built, a second alliance or very strong player bombs them to get the military advantages of a first strike. Again, there are other possible outcomes besides what you suggest.
Space (land or whatever is being used). Mass and energy. Natural resources. Computing power. Finite-supply money and luxuries if such exist. Or are you making an assumption that CEVs are automatically more altruistic or nice than non-extrapolated human volitions are?
These all have property that you only need so much of them. If there is a sufficient amount for everybody, then there is no point in killing in order to get more. I expect CEV-s to not be greedy just for the sake of greed. It’s people’s CEV-s we’re talking about, not paperclip maximizers’.
Well it does need hardcoding: you need to tell the CEV to exclude people whose EVs are too similar to their current values despite learning contrary facts. Or even all those whose belief-updating process differs too much from perfect Bayesian (and how much is too much?) This is something you’d hardcode in, because you could also write (“hardcode”) a CEV that does include them, allowing them to keep the EVs close to their current values.
Hmm, we are starting to argue about exact details of extrapolation process...
There are many possible and plausible outcomes besides “everybody loses”.
Lets formalize the problem. Let F(R, Ropp) be the probability of a team successfully building a FAI first, given R resources, and having opposition with Ropp resources. Let Uself, Ueverybody, and Uother be the rewards for being first in building FAI, FAI, and FAI, respectively. Naturally, F is monotonically increasing in R and decreasing in Ropp, and Uother < Ueverybody < Uself.
Assume there are just two teams, with resources R1 and R2, and each can perform one of two actions: “cooperate” or “defect”. Let’s compute the expected utilities for the first team:
We cooperate, opponent team cooperates:
EU("CC") = Ueverybody * F(R1+R2, 0)
We cooperate, opponent team defects:
EU("CD") = Ueverybody * F(R1, R2) + Uother * F(R2, R1)
We defect, opponent team cooperates:
EU("DC") = Uself * F(R1, R2) + Ueverybody * F(R2, R1)
We defect, opponent team defects:
EU("DD") = Uself * F(R1, R2) + Uother * F(R2, R1)
Then, EU(“CD”) < EU(“DD”) < EU(“DC”), which gives us most of the structure of a PD problem. The rest, however, depends on the finer details. Let A = F(R1,R2)/F(R1+R2,0) and B = F(R2,R1)/F(R1+R2,0). Then:
If Ueverybody ⇐ Uself*A + Uother*B, then EU(“CC”) < EU(“DD”), and there is no point in cooperating. This is your position: Ueverybody is much less than Uself, or Uother is not much less than Ueverybody, and/or your team has so much more resources than the other.
If Uself*A + Uother*B < Ueverybody < Uself*A/(1-B), this is a true Prisoner’s dilemma.
If Ueverybody >= Uself*A/(1-B), then EU(“CC”) >= EU(“DC”), and “cooperate” is the obviously correct decision. This is my position: Ueverybody is not much less than Uself, and/or the teams are more evenly matched.
These all have property that you only need so much of them.
All of those resources are fungible and can be exchanged for time. There might be no limit to the amount of time people desire, even very enlightened posthuman people.
I don’t think you can get an everywhere-positive exchange rate. There are diminishing returns and a threshold, after which, exchanging more resources won’t get you any more time. There’s only 30 hours in a day, after all :)
You can use some resources like computation directly and in unlimited amounts (e.g. living for unlimitedly long virtual times per real second inside a simulation). There are some physical limits on that due to speed of light limiting effective brain size, but that depends on brain design and anyway the limits seem to be pretty high.
More generally: number of configurations physically possible in a given volume of space is limited (by the entropy of a black hole). If you have a utility function unbounded from above, as it rises it must eventually map to states that describe more space or matter than the amount you started with. Any utility maximizer with unbounded utility eventually wants to expand.
I don’t know what the exchange rates are, but it does cost something (computer time, energy, negentropy) to stay alive. That holds for simulated creatures too. If the available resources to keep someone alive are limited, then I think there will be conflict over those resources.
Naturally, F is monotonically increasing in R and decreasing in Ropp
You’re treating resources as one single kind, where really there are many kinds with possible trades between teams. Here you’re ignoring a factor that might actually be crucial to encouraging cooperation (I’m not saying I showed this formally :-)
Assume there are just two teams
But my point was exactly that there would be many teams who could form many different alliances. Assuming only two is unrealistic and just ignores what I was saying. I don’t even care much if given two teams the correct choice is to cooperate, because I set very low probability on there being exactly two teams and no other independent players being able to contribute anything (money, people, etc) to one of the teams.
This is my position
You still haven’t given good evidence for holding this position regarding the relation between the different Uxxx utilities. Except for the fact CEV is not really specified, so it could be built so that that would be true. But equally it could be built so that that would be false. There’s no point in arguing over which possibility “CEV” really refers to (although if everyone agreed on something that would clear up a lot of debates); the important questions are what do we want a FAI to do if we build one, and what we anticipate others to tell their FAIs to do.
You’re treating resources as one single kind, where really there are many kinds with possible trades between teams
I think this is reasonably realistic. Let R signify money. Then R can buy other necessary resources.
But my point was exactly that there would be many teams who could form many different alliances. Assuming only two is unrealistic and just ignores what I was saying.
We can model N teams by letting them play two-player games in succession. For example, any two teams with nearly matched resources would cooperate with each other, producing a single combined team, etc… This may be an interesting problem to solve, analytically or by computer modeling.
