Questions about what is ‘desirable’ or ‘good’ for X only require (and only have) answers when there’s some kind of information flow between the thinker and X.
No, that’s not non-interacting, because as you say later, you want to interact with it. I mean really strictly noninteracing: no information flow either way. Imagine it’s over the cosmic horizon.
So I’m assuming, in this case, that the scenario to judge is a material copy of everything in our “recursive cosmic horizon” (that is, including the stuff at the edge of the cosmic horizon of the stuff at the edge of our cosmic horizon and so on until everything morally significant has either been firmly excluded or included so no one at the “very outer edge” relative to Sol has a change in experience either over their entire “cosmic existence” for the next N trillion years so we and everything that will eventually be in our light cone has identical but non-interacting astronomical waste issues to deal with) and then that physical system is moved off to its own place that is unimaginably far away and isolated.
This triggers my platospace intuitions because, as near as I can figure, every possible such universe already “exists” as a mathematical object in platospace (given that weak version of that kind of “existence” predicate) and copies in platospace are the one situation where the identity of indiscernibles is completely meaningful.
That kind of duplication is a no-op (like adding zero in a context where the there are no opportunity costs because you could have computed something meaningful instead) and has no value.
For reference, I one-box on Newcombe’s paradox if there really is something that can verifiably predict what I’ll do (and I’m not being scammed by a huckster in an angel costume with some confederates who have pre-arranged to one-box or two-box or to signal intent via backchannels if I randomly instruct them how to pick for experimental purposes, etc, etc).
Back in like 2001 I tried to build a psych instrument that had better than random chance of predicting whether someone would one-box or two-box in a specific, grounded and controlled Newcombe’s Paradox situation—and that retained its calibration even when the the situation was modified into its probabilistic form and the actual measured calibration was honestly reported to the participants.
Eventually I moved on, because I ran into numerous practical difficulties (money, time, competing interests, etc) trying to implement a research program of this sort as an undergrad in my spare time :-P
Still, I’ve retained the orientation towards reality that makes me more interested in psychological instruments on the subject of Newcombe’s Paradox than about abstract models of rationality that are capable of getting the right answer. I don’t think it requires special math to get the right answer, it just requires clear thinking.
The gritty details of practical reality always seems to matter in the end, and questions about the terminal value of imaginary nothings have, in my experience, never mattered except for signaling tribal membership… where imaginary nothings matter a lot, but in complicated ways with their own quirky domain details that are probably pretty far off topic :-P
Back in like 2001 I tried to build a psych instrument that had better than random chance of predicting whether someone would one-box or two-box in a specific, grounded and controlled Newcomb’s Paradox situation—and that retained its calibration even when the the situation was modified into its probabilistic form and the actual measured calibration was honestly reported to the participants.
That sounds incredibly interesting and I’m curious what else one-boxing correlates with. By “instrument”, you mean a questionnaire? What kinds of things did you try asking? Wouldn’t the simplest way of doing that be to just ask “Would you one-box on Newcomb’s Problem?”
I think I might end up disappointing because I have almost no actual data...
By an instrument I meant a psychological instrument, probably initially just a quiz and if that didn’t work then perhaps some stroop-like measurements of millisecond delay when answering questions on a computer.
Most of my effort went into working out a strategy for iterative experimental design and brainstorming questions for the very first draft of the questionnaire. I didn’t really have a good theory about what pre-existing dispositions or “mental contents” might correlate with dispositions one way or the other.
I thought it would be funny if people who “believed in free will” in the manner of Martin Gardner (an avowed mysterian) turned out to be mechanically predictable on the basis of inferring that they are philosophically confused in ways that lead to two-boxing. Gardner said he would two box… but also predicted that it was impossible for anyone to successfully predict that he would two box.
In his 1974 “Mathematical Games” article in Scientific American he ended with a question:
But has either side really done more than just repeat its case “loudly and slowly”? Can it be that Newcombe’s paradox validates free will by invalidating the possibility, in principle, of a predictor capable of guessing a person’s choice between two equally rational actions* with better than 50% probability?
= From context I infer that he means that one and two boxing are equally rational and doesn’t mean to cover the more general claim this seems to imply.
In his post script to the same article, reprinted in The Night Is Large he wrote:
It is my view that Newcomb’s predictor, even if accurate only 51% of the time, forces a logical contradiction that makes such a prediction, like Russell’s barber, impossible. We can avoid the contradiction arising from two different “shoulds” (should you take one or two boxes?) by stating the contradiction as follows. One flawless argument implies that the best way to maximize your reward is to take only the closed box. Another flawless argument implies that the best way to maximize your reward is to take both boxes. Because the two conclusions are contradictory, the prediction cannot be even probably valid. Faced with a Newcomb decision, I would share the suspicions of Max Black and others that I was either the victim of a hoax or of a badly controlled experiment that had yielded false data about the predictor’s accuracy. On this assumption, I would take both boxes.
