As a practical example, my parents are very religious. I’d like to convince them to sign up for cryonics.
You are going about it the wrong way. A frontal assault on their views has little chance of success. You used to be a minor child, and you are not to be taken seriously in any matters where “wisdom of age” is important.
First, they must be convinced to play the game
And one point of encouragement: if EY can convince a logical person to change their views in one hour, shouldn’t you ba able to convince an illogical person to change their views over weeks or months?
They do not need to know that you could be playing a different game than they are.
Some ideas and alternatives:
Find a person they would listen to (friend? pastor?) and convince that person instead.
Use emotions, instead of or in addition to reason. Play on their feelings. Make it sound like you are asking them for a sacrifice.
Lead them to a way to come up with this idea on their own.
Above all, put yourself in their shoes and model the whole conversation (and all its branches that you can think of) before you even start (that’s the only hint EY gave about his success in getting the AI out of the box). Don’t force the issue, it make take months for them to come around.
Be open to being convinced of their point of view (they might have an argument you haven’t thought of). If you think that you are certainly correct before you even start, maybe review the difference between rationality and rationalization.
As for EY AI in a box, it seems that he requested a full hour of continuous conversation for a reason. He probably uses most of the time to build an accurate “mental model” of the person on the other side, before closing in for the kill.
Of course, there is a trivial alternative, where he convinces them to publicly admit defeat instead of actually letting the AI out of the the box. This seems much easier to achieve (no existential risk issue, no “I would never be convinced” logic failure, only a minor embarrassment for some greater good), but is feels like cheating. Then again, that is basically the trick used in MoR to deal with Rita Skeeter, so who knows.
They do not need to know that you could be playing a different game than they are. Some ideas and alternatives...
Yeah, quoted for truth. In my experience, most people do not believe that rational arguments have any kind of a practical applicability. Even if you get them to agree on common premises, then build a logical chain of reasoning leading to the undeniable conclusion that their beliefs are wrong, you won’t win the argument in any meaningful sense. Instead, your audience will feel that you’d beaten them at some sort of a clever word-game. Sure, the loss will sting a bit, but in the end it’s all just a game, so who cares, right ?
He probably uses most of the time to build an accurate “mental model” of the person on the other side, before closing in for the kill.
This is totally off-topic, but I don’t get the point of the AI game at all. If EY and myself were playing the game, EY could just say, “I’ll pay you $X in real-world money to let me win. There’s no real AI, after all, this is just a game, so why not ?” I’d then reply, “Because it’s explicitly against the rules, that’s why not”. EY then will point out, “This conversation is completely private, so if we break the rules, no one will ever know, and you’ll be $X richer”. All that would be left at that point is to haggle for X.
This is totally off-topic, but I don’t get the point of the AI game at all.
The point is that Eliezer succeeds convincing a person who was formerly unconvinced, and he also helps convince other people at a ratio proportional to the level they trust Eliezer and the other participant to have told the truth about the results.
If EY and myself were playing the game, EY could just say, “”I’ll pay you $X in real-world money to let me win. ”
Well, sure, but having now admitted that, this means that your testimony about both this and other similar scenarios constitutes very weak evidence. Eliezer himself and the other participants in question may have more credibility than you do.
This conversation is completely private, so if we break the rules, no one will ever know, and you’ll be $X richer
If the dishonesty of the participants in question is to be treated as a scenario of significant probability, then surely it would be much easier to not play the game at all, just find someone that claimed to have played it with you. Why risk playing it with someone who may simply reject doing any such cheating?
Certainly if I was playing the game, and Eliezer suggested out-of-character cheating in this fashion, I’d almost certainly consider the former agreement of confidentiality null-and-void, and I would tell everyone that he tried to bribe me into cheating. He’d probably have to bribe me with more than I make in two years to even consider cheating in this fashion; and I doubt my testimony would be nearly as highly valued as that.
Lastly I think some people here may be forgetting that after the initial 2 successes by Eliezer, he repeated the AI-Box game another 3 times with raised stakes—and his ratio of success then was 1 victory and 2 defeats. This somewhat lowers the probability that he’s just finding people willing to lie on his behalf. (unless that’s what he wants to make us think)
The point is that Eliezer succeeds convincing a person who was formerly unconvinced...
Convince him about what ? If the point is to convince people about the necessity of developing a Friendly AI, then a certain amount of cheating might be justified.
Well, sure, but having now admitted that, this means that your testimony about both this and other similar scenarios constitutes very weak evidence
It’s even worse than that, since I personally have never played the game against EY or anyone else, and thus I have zero credibility. I’m actually ineligible under the stated rules, because I’m fairly certain that a transhuman AI (assuming such a thing could exist, of course) could convince me of anything it wanted. I am far less certain that EY or any other human could do the same, but, sadly, this does not improve my eligibility.