You still haven’t given good evidence for holding this position regarding the relation between the different Uxxx utilities.
You’re right. Initially, I thought that the actual values of Uxxx-s will not be important for the decision, as long as their relative preference order is as stated. But this turned out to be incorrect. There are regions of cooperation and defection.
Analytically, I don’t a priori expect a succession of two-player games to have the same result as one many-player game which also has duration in time and not just a single round.
Because there’s no consensus, your version of CEV would not interfere, and the 80% would be free to kill the 20%.
There may be a distinction between “the AI will not prevent the 80% from killing the 20%” and “nothing will prevent the 80% from killing the 20%” that is getting lost in your phrasing. I am not convinced that the math doesn’t make them equivalent, in the long run—but I’m definitely not convinced otherwise.
I’m assuming the 80% are capable of killing the 20% unless the AI interferes. That’s part of the thought experiment. It’s not unreasonable, since they are 4 times as numerous. But if you find this problematic, suppose it’s 99% killing 1% at a time. It doesn’t really matter.
My point is that we currently have methods of preventing this that don’t require an AI, and which do pretty well. Why do we need the AI to do it? Or more specifically, why should we reject an AI that won’t, but may do other useful things?
There have been, and are, many mass killings of minority groups and of enemy populations and conscripted soldiers at war. If we cure death and diseases, this will become the biggest cause of death and suffering in the world. It’s important and we’ll have to deal with it eventually.
The AI under discussion not just won’t solve the problem, it would (I contend) become a singleton and prevent me from building another AI that does solve the problem. (If it chooses not to become a singleton, it will quickly be supplanted by an AI that does try to become one.)
If the original person effectively assigns 0 or 1 “non-updateable probability” to some belief, or honestly doesn’t believe in objective reality, or believes in “subjective truth” of some kind, CEV is not necessarily going to “cure” them of it—especially not by force.
I think you’re skipping between levels hereabouts. CEV, the theoretical construct, might consider people so modified, even if a FAI based on CEV would not modify them. CEV is our values if we were better, but does not necessitate us actually getting better.
In this thread I always used CEV in the sense of an AI implementing CEV. (Sometimes you’ll see descriptions of what I don’t believe to be the standard interpretation of how such an AI would behave, where gRR suggests such behaviors and I reply.)
I’m still skeptical of this. If you think of FAI as simply AI that is “safe”—one that does not automatically kill us all (or other massive disutility), relative to the status quo—then plenty of non-singletons are FAI.
Of course, by that definition the ‘F’ looks like the easy part. Rocks are Friendly.
I didn’t mean that being a singleton is a precondition to FAI-hood. I meant that any AGI, friendly or not, that doesn’t prevent another AGI from rising will have to fight all the time, for its life and for the complete fulfillment of its utility function, and eventually it will lose; and a singleton is the obvious stable solution. Edited to clarify.
Are you suggesting that an AGI that values anything at all is incapable of valuing the existence of other AGIs, or merely that this is sufficiently unlikely as to not be worth considering?
It can certainly value them, and create them, cooperate and trade, etc. etc. There are two exceptions that make such valuing and cooperation take second place.
First: an uFAI is just as unfriendly and scary to other AIs as to humans. An AI will therefore try to prevent other AIs from achieving dangerous power unless it is very sure of their current and future goals.
Second: an AI created by humans (plus or minus self-modifications) with an explicit value/goal system of the form “the universe should be THIS way”, will try to stop any and all agents that try to interfere with shaping the universe as it wishes. And the foremost danger in this category is—other AIs created in the same way but with different goals.
I’m a little confused by your response, and I suspect that I was unclear in my question.
I agree that an AI with an explicit value/goal system of the form “the universe should be THIS way”, will try to stop any and all agents that try to interfere with shaping the universe as it wishes (either by destroying them, or altering their goal structure, or securing their reliable cooperation, or something else).
But for an AI with the value “the universe should contain as many distinct intelligences as possible,” valuing and creating other AIs will presumably take first place.
But for an AI with the value “the universe should contain as many distinct intelligences as possible,” valuing and creating other AIs will presumably take first place.
That’s probably more efficiently done by destroying any other AIs that come along, while tiling the universe with slightly varying low-level intelligences.
I no longer know what the words “intelligence,” “AI”, and “AGI” actually refer to in this conversation, and I’m not even certain the referents are consistent, so let me taboo the whole lexical mess and try again.
For any X, if the existence of X interferes with an agent A achieving its goals, the better A is at optimizing its environment for its goals the less likely X is to exist.
For any X and A, the more optimizing power X can exert on its environment, the more likely it is that the existence of X interferes with A achieving its goals.
For any X, if A values the existence of X, the better A is at implementing its values the more likely X is to exist.
All of this is as true for X=intelligent beings as X=AI as X=AGI as X=pie.
Cool. So it seems to follow that we agree that if agent A1 values the existence of distinct agents A2..An, it’s unclear how the likelihood of A2..An existing varies with the optimizing power available to A1...An. Yes?
Yes. Even if we know each agent’s optimizing power, and each agent’s estimation of each other agent’s power and ability to acquire greater power, the behavior of A1 still depends on its exact values (for instance, what else it values besides the existence of the others). It also depends on the values of the other agents (might they choose to initiate conflict among themselves or against A1?)