This obviously suggests a great opportunity for falsification by rolling up one’s sleeves and just doing it. But I didn’t get very far...
One reason I didn’t get very far is that I was a very poor college student and I had a number of worries about ecological validity if there wasn’t really some money on the line which I couldn’t put up.
A quick and dirty idea I had to just get moving was to just get a bunch of prefab psych instruments (like the MMPI and big five stuff but I tried tracking down other things like religious belief inventories and such) and then also make up a Newcomb’s quiz of my own, that explained the situation, had some comprehension questions, and then asked for “what would you do”.
The Newcomb’s quiz would just be “one test among many”, but I could score the quizes and come back to give the exact same Newcomb’s quiz a second time with a cover sheet explaining that the answer to the final question was actually going to determine payoffs for the subject. All the other tests would give a plausible reason that the prediction might be possible, act as a decoy (soliciting an unvarnished Newcomb’s opinion because it wouldn’t leap out), and provide fascinating side material to see what might be correlated with opinions about Newcomb’s paradox.
This plan foundered on my inability to find any other prefab quizes. I had thought, you know?… science? …openness? But in my context at that time and place (with the internet not nearly as mature as it is now, not having the library skills I now have, and so on) all my attempts to acquire such tests were failures.
One of the things I realized is that the claim about the accuracy might substantially change the behavior of the subject so I potentially had a chick and egg problem—even nonverbals could influence things as I handed the second set of papers claiming success rates of 1%, 50%, or 99% to help explore the stimulus-reaction space… it would be tricky. I considered eventually bringing in some conformity experiment stuff, like with confederates who one-box or two-box in a way the real subject could watch and maybe be initially fooled by, but that was just getting silly, given my resources.
Another issue is that, if a subject of prediction isn’t sure what the predictor may have determined about your predicted action, it seems plausible that the very first time you faced the situation you might have a unique opportunity to do something moderately creative like flipping a coin, having it tell you to two box, and coming out with both prizes, so one of the questions I wanted to stick in was something to very gently probe the possibility that the person would “go random” like this and optimize over this possibility. Do you test for this propensity in advance? How? How without suggesting the very possibility?
This also raises the interesting question about iterated prediction. Its one thing to predict that a smart 12 year old who has just been introduced to the paradox will do, and a different thing to give the test to people who have publicly stated what they would do, and still a different thing to run someone through the system five times in a row so the system and the subject started gaining mutually reflective information (for example, the results on the fifth attempt would probably wash out any information from confederates and giving the subject first hand experience with the success rate, but it creates all kinds of opportunities for the instrument to be hacked by a clever subject).
Or another angle, what about doing the experiment on a group of people where they get to talk about it and watch each other as they go through. Do people influence each other on this subject? Would the instrument have to know what the social context was to predict subject behavior?
This leads to the conclusion that I’d probably have to add some metadata to my instrument so that it could ask the question “how many times have you seen this specific questionnaire?” or “when was the first time you heard about Newcomb’s paradox?” and possibly also have the person giving the questionnaire fill in some metadata about recent history of previous attempts and/or the experimenter’s subjective estimate of the answers the person should give to familiarity questions.
Another problem was simply finding people with the patience to take the quizes :-P
I ended up never having a final stable version of the quiz and a little bit after that I became way more interested in complex system theory and then later more practical stuff like machine learning and bioinformatics and business models and whatnot—aiming for domain expertise in AI and nano for “save with world” purposes in the coming decades.
I think I did the right thing by leaving the academic psych track, but I was so young and foolish then that I’m still not sure. In any case, I haven’t seriously worked on Newcomb’s paradox in almost a decade.
Nowadays, I have a sort of suspicion that upon seeing Newcomb’s paradox, some people see that the interaction with the predictor is beneficial to them and that it would be good if they could figure out some way to get the most possible benefit from the situation, which would involve at least being predicted to one-box. Then its an open question as to whether its possible (or moral) to “cheat” and two-box on top of that.
So the suspicious part of this idea is that a lot of people’s public public claims in this area are a very very geeky form of signaling, with one-box being a way to say “Its simple: I want to completely win… but I won’t cheat”. I think the two-box choice is also probably a matter of signaling, and it functions as a way to say something like “I have put away childish aspirations of vaguely conceived get rich quick schemes, accepted that in the real world everyone can change their mind at any moment, and am really trustworthy because I’m neither lean nor hungry nor trying to falsely signal otherwise”.
Like… I bet C.S. Lewis would have two-boxed. The signaling theory reminds me of his version of moral jujitsu, though personally, I still one box—I want to win :-P
Back in like 2001 I tried to build a psych instrument that had better than random chance of predicting whether someone would one-box or two-box in a specific, grounded and controlled Newcombe’s Paradox situation—and that retained its calibration even when the the situation was modified into its probabilistic form and the actual measured calibration was honestly reported to the participants.