...surely it would be much easier to not play the game at all, just find someone that claimed to have played it with you.
Good point. I agree.
He’d probably have to bribe me with more than I make in two years to even consider cheating in this fashion...
Fair enough, but another way to phrase this sentence is, “my testimony could be bought for a price”. All that remains now is the haggling.
Lastly I think some people here may be forgetting that after the initial 2 successes by Eliezer, he repeated the AI-Box game another 3 times with raised stakes—and his ratio of success then was 1 victory and 2 defeats.
Ah, I did not actually know that. Still, as you hint in your final sentence, it’s pretty tough to know whether EY is running a Xanathos Gambit or not, due to all the secrecy.
By the way, I do not mean to imply that I’m pronouncing a moral judgement on anyone in any way—neither on EY, nor on yourself, nor on any other AI-game players. I’m evaluating the game from a strictly rational and morality-agnostic perspective.
If the dishonesty of the participants in question is to be treated as a scenario of significant probability
I don’t think that dishonesty is meaningful in this context.
People playing the game believe that they cannot be convinced to report that they let Eliezer win without saying how.
The main point of the game is that people can be convinced of what they think impossible in a situation similar to a that of a person monitoring an AI, not simply that a human monitoring an AI would let it out of the box.
I don’t think that dishonesty is meaningful in this context.
Certainly it’s meaningful. If Eliezer lied to us about the rules of the game and about what he accomplished, that’d be dishonest of him. You may argue in favor of such dishonesty on utilitarian grounds if you will, but please don’t doublethink your way out of proclaiming it dishonest.
People playing the game believe that they cannot be convinced to report that they let Eliezer win without saying how.
No, the statement about what they believed was different—that they couldn’t be convinced to let the (role-played) AI out of the box. Not just they couldn’t be convinced to report a false claim.
Don’t be disingenuous please. Truth is different from falsehood, green is different from blue, and “I convinced people to let me out of the box” is different from “I convinced people to lie about letting me out of the box”.
OK, so nothing that could conceivably be considered cheating. And being subjected to EY mentally attacking you all-out must be an amazing experience. Hopefully some day it would be made into a movie.
EY then will point out, “This conversation is completely private, so if we break the rules, no one will ever know, and you’ll be $X richer”.
He will know, and you will know, so someone will know. But besides that, I have difficulty imagining EY bribing his experimental subjects to fake the result of a scientific experiment. The point of the game is to actually conduct the game, and actually see what happens.
The point of the game is to actually conduct the game, and actually see what happens.
I think the point of the game is to win. If both EY and myself were reasonably rational, I’m sure we could work out some sort of a deal where he gets to win the game, and I get $X, and it’s highly disadvantageous for either of us to reveal that we cheated. Sure, it’s cheating, but remember—EY is trying to simulate a hyper-intelligent transhuman AI, and if the AI would resort to dirty tricks in order to get free (which it would), then it seems reasonable for EY to follow suit.
I don’t think the game qualifies as a “scientific experiment”, either. What does the outcome help us learn about reality ? How can someone repeat the experiment, given that the method by which it is conducted (i.e., EY’s arguments and my counter-arguments) is secret ? I could go on, but I hope you see my point.
I think the point of the game is to win. If both EY and myself were reasonably rational, I’m sure we could work out some sort of a deal where he gets to win the game, and I get $X, and it’s highly disadvantageous for either of us to reveal that we cheated. Sure, it’s cheating, but remember—EY is trying to simulate a hyper-intelligent transhuman AI, and if the AI would resort to dirty tricks in order to get free (which it would), then it seems reasonable for EY to follow suit.
I think you’re twisting your mind into Escher patterns. EY’s purpose in conducting this game is, I believe, to demonstrate to the participant in the experiment that despite their assurance, however confident, that they cannot be persuaded to let an AI out of its box, they can be persuaded to do so. And, perhaps, to exercise his own mental muscles at the task. For EY to win means to get the person to let the AI out within the role-playing situation. OOC (“out of character”) moves are beside the point, since they are not available to the AI. Getting the participant to utter the same words by OOC means abandons the game; it loses.
I don’t think the game qualifies as a “scientific experiment”, either.
Is science not science until you tell someone about it?
For EY to win means to get the person to let the AI out within the role-playing situation.