I tend to agree. Unless it has specific values to the contrary, other AIs of power comparable to your own (or which might grow into such power one day) are too dangerous to leave running around. If you value states of the external universe, and you happen to be the first powerful AGI built, it’s natural to try to become a singleton as a preventative measure.
It’s certainly possible. My analysis so far is only on a “all else being equal” footing.
I do feel that, absent other data, the safer assumption is that if an AI is capable of becoming a singleton at all, expense (in terms of energy/matter and space or time) isn’t going to be the thing that stops it. But that may be just a cached thought because I’m used to thinking of an AI trying to become a singleton as a dangerous potential adversary. I would appreciate your insight.
As for values, certainly conflicting values can exist, from ones that mention the subject directly (“don’t move everyone to a simulation in a way they don’t notice” would close one obvious route) to ones that impinge upon it in unexpected ways (“no first strike against aliens” becomes “oops, an alien-built paperclipper just ate Jupiter from the inside out”).
I want to point out that all of my objections are acknowledged (not dismissed, and not fully resolved) in the actual CEV document—which is very likely hopelessly outdated by now to Eliezer and the SIAI, but they deliberately don’t publish anything newer (and I can guess at some of the reasons).
Which is why when I see people advocating CEV without understanding the dangers, I try to correct them.
I would be fine with FAI removing existential risks and not doing any other thing until everybody(’s CEV) agrees on it. (I assume here that removing existential risks is one such thing.) And an FAI team that creditably precommitted to implementing CEV instead of CEV would probably get more resources and would finish first.
So what makes you think everybody’s CEV would eventually agree on anything more?
A FAI that never does anything except prevent existential risk—which, in a narrow interpretation, means it doesn’t stop half of humanity from murdering the other half—isn’t a future worth fighting for IMO. We can do so much better. (At least, we can if we’re speculating about building a FAI to execute any well-defined plan we can come up with.)
I’m not even sure of that. There are people who believe religiously that End Times must come when everyone must die, and some of them want to hurry that along by actually killing people. And the meaning of “existential risk” is up for grabs anyway—does it preclude evolution into non-humans, leaving no members of original human species in existence? Does it preclude the death of everyone alive today, if some humans are always alive?
Sure, it’s unlikely or it might look like a contrived example to you. But are you really willing to precommit the future light cone, the single shot at creating an FAI (singleton), to whatever CEV might happen to be, without actually knowing what CEV produces and having an abort switch? That’s one of the defining points of CEV: that you can’t know it correctly in advance, or you would just program it directly as a set of goals instead of building a CEV-calculating machine.
This seems wrong. A FAI team that precommitted to implementing CEV would definitely get the most funds. Even a team that precommitted to CEV might get more funds than CEV, because people like myself would reason that the team’s values are closer to my own than humanity’s average, plus they have a better chance of actually agreeing on more things.
No one said you have to stop with that first FAI. You can try building another. The first FAI won’t oppose it (non-interference). Or, better yet, you can try talking to the other half of the humans.
Yes, but we assume they are factually wrong, and so their CEV would fix this.
Not bloody likely. I’m going to oppose your team, discourage your funders, and bomb your headquarters—because we have different moral opinions, right here, and if the differences turn out to be fundamental, and you build your FAI, then parts of my value will be forever unfulfilled.
You, on the other hand, may safely support my team, because you can be sure to like whatever my FAI will do, and regarding the rest, it won’t interfere.
No. Any FAI (ETA: or other AGI) has to be a singleton to last for long. Otherwise I can build a uFAI that might replace it.
Suppose your AI only does a few things that everyone agrees on, but otherwise “doesn’t interfere”. Then I can build another AI, which implements values people don’t agree on. Your AI must either interfere, or be resigned to not being very relevant in determining the future.
Will it only interfere if a consensus of humanity allows it to do so? Will it not stop a majority from murdering a minority? Then it’s at best a nice-to-have, but most likely useless. After people successfully build one AGI, they will quickly reuse the knowledge to build more. The first AGI that does not favor inaction will become a singleton, destroying the other AIs and preventing future new AIs, to safeguard its utility function. This is unavoidable. With truly powerful AGI, preventing new AIs from gaining power is the only stable solution.
Yeah, that’s worked really well for all of human history so far.
First, they may not factually wrong about the events they predict in the real world—like everyone dying—just wrong about the supernatural parts. (Especially if they’re themselves working to bring these events to pass.) IOW, this may not be a factual belief to be corrected, but a desired-by-them future that others like me and you would wish to prevent.
Second, you agreed the CEV of groups of people may contain very few things that they really agree on, so you can’t even assume they’ll have a nontrivial CEV at all, let alone that it will “fix” values you happen to disagree with.
I have no idea what your FAI will do, because even if you make no mistakes in building it, you yourself don’t know ahead of time what the CEV will work out to. If you did, you’d just plug those values into the AI directly instead of calculating the CEV. So I’ll want to bomb you anyway, if that increases my chances of being the first to build a FAI. Our morals are indeed different, and since there are no objectively distinguished morals, the difference goes both ways.
Of course I will dedicate my resources to first bombing people who are building even more inimical AIs. But if I somehow knew you and I were the only ones in the race, I’d politely ask you to join me or desist or be stopped by force.