This sounds very interesting, so I second Blueberry’s questions.
One, if it’s set up in what I think is the standard way. (Though of course one can devise very similar problems like the Smoking Lesion where the ‘right’ answer would be two.)
I’m not entirely sure how you’re connecting this with the statement you quoted, but I will point out that there is information flow between the Newcomb player’s predisposition to one-box or two-box and the predictor’s prediction. And that without some kind of information flow there couldn’t be a correlation between the two (short of a Cosmic Coincidence.)
How many boxes do you take on Newcomb’s problem?
So I’m assuming, in this case, that the scenario to judge is a material copy of everything in our “recursive cosmic horizon” (that is, including the stuff at the edge of the cosmic horizon of the stuff at the edge of our cosmic horizon and so on until everything morally significant has either been firmly excluded or included so no one at the “very outer edge” relative to Sol has a change in experience either over their entire “cosmic existence” for the next N trillion years so we and everything that will eventually be in our light cone has identical but non-interacting astronomical waste issues to deal with) and then that physical system is moved off to its own place that is unimaginably far away and isolated.
This triggers my platospace intuitions because, as near as I can figure, every possible such universe already “exists” as a mathematical object in platospace (given that weak version of that kind of “existence” predicate) and copies in platospace are the one situation where the identity of indiscernibles is completely meaningful.
That kind of duplication is a no-op (like adding zero in a context where the there are no opportunity costs because you could have computed something meaningful instead) and has no value.
For reference, I one-box on Newcombe’s paradox if there really is something that can verifiably predict what I’ll do (and I’m not being scammed by a huckster in an angel costume with some confederates who have pre-arranged to one-box or two-box or to signal intent via backchannels if I randomly instruct them how to pick for experimental purposes, etc, etc).
Back in like 2001 I tried to build a psych instrument that had better than random chance of predicting whether someone would one-box or two-box in a specific, grounded and controlled Newcombe’s Paradox situation—and that retained its calibration even when the the situation was modified into its probabilistic form and the actual measured calibration was honestly reported to the participants.
Eventually I moved on, because I ran into numerous practical difficulties (money, time, competing interests, etc) trying to implement a research program of this sort as an undergrad in my spare time :-P
Still, I’ve retained the orientation towards reality that makes me more interested in psychological instruments on the subject of Newcombe’s Paradox than about abstract models of rationality that are capable of getting the right answer. I don’t think it requires special math to get the right answer, it just requires clear thinking.
The gritty details of practical reality always seems to matter in the end, and questions about the terminal value of imaginary nothings have, in my experience, never mattered except for signaling tribal membership… where imaginary nothings matter a lot, but in complicated ways with their own quirky domain details that are probably pretty far off topic :-P
That sounds incredibly interesting and I’m curious what else one-boxing correlates with. By “instrument”, you mean a questionnaire? What kinds of things did you try asking? Wouldn’t the simplest way of doing that be to just ask “Would you one-box on Newcomb’s Problem?”
I think I might end up disappointing because I have almost no actual data...
By an instrument I meant a psychological instrument, probably initially just a quiz and if that didn’t work then perhaps some stroop-like measurements of millisecond delay when answering questions on a computer.
Most of my effort went into working out a strategy for iterative experimental design and brainstorming questions for the very first draft of the questionnaire. I didn’t really have a good theory about what pre-existing dispositions or “mental contents” might correlate with dispositions one way or the other.
I thought it would be funny if people who “believed in free will” in the manner of Martin Gardner (an avowed mysterian) turned out to be mechanically predictable on the basis of inferring that they are philosophically confused in ways that lead to two-boxing. Gardner said he would two box… but also predicted that it was impossible for anyone to successfully predict that he would two box.
In his 1974 “Mathematical Games” article in Scientific American he ended with a question:
= From context I infer that he means that one and two boxing are equally rational and doesn’t mean to cover the more general claim this seems to imply.
In his post script to the same article, reprinted in The Night Is Large he wrote:
This obviously suggests a great opportunity for falsification by rolling up one’s sleeves and just doing it. But I didn’t get very far...
One reason I didn’t get very far is that I was a very poor college student and I had a number of worries about ecological validity if there wasn’t really some money on the line which I couldn’t put up.
A quick and dirty idea I had to just get moving was to just get a bunch of prefab psych instruments (like the MMPI and big five stuff but I tried tracking down other things like religious belief inventories and such) and then also make up a Newcomb’s quiz of my own, that explained the situation, had some comprehension questions, and then asked for “what would you do”.