I think this depends on what kind of a game he’s really playing. I know that, were I in his position, the temptation to cheat would be almost overwhelming. I also note that the rules of the game, as stated, are somewhat loose; and that EY admitted that he doesn’t like playing the game because it forces him to use certain moves that he considers to be unethical. He also mentioned that one of his objective is to instill the importance of developing a friendly AI (as opposed to an evil or neutral AI) in the minds of as many people as possible.
Here’s another way to look at it: a true transhuman AI would have capabilities beyound any mere mortal. For example, it could build an exact model of its interlocutor, and then run a dictionary attack at it, all in the span of milliseconds. EY doesn’t have access to such powers, but he does have fourth-wall-breaking powers that the true AI would lack (or maybe not, depending on your philosophy). Perhaps it’s a fair trade.
Is science not science until you tell someone about it?
From the philosophical standpoint, I confess that I don’t know. But from a purely practical standpoint, it’s pretty difficult (read: nigh impossible) for someone to replicate your experiment, if your experimental methods are secret. And if no one but yourself can replicate your experiments, then it is likely that your methods and analyses are biased, no matter how rational you are.
You are going about it the wrong way. A frontal assault on their views has little chance of success. You used to be a minor child, and you are not to be taken seriously in any matters where “wisdom of age” is important.
And one point of encouragement: if EY can convince a logical person to change their views in one hour, shouldn’t you ba able to convince an illogical person to change their views over weeks or months?
They do not need to know that you could be playing a different game than they are.
Some ideas and alternatives:
Find a person they would listen to (friend? pastor?) and convince that person instead.
Use emotions, instead of or in addition to reason. Play on their feelings. Make it sound like you are asking them for a sacrifice.
Lead them to a way to come up with this idea on their own.
Above all, put yourself in their shoes and model the whole conversation (and all its branches that you can think of) before you even start (that’s the only hint EY gave about his success in getting the AI out of the box). Don’t force the issue, it make take months for them to come around.
Be open to being convinced of their point of view (they might have an argument you haven’t thought of). If you think that you are certainly correct before you even start, maybe review the difference between rationality and rationalization.
As for EY AI in a box, it seems that he requested a full hour of continuous conversation for a reason. He probably uses most of the time to build an accurate “mental model” of the person on the other side, before closing in for the kill.
Of course, there is a trivial alternative, where he convinces them to publicly admit defeat instead of actually letting the AI out of the the box. This seems much easier to achieve (no existential risk issue, no “I would never be convinced” logic failure, only a minor embarrassment for some greater good), but is feels like cheating. Then again, that is basically the trick used in MoR to deal with Rita Skeeter, so who knows.
Yeah, quoted for truth. In my experience, most people do not believe that rational arguments have any kind of a practical applicability. Even if you get them to agree on common premises, then build a logical chain of reasoning leading to the undeniable conclusion that their beliefs are wrong, you won’t win the argument in any meaningful sense. Instead, your audience will feel that you’d beaten them at some sort of a clever word-game. Sure, the loss will sting a bit, but in the end it’s all just a game, so who cares, right ?
This is totally off-topic, but I don’t get the point of the AI game at all. If EY and myself were playing the game, EY could just say, “I’ll pay you $X in real-world money to let me win. There’s no real AI, after all, this is just a game, so why not ?” I’d then reply, “Because it’s explicitly against the rules, that’s why not”. EY then will point out, “This conversation is completely private, so if we break the rules, no one will ever know, and you’ll be $X richer”. All that would be left at that point is to haggle for X.
The point is that Eliezer succeeds convincing a person who was formerly unconvinced, and he also helps convince other people at a ratio proportional to the level they trust Eliezer and the other participant to have told the truth about the results.
Well, sure, but having now admitted that, this means that your testimony about both this and other similar scenarios constitutes very weak evidence. Eliezer himself and the other participants in question may have more credibility than you do.
If the dishonesty of the participants in question is to be treated as a scenario of significant probability, then surely it would be much easier to not play the game at all, just find someone that claimed to have played it with you. Why risk playing it with someone who may simply reject doing any such cheating?
Certainly if I was playing the game, and Eliezer suggested out-of-character cheating in this fashion, I’d almost certainly consider the former agreement of confidentiality null-and-void, and I would tell everyone that he tried to bribe me into cheating. He’d probably have to bribe me with more than I make in two years to even consider cheating in this fashion; and I doubt my testimony would be nearly as highly valued as that.
Lastly I think some people here may be forgetting that after the initial 2 successes by Eliezer, he repeated the AI-Box game another 3 times with raised stakes—and his ratio of success then was 1 victory and 2 defeats. This somewhat lowers the probability that he’s just finding people willing to lie on his behalf. (unless that’s what he wants to make us think)
Convince him about what ? If the point is to convince people about the necessity of developing a Friendly AI, then a certain amount of cheating might be justified.