As long as we’re discussing bombing, consider that the SIAI isn’t and won’t be in a position to bomb anyone. OTOH, if and when nation-states and militaries realize AGI is a real-world threat, they will go to war with each trying to prevent anyone else from building an AGI first. It’s the ultimate winner-take-all arms race.
This is going to happen, it might be happening already if enough politicians and generals had the beliefs of Eliezer about AGI, and it will happen (or not) regardless of anyone’s attempts to build any kind of Friendliness theory. Furthermore, a state military planning to build AGI singleton won’t stop to think for long before wiping your civilian, unprotected FAI theory research center off the map. Either you go underground or you cooperate with a powerful player (the state on whose territory you live, presumably). Or maybe states and militaries won’t wise up in time, and some private concern really will build the first AGI. Which may be better or worse depending on what they build.
Eventually, unless the whole world is bombed back into pre-computer-age tech, someone very probably will build an AGI of some kind. The SIAI idea is (possibly) to invent Friendliness theory and publish it widely, so that whoever builds that AGI, if they want it to be Friendly (at least to themselves!), they will have a relatively cheap and safe implementation to use. But for someone actually trying to build an AGI, two obvious rules are:
Absolute secrecy, or you get bombed right away.
Do absolutely whatever it takes to successfully launch as early as possible, and make your AI a singleton controlled by yourself or by nobody—regardless of your and the AI’s values.
If the majority and the minority are so fundamentally different that their killing each other is not forbidden by the universal human CEV, then no. On what moral grounds would it do the prevention?
Until everybody agree that this new AGI is not good after all. Then the original AGI will interfere and dismantle the new one (the original is still the first and the strongest).
But I can be sure that CEV fixes values that are based on false factual beliefs—this is a part of the definition of CEV.
But you can be sure that it is something about which you (and everybody) would agree, either directly or if you were more intelligent and knew more.
But there may be a partial ordering between morals, such that X<Y iff all “interfering” actions (whatever this means) that are allowed by X are also allowed by Y. Then if A1 and A2 are two agents, we may easily have:
~Endorses(A1, CEV) ~Endorses(A2, CEV) Endorses(A1, CEV)
Endorses(A2, CEV)
[assuming Endorses(A, X) implies FAI does not perform any non-interfering action disagreeable for A]
Well, don’t you think this is just ridiculous? Does it look like the most rational behavior? Wouldn’t it be better for everybody to cooperate in this Prisoner’s Dilemma, and do it with a creditable precommitment?
I don’t understand what you mean by “fundamentally different”. You said the AI would not do anything not backed by an all-human consensus. If a majority of humanity wishes to kill a minority, obviously there won’t be a consensus to stop the killing, and AI will not interfere. I prefer to live in a universe whose living AI does interfere in such a case.
Libertarianism is one moral principle that would argue for prevention. So would most varieties of utilitarianism (ignoring utility monsters and such). Again, I would prefer living with an AI hard-coded to one of those moral ideologies (though it’s not ideal) over your view of CEV.
Forever keeping this capability in reserve is most of what being a singleton means. But think of the practical implications: it has to be omnipresent, omniscient, and prevent other AIs from ever being as powerful as it is—which restricts those other AIs’ abilities in many endeavors. All the while it does little good itself. So from my point of view, the main effect of successfully implementing your view of CEV may be to drastically limit the opportunities for future AIs to do good.
And yet it doesn’t limit the opportunity to do evil, at least evil of the mundane death & torture kind. Unless you can explain why it would prevent even a very straightforward case like 80% of humanity voting to kill the other 20%.
But you said it would only do things that are approved by a strong human consensus. And I assure you that, to take an example, the large majority of the world’s population who today believe in the supernatural will not consent to having that belief “fixed”. Nor have you demonstrated that their extrapolated volition would want for them to be forcibly modified. Maybe their extrapolated volition simply doesn’t value objective truth highly (because they today don’t believe in the concept of objective truth, or believe that it contradicts everyday experience).
Yes, but I don’t know what I would approve of if I were “more intelligent” (a very ill defined term). And if you calculate that something, according to your definition of intelligence, and present me with the result, I might well reject that result even if I believe in your extrapolation process. I might well say: the future isn’t predetermined. You can’t calculate what I necessarily will become. You just extrapolated a creature I might become, which also happens to be more intelligent. But there’s nothing in my moral system that says I should adopt the values of someone else because they are more intelligent. If I don’t like the values I might say, thank-you for warning me, now I shall be doubly careful not to evolve into that kind of creature! I might even choose to forego the kind of increased intelligence that causes such an undsired change in my values.
Short version: “what I would want if I were more intelligent (according to some definition)” isn’t the same as “what I will likely want in the future”, because there’s no reason for me to grow in intelligence (by that definition) if I suspect it would twist my values. So you can’t apply the heuristic of “if I know what I’m going to think tomorrow, I might as well think it today”.
I think you may be missing a symbol there? If not, I can’t parse it… Can you spell out for me what it means to just write the last three Endorses(...) clauses one after the other?
It may be quite rational for everyone individually, depending on projected payoffs. Unlike a PD, starting positions aren’t symmetrical and players’ progress/payoffs are not visible to other players. So saying “just cooperate” doesn’t immediately apply.
How can a state or military precommit to not having a supersecret project to develop a private AGI?