The Newcomb’s quiz would just be “one test among many”, but I could score the quizes and come back to give the exact same Newcomb’s quiz a second time with a cover sheet explaining that the answer to the final question was actually going to determine payoffs for the subject. All the other tests would give a plausible reason that the prediction might be possible, act as a decoy (soliciting an unvarnished Newcomb’s opinion because it wouldn’t leap out), and provide fascinating side material to see what might be correlated with opinions about Newcomb’s paradox.
This plan foundered on my inability to find any other prefab quizes. I had thought, you know?… science? …openness? But in my context at that time and place (with the internet not nearly as mature as it is now, not having the library skills I now have, and so on) all my attempts to acquire such tests were failures.
One of the things I realized is that the claim about the accuracy might substantially change the behavior of the subject so I potentially had a chick and egg problem—even nonverbals could influence things as I handed the second set of papers claiming success rates of 1%, 50%, or 99% to help explore the stimulus-reaction space… it would be tricky. I considered eventually bringing in some conformity experiment stuff, like with confederates who one-box or two-box in a way the real subject could watch and maybe be initially fooled by, but that was just getting silly, given my resources.
Another issue is that, if a subject of prediction isn’t sure what the predictor may have determined about your predicted action, it seems plausible that the very first time you faced the situation you might have a unique opportunity to do something moderately creative like flipping a coin, having it tell you to two box, and coming out with both prizes, so one of the questions I wanted to stick in was something to very gently probe the possibility that the person would “go random” like this and optimize over this possibility. Do you test for this propensity in advance? How? How without suggesting the very possibility?
This also raises the interesting question about iterated prediction. Its one thing to predict that a smart 12 year old who has just been introduced to the paradox will do, and a different thing to give the test to people who have publicly stated what they would do, and still a different thing to run someone through the system five times in a row so the system and the subject started gaining mutually reflective information (for example, the results on the fifth attempt would probably wash out any information from confederates and giving the subject first hand experience with the success rate, but it creates all kinds of opportunities for the instrument to be hacked by a clever subject).
Or another angle, what about doing the experiment on a group of people where they get to talk about it and watch each other as they go through. Do people influence each other on this subject? Would the instrument have to know what the social context was to predict subject behavior?
This leads to the conclusion that I’d probably have to add some metadata to my instrument so that it could ask the question “how many times have you seen this specific questionnaire?” or “when was the first time you heard about Newcomb’s paradox?” and possibly also have the person giving the questionnaire fill in some metadata about recent history of previous attempts and/or the experimenter’s subjective estimate of the answers the person should give to familiarity questions.
Another problem was simply finding people with the patience to take the quizes :-P
I ended up never having a final stable version of the quiz and a little bit after that I became way more interested in complex system theory and then later more practical stuff like machine learning and bioinformatics and business models and whatnot—aiming for domain expertise in AI and nano for “save with world” purposes in the coming decades.
I think I did the right thing by leaving the academic psych track, but I was so young and foolish then that I’m still not sure. In any case, I haven’t seriously worked on Newcomb’s paradox in almost a decade.
Nowadays, I have a sort of suspicion that upon seeing Newcomb’s paradox, some people see that the interaction with the predictor is beneficial to them and that it would be good if they could figure out some way to get the most possible benefit from the situation, which would involve at least being predicted to one-box. Then its an open question as to whether its possible (or moral) to “cheat” and two-box on top of that.
So the suspicious part of this idea is that a lot of people’s public public claims in this area are a very very geeky form of signaling, with one-box being a way to say “Its simple: I want to completely win… but I won’t cheat”. I think the two-box choice is also probably a matter of signaling, and it functions as a way to say something like “I have put away childish aspirations of vaguely conceived get rich quick schemes, accepted that in the real world everyone can change their mind at any moment, and am really trustworthy because I’m neither lean nor hungry nor trying to falsely signal otherwise”.
Like… I bet C.S. Lewis would have two-boxed. The signaling theory reminds me of his version of moral jujitsu, though personally, I still one box—I want to win :-P
Second Blueberry’s question.
This sounds very interesting, so I second Blueberry’s questions.
(Edit—beaten by Eliezer—I guess I third them.)
One, if it’s set up in what I think is the standard way. (Though of course one can devise very similar problems like the Smoking Lesion where the ‘right’ answer would be two.)
I’m not entirely sure how you’re connecting this with the statement you quoted, but I will point out that there is information flow between the Newcomb player’s predisposition to one-box or two-box and the predictor’s prediction. And that without some kind of information flow there couldn’t be a correlation between the two (short of a Cosmic Coincidence.)
Do you, in general, use acausal (e.g. Wei Dai’s timeless) decision theory?
Wei Dai’s Updateles. Eliezer’s timeless.
Well, I do have an affinity towards it—I think it ‘gets the answers right’ in cases like Counterfactual Mugging.