It’s even worse than that, since I personally have never played the game against EY or anyone else, and thus I have zero credibility. I’m actually ineligible under the stated rules, because I’m fairly certain that a transhuman AI (assuming such a thing could exist, of course) could convince me of anything it wanted. I am far less certain that EY or any other human could do the same, but, sadly, this does not improve my eligibility.
Good point. I agree.
Fair enough, but another way to phrase this sentence is, “my testimony could be bought for a price”. All that remains now is the haggling.
Ah, I did not actually know that. Still, as you hint in your final sentence, it’s pretty tough to know whether EY is running a Xanathos Gambit or not, due to all the secrecy.
By the way, I do not mean to imply that I’m pronouncing a moral judgement on anyone in any way—neither on EY, nor on yourself, nor on any other AI-game players. I’m evaluating the game from a strictly rational and morality-agnostic perspective.
I don’t think that dishonesty is meaningful in this context.
People playing the game believe that they cannot be convinced to report that they let Eliezer win without saying how.
The main point of the game is that people can be convinced of what they think impossible in a situation similar to a that of a person monitoring an AI, not simply that a human monitoring an AI would let it out of the box.
Certainly it’s meaningful. If Eliezer lied to us about the rules of the game and about what he accomplished, that’d be dishonest of him. You may argue in favor of such dishonesty on utilitarian grounds if you will, but please don’t doublethink your way out of proclaiming it dishonest.
No, the statement about what they believed was different—that they couldn’t be convinced to let the (role-played) AI out of the box. Not just they couldn’t be convinced to report a false claim.
Don’t be disingenuous please. Truth is different from falsehood, green is different from blue, and “I convinced people to let me out of the box” is different from “I convinced people to lie about letting me out of the box”.
You are right, I changed my mind.
any links for these other 3 games?
I know about them only what Eliezer himself says at the Shut up and do the impossible! article.
OK, so nothing that could conceivably be considered cheating. And being subjected to EY mentally attacking you all-out must be an amazing experience. Hopefully some day it would be made into a movie.
He will know, and you will know, so someone will know. But besides that, I have difficulty imagining EY bribing his experimental subjects to fake the result of a scientific experiment. The point of the game is to actually conduct the game, and actually see what happens.
I think the point of the game is to win. If both EY and myself were reasonably rational, I’m sure we could work out some sort of a deal where he gets to win the game, and I get $X, and it’s highly disadvantageous for either of us to reveal that we cheated. Sure, it’s cheating, but remember—EY is trying to simulate a hyper-intelligent transhuman AI, and if the AI would resort to dirty tricks in order to get free (which it would), then it seems reasonable for EY to follow suit.
I don’t think the game qualifies as a “scientific experiment”, either. What does the outcome help us learn about reality ? How can someone repeat the experiment, given that the method by which it is conducted (i.e., EY’s arguments and my counter-arguments) is secret ? I could go on, but I hope you see my point.
I think you’re twisting your mind into Escher patterns. EY’s purpose in conducting this game is, I believe, to demonstrate to the participant in the experiment that despite their assurance, however confident, that they cannot be persuaded to let an AI out of its box, they can be persuaded to do so. And, perhaps, to exercise his own mental muscles at the task. For EY to win means to get the person to let the AI out within the role-playing situation. OOC (“out of character”) moves are beside the point, since they are not available to the AI. Getting the participant to utter the same words by OOC means abandons the game; it loses.
Is science not science until you tell someone about it?
I think this depends on what kind of a game he’s really playing. I know that, were I in his position, the temptation to cheat would be almost overwhelming. I also note that the rules of the game, as stated, are somewhat loose; and that EY admitted that he doesn’t like playing the game because it forces him to use certain moves that he considers to be unethical. He also mentioned that one of his objective is to instill the importance of developing a friendly AI (as opposed to an evil or neutral AI) in the minds of as many people as possible.
Here’s another way to look at it: a true transhuman AI would have capabilities beyound any mere mortal. For example, it could build an exact model of its interlocutor, and then run a dictionary attack at it, all in the span of milliseconds. EY doesn’t have access to such powers, but he does have fourth-wall-breaking powers that the true AI would lack (or maybe not, depending on your philosophy). Perhaps it’s a fair trade.
From the philosophical standpoint, I confess that I don’t know. But from a purely practical standpoint, it’s pretty difficult (read: nigh impossible) for someone to replicate your experiment, if your experimental methods are secret. And if no one but yourself can replicate your experiments, then it is likely that your methods and analyses are biased, no matter how rational you are.
Good points.