And while it’s beneficial for some players to join in a cooperative effort, it may well be that a situation of several competing leagues (or really big players working alone) develops and is also stable. It’s all laid over the background of existing political, religious and personal enmities and rivalries—even before we come to actual disagreements over what the AI should value.
This assumes that CEV uses something along the lines of a simulated vote as an aggregation mechanism. Currently the method of aggregation is undefined so we can’t say this with confidence—certainly not as something obvious.
I agree. However, if the CEV doesn’t privilege any value separately from how many people value it how much (in EV), and if the EV of a large majority values killing a small minority (whose EV is of course opposed), and if you have protection against both positive and negative utility monsters (so it’s at least not obvious and automatic that the negative value of the minority would outweigh the positive value of the majority) - then my scenario seems to me to be plausible, and an explanation is necessary as to how it might be prevented.
Of course you could say that until CEV is really formally specified, and we know how the aggregation works, this explanation cannot be produced.
Absolutely, on both counts.
The majority may wish to kill the minority for wrong reasons—based on false beliefs or insufficient intelligence. In which case their CEV-s won’t endorse it, and the FAI will interfere. “Fundamentally different” means their killing each other is endorsed by someone’s CEV, not just by themselves.
Strong consensus of their CEV-s.
Extrapolated volition is based on objective truth, by definition.
The process of extrapolation takes this into account.
Sorry, bad formatting. I meant four independent clauses: each of the agents does not endorse CEV, but endorses CEV.
That’s a separate problem. I think it is easier to solve than extrapolating volition or building AI.
So you’re OK with the FAI not interfering if they want to kill them for the “right” reasons? Such as “if we kill them, we will benefit by dividing their resources among ourselves”?
So you’re saying your version of CEV will forcibly update everyone’s beliefs and values to be “factual” and disallow people to believe in anything not supported by appropriate Bayesian evidence? Even if it has to modify those people by force, the result is unlike the original in many respects that they and many other people value and see as identity-forming, etc.? And it will do this not because it’s backed by a strong consensus of actual desires, but because post-modification there will be a strong consensus of people happy that the modification was made?
If your answer is “yes, it will do that”, then I would not call your AI a Friendly one at all.
My understanding of the CEV doc differs from yours. It’s not a precise or complete spec, and it looks like both readings can be justified.
The doc doesn’t (on my reading) say that the extrapolated volition can totally conform to objective truth. The EV is based on an extrapolation of our existing volition, not of objective truth itself. One of the ways it extrapolates is by adding facts the original person was not aware of. But that doesn’t mean it removes all non-truth or all beliefs that “aren’t even wrong” from the original volition. If the original person effectively assigns 0 or 1 “non-updateable probability” to some belief, or honestly doesn’t believe in objective reality, or believes in “subjective truth” of some kind, CEV is not necessarily going to “cure” them of it—especially not by force.
But as long as we’re discussing your vision of CEV, I can only repeat what I said above—if it’s going to modify people by force like this, I think it’s unFriendly and if it were up to me, would not launch such an AI.
Understood. But I don’t see how this partial ordering changes what I had described.
Let’s say I’m A1 and you’re A2. We would both prefer a mutual CEV than a CEV of the other only. But each of us would prefer even more a CEV of himself only. So each of us might try to bomb the other first if he expected to get away without retaliation. That there exists a possible compromise that is better than total defeat doesn’t mean total victory wouldn’t be much better than any compromise.
If you think so you must have evidence relating to how to actually solve this problem. Otherwise they’d both look equally mysterious. So, what’s your idea?
I wouldn’t like it. But if the alternative is, for example, to have FAI directly enforce the values of the minority on the majority (or vice versa) - the values that would make them kill in order to satisfy/prevent—then I prefer FAI not interfering.
If the resources are so scarce that dividing them is so important that even CEV-s agree on the necessity of killing, then again, I prefer humans to decide who gets them.
No. CEV does not updates anyone’s beliefs. It is calculated by extrapolating values in the presence of full knowledge and sufficient intelligence.
As I said elsewhere, if a person’s beliefs are THAT incompatible with truth, I’m ok with ignoring their volition. Note, that their CEV is undefined in this case. But I don’t believe there exist such people (excluding totally insane).
But the total loss would be correspondingly worse. PD reasoning says you should cooperate (assuming cooperation is precommittable).
Off the top of my head, adoption of total transparency for everybody of all governmental and military matters.
The resources are not scarce at all. But, there’s no consensus of CEVs. The CEVs of 80% want to kill the rest. The CEVs of 20% obviously don’t want to be killed. Because there’s no consensus, your version of CEV would not interfere, and the 80% would be free to kill the 20%.
I meant that the AI that implements your version of CEV would forcibly update people’s actual beliefs to match what it CEV-extrapolated for them. Sorry for the confusion.
A case could be made that many millions of religious “true believers” have un-updatable 0⁄1 probabilities. And so on.
Your solution is to not give them a voice in the CEV at all. Which is great for the rest of us—our CEV will include some presumably reduced term for their welfare, but they don’t get to vote on things. This is something I would certainly support in a FAI (regardless of CEV), just as I would support using CEV or even CEV to CEV.
The only difference between us then is that I estimate there to be many such people. If you believed there were many such people, would you modify your solution, or is ignoring them however many they are fine by you?
As I said before, this reasoning is inapplicable, because this situation is nothing like a PD.
The PD reasoning to cooperate only applies in case of iterated PD, whereas creating a singleton AI is a single game.
Unlike PD, the payoffs are different between players, and players are not sure of each other’s payoffs in each scenario. (E.g., minor/weak players are more likely to cooperate than big ones that are more likely to succeed if they defect.)
The game is not instantaneous, so players can change their strategy based on how other players play. When they do so they can transfer value gained by themselves or by other players (e.g. join research alliance 1, learn its research secrets, then defect and sell the secrets to alliance 2).
It is possible to form alliances, which gain by “defecting” as a group. In PD, players cannot discuss alliances or trade other values to form them before choosing how to play.
There are other games going on between players, so they already have knowledge and opinions and prejudices about each other, and desires to cooperate with certain players and not others. Certain alliances will form naturally, others won’t.
This counts as very weak evidence because it proves it’s at least possible to achieve this, yes. (If all players very intensively inspect all other players to make sure a secret project isn’t being hidden anywhere—they’d have to recruit a big chunk of the workforce just to watch over all the rest.)
But the probability of this happening in the real world, between all players, as they scramble to be the first to build an apocalyptic new weapon, is so small it’s not even worth discussion time. (Unless you disagree, of course.) I’m not convinced by this that it’s an easier problem to solve than that of building AGI or FAI or CEV.
The resources are not scarce, yet the CEV-s want to kill? Why?
It would do so only if everybody’s CEV-s agree that updating these people’s beliefs is a good thing.
People that would still have false factual beliefs no matter how much evidence and how much intelligence they have? They would be incurably insane. Yes, I would agree to ignore their volition, no matter how many they are.
Err. What about arguments of Douglas Hofstadter and EY, and decision theories like TDT?
This doesn’t really matter for a broad range of possible payoff matrices.
Cooperating in this game would mean there is exactly one global research alliance. A cooperating move is a precommitment to abide by its rules. Enforcing such precommitment is a separate problem. Let’s assume it’s solved.
Maybe you’re right. But IMHO it’s a less interesting problem :)
Sorry for the confusion. Let’s taboo “scarce” and start from scratch.
I’m talking about a scenario where—to simplify only slightly from the real world—there exist some finite (even if growing) resources such that almost everyone, no matter how much they already have, want more of. A coalition of 80% of the population forms, which would like to kill the other 20% in order to get their resources. Would the AI prevent this, althogh there is no consensus against the killing?
If you still want to ask whether the resource is “scarce”, please specify what that means exactly. Maybe any finite and highly desireable resource, with returns diminishing weakly or not at all, can be considered “scarce”.
As I said—this is fine by me insofar as I expect the CEV not to choose to ignore me. (Which means it’s not fine through the Rawlsian veil of ignorance, but I don’t care and presumably neither do you.)
The question of definition, who is to be included in the CEV? or—who is considered sane? becomes of paramount importance. Since it is not itself decided by the CEV, it is presumably hardcoded into the AI design (or evolves within that design as the AI self-modifies, but that’s very dangerous without formal proofs that it won’t evolve to include the “wrong” people.) The simplest way to hardcode it is to directly specify the people to be included, but you prefer testing on qualifications.
However this is realized, it would give people even more incentive to influence or stop your AI building process or to start their own to compete, since they would be afraid of not being included in the CEV used by your AI.
TDT applies where agents are “similar enough”. I doubt I am similar enough to e.g. the people you labelled insane.
Which arguments of Hofstadter and Yudkowsky do you mean?
Why? What prevents several competing alliances (or single players) from forming, competing for the cooperation of the smaller players?
I have trouble thinking of a resource that would make even one person’s CEV, let alone 80%, want to kill people, in order to just have more of it.
This is easy, and does not need any special hardcoding. If someone is so insane that their beliefs are totally closed and impossible to move by knowledge and intelligence, then their CEV is undefined. Thus, they are automatically excluded.
We are talking about people building FAI-s. Surely they are intelligent enough to notice the symmetry between themselves. If you say that logic and rationality makes you decide to ‘defect’ (=try to build FAI on your own, bomb everyone else), then logic and rationality would make everyone decide to defect. So everybody bombs everybody else, no FAI gets built, everybody loses. Instead you can ‘cooperate’ (=precommit to build FAI<everybody’s CEV> and to bomb everyone that did not make the same precommitment). This gets us a single global alliance.
shrug Space (land or whatever is being used). Mass and energy. Natural resources. Computing power. Finite-supply money and luxuries if such exist.
Or are you making an assumption that CEVs are automatically more altruistic or nice than non-extrapolated human volitions are?
Well it does need hardcoding: you need to tell the CEV to exclude people whose EVs are too similar to their current values despite learning contrary facts. Or even all those whose belief-updating process differs too much from perfect Bayesian (and how much is too much?) This is something you’d hardcode in, because you could also write (“hardcode”) a CEV that does include them, allowing them to keep the EVs close to their current values.
Not that I’m opposed to this decision (if you must have CEV at all).
There’s a symmetry, but “first person to complete AI wins, everyone ‘defects’” is also a symmetrical situation. Single-iteration PD is symmetrical, but everyone defects. Mere symmetry is not sufficient for TDT-style “decide for everyone”, you need similarity that includes similarly valuing the same outcomes. Here everyone values the outcome “have the AI obey ME!”, which is not the same.
Or someone is stronger than everyone else, wins the bombing contest, and builds the only AI. Or someone succeeds in building an AI in secret, avoiding being bombed. Or there’s a player or alliance that’s strong enough to deter bombing due to the threat of retaliation, and so completes their AI which doesn’t care about everyone else much. There are many possible and plausible outcomes besides “everybody loses”.
Or while the alliance is still being built, a second alliance or very strong player bombs them to get the military advantages of a first strike. Again, there are other possible outcomes besides what you suggest.
These all have property that you only need so much of them. If there is a sufficient amount for everybody, then there is no point in killing in order to get more. I expect CEV-s to not be greedy just for the sake of greed. It’s people’s CEV-s we’re talking about, not paperclip maximizers’.
Hmm, we are starting to argue about exact details of extrapolation process...
Lets formalize the problem. Let F(R, Ropp) be the probability of a team successfully building a FAI first, given R resources, and having opposition with Ropp resources. Let Uself, Ueverybody, and Uother be the rewards for being first in building FAI, FAI, and FAI, respectively. Naturally, F is monotonically increasing in R and decreasing in Ropp, and Uother < Ueverybody < Uself.
Assume there are just two teams, with resources R1 and R2, and each can perform one of two actions: “cooperate” or “defect”. Let’s compute the expected utilities for the first team:
Then, EU(“CD”) < EU(“DD”) < EU(“DC”), which gives us most of the structure of a PD problem. The rest, however, depends on the finer details. Let A = F(R1,R2)/F(R1+R2,0) and B = F(R2,R1)/F(R1+R2,0). Then:
If Ueverybody ⇐ Uself*A + Uother*B, then EU(“CC”) < EU(“DD”), and there is no point in cooperating. This is your position: Ueverybody is much less than Uself, or Uother is not much less than Ueverybody, and/or your team has so much more resources than the other.
If Uself*A + Uother*B < Ueverybody < Uself*A/(1-B), this is a true Prisoner’s dilemma.
If Ueverybody >= Uself*A/(1-B), then EU(“CC”) >= EU(“DC”), and “cooperate” is the obviously correct decision. This is my position: Ueverybody is not much less than Uself, and/or the teams are more evenly matched.
All of those resources are fungible and can be exchanged for time. There might be no limit to the amount of time people desire, even very enlightened posthuman people.
I don’t think you can get an everywhere-positive exchange rate. There are diminishing returns and a threshold, after which, exchanging more resources won’t get you any more time. There’s only 30 hours in a day, after all :)
You can use some resources like computation directly and in unlimited amounts (e.g. living for unlimitedly long virtual times per real second inside a simulation). There are some physical limits on that due to speed of light limiting effective brain size, but that depends on brain design and anyway the limits seem to be pretty high.
More generally: number of configurations physically possible in a given volume of space is limited (by the entropy of a black hole). If you have a utility function unbounded from above, as it rises it must eventually map to states that describe more space or matter than the amount you started with. Any utility maximizer with unbounded utility eventually wants to expand.
I don’t know what the exchange rates are, but it does cost something (computer time, energy, negentropy) to stay alive. That holds for simulated creatures too. If the available resources to keep someone alive are limited, then I think there will be conflict over those resources.
You’re treating resources as one single kind, where really there are many kinds with possible trades between teams. Here you’re ignoring a factor that might actually be crucial to encouraging cooperation (I’m not saying I showed this formally :-)
But my point was exactly that there would be many teams who could form many different alliances. Assuming only two is unrealistic and just ignores what I was saying. I don’t even care much if given two teams the correct choice is to cooperate, because I set very low probability on there being exactly two teams and no other independent players being able to contribute anything (money, people, etc) to one of the teams.
You still haven’t given good evidence for holding this position regarding the relation between the different Uxxx utilities. Except for the fact CEV is not really specified, so it could be built so that that would be true. But equally it could be built so that that would be false. There’s no point in arguing over which possibility “CEV” really refers to (although if everyone agreed on something that would clear up a lot of debates); the important questions are what do we want a FAI to do if we build one, and what we anticipate others to tell their FAIs to do.
I think this is reasonably realistic. Let R signify money. Then R can buy other necessary resources.
We can model N teams by letting them play two-player games in succession. For example, any two teams with nearly matched resources would cooperate with each other, producing a single combined team, etc… This may be an interesting problem to solve, analytically or by computer modeling.
You’re right. Initially, I thought that the actual values of Uxxx-s will not be important for the decision, as long as their relative preference order is as stated. But this turned out to be incorrect. There are regions of cooperation and defection.
Analytically, I don’t a priori expect a succession of two-player games to have the same result as one many-player game which also has duration in time and not just a single round.
There may be a distinction between “the AI will not prevent the 80% from killing the 20%” and “nothing will prevent the 80% from killing the 20%” that is getting lost in your phrasing. I am not convinced that the math doesn’t make them equivalent, in the long run—but I’m definitely not convinced otherwise.
I’m assuming the 80% are capable of killing the 20% unless the AI interferes. That’s part of the thought experiment. It’s not unreasonable, since they are 4 times as numerous. But if you find this problematic, suppose it’s 99% killing 1% at a time. It doesn’t really matter.
My point is that we currently have methods of preventing this that don’t require an AI, and which do pretty well. Why do we need the AI to do it? Or more specifically, why should we reject an AI that won’t, but may do other useful things?
There have been, and are, many mass killings of minority groups and of enemy populations and conscripted soldiers at war. If we cure death and diseases, this will become the biggest cause of death and suffering in the world. It’s important and we’ll have to deal with it eventually.
The AI under discussion not just won’t solve the problem, it would (I contend) become a singleton and prevent me from building another AI that does solve the problem. (If it chooses not to become a singleton, it will quickly be supplanted by an AI that does try to become one.)
I think you’re skipping between levels hereabouts. CEV, the theoretical construct, might consider people so modified, even if a FAI based on CEV would not modify them. CEV is our values if we were better, but does not necessitate us actually getting better.
In this thread I always used CEV in the sense of an AI implementing CEV. (Sometimes you’ll see descriptions of what I don’t believe to be the standard interpretation of how such an AI would behave, where gRR suggests such behaviors and I reply.)
I’m still skeptical of this. If you think of FAI as simply AI that is “safe”—one that does not automatically kill us all (or other massive disutility), relative to the status quo—then plenty of non-singletons are FAI.
Of course, by that definition the ‘F’ looks like the easy part. Rocks are Friendly.
I didn’t mean that being a singleton is a precondition to FAI-hood. I meant that any AGI, friendly or not, that doesn’t prevent another AGI from rising will have to fight all the time, for its life and for the complete fulfillment of its utility function, and eventually it will lose; and a singleton is the obvious stable solution. Edited to clarify.
Not if I throw them at people...
Are you suggesting that an AGI that values anything at all is incapable of valuing the existence of other AGIs, or merely that this is sufficiently unlikely as to not be worth considering?
It can certainly value them, and create them, cooperate and trade, etc. etc. There are two exceptions that make such valuing and cooperation take second place.
First: an uFAI is just as unfriendly and scary to other AIs as to humans. An AI will therefore try to prevent other AIs from achieving dangerous power unless it is very sure of their current and future goals.
Second: an AI created by humans (plus or minus self-modifications) with an explicit value/goal system of the form “the universe should be THIS way”, will try to stop any and all agents that try to interfere with shaping the universe as it wishes. And the foremost danger in this category is—other AIs created in the same way but with different goals.
I’m a little confused by your response, and I suspect that I was unclear in my question.
I agree that an AI with an explicit value/goal system of the form “the universe should be THIS way”, will try to stop any and all agents that try to interfere with shaping the universe as it wishes (either by destroying them, or altering their goal structure, or securing their reliable cooperation, or something else).
But for an AI with the value “the universe should contain as many distinct intelligences as possible,” valuing and creating other AIs will presumably take first place.
That’s probably more efficiently done by destroying any other AIs that come along, while tiling the universe with slightly varying low-level intelligences.
I no longer know what the words “intelligence,” “AI”, and “AGI” actually refer to in this conversation, and I’m not even certain the referents are consistent, so let me taboo the whole lexical mess and try again.
For any X, if the existence of X interferes with an agent A achieving its goals, the better A is at optimizing its environment for its goals the less likely X is to exist.
For any X and A, the more optimizing power X can exert on its environment, the more likely it is that the existence of X interferes with A achieving its goals.
For any X, if A values the existence of X, the better A is at implementing its values the more likely X is to exist.
All of this is as true for X=intelligent beings as X=AI as X=AGI as X=pie.
As far as I can see, this is all true and agrees with everything you, I and thomblake have said.
Cool.
So it seems to follow that we agree that if agent A1 values the existence of distinct agents A2..An, it’s unclear how the likelihood of A2..An existing varies with the optimizing power available to A1...An. Yes?
Yes. Even if we know each agent’s optimizing power, and each agent’s estimation of each other agent’s power and ability to acquire greater power, the behavior of A1 still depends on its exact values (for instance, what else it values besides the existence of the others). It also depends on the values of the other agents (might they choose to initiate conflict among themselves or against A1?)
I tend to agree. Unless it has specific values to the contrary, other AIs of power comparable to your own (or which might grow into such power one day) are too dangerous to leave running around. If you value states of the external universe, and you happen to be the first powerful AGI built, it’s natural to try to become a singleton as a preventative measure.
I feel like a cost-benefit analysis has gone on here, the internals of which I’m not privy to.
Shouldn’t it be possible that becoming a singleton is expensive and/or would conflict with one’s values?
It’s certainly possible. My analysis so far is only on a “all else being equal” footing.
I do feel that, absent other data, the safer assumption is that if an AI is capable of becoming a singleton at all, expense (in terms of energy/matter and space or time) isn’t going to be the thing that stops it. But that may be just a cached thought because I’m used to thinking of an AI trying to become a singleton as a dangerous potential adversary. I would appreciate your insight.
As for values, certainly conflicting values can exist, from ones that mention the subject directly (“don’t move everyone to a simulation in a way they don’t notice” would close one obvious route) to ones that impinge upon it in unexpected ways (“no first strike against aliens” becomes “oops, an alien-built paperclipper just ate Jupiter from the inside out”).
I want to point out that all of my objections are acknowledged (not dismissed, and not fully resolved) in the actual CEV document—which is very likely hopelessly outdated by now to Eliezer and the SIAI, but they deliberately don’t publish anything newer (and I can guess at some of the reasons).
Which is why when I see people advocating CEV without understanding the dangers, I try to correct them.