Decision Theory FAQ — LessWrong

120

Decision Theory FAQ

28th Feb 2013

69 min read

120

Decision theoryQ&A (format)Bayesian Decision TheoryTimeless Decision Theory

120

Decision Theory FAQ

1patrickscottshields

0patrickscottshields

3private_messaging

0private_messaging

2private_messaging

-1private_messaging

0private_messaging

1private_messaging

-2private_messaging

-2private_messaging

-2private_messaging

-2private_messaging

-2private_messaging

11loup-vaillant

4Vladimir_Nesov

0patrickscottshields

1Joseph Greenwood

0alexander_poddiakov

0Stuart_Armstrong

0adam_strandberg

0buybuydandavis

17Eliezer Yudkowsky

21Eliezer Yudkowsky

2Eliezer Yudkowsky

0JonatasMueller

3Eliezer Yudkowsky

4Eliezer Yudkowsky

0JonatasMueller

16Eliezer Yudkowsky

16Eliezer Yudkowsky

9Eliezer Yudkowsky

4Eliezer Yudkowsky

5Eliezer Yudkowsky

16Eliezer Yudkowsky

0Stuart_Armstrong

6Eliezer Yudkowsky

2Stuart_Armstrong

4Eliezer Yudkowsky

0Stuart_Armstrong

2Eliezer Yudkowsky

-2buybuydandavis

6Eliezer Yudkowsky

2JonatasMueller

25Eliezer Yudkowsky

16Eliezer Yudkowsky

7Eliezer Yudkowsky

New Comment

Rendering 485/487 comments, sorted by

(show more) Click to highlight new comments since: Today at 12:33 AM

Some comments are truncated due to high volume. (⌘F to expand all)Change truncation settings

[-]incogn13y163

I don't really think Newcomb's problem or any of its variations belong in here. Newcomb's problem is not a decision theory problem, the real difficulty is translating the underspecified English into a payoff matrix.

The ambiguity comes from the the combination of the two claims, (a) Omega being a perfect predictor and (b) the subject being allowed to choose after Omega has made its prediction. Either these two are inconsistent, or they necessitate further unstated assumptions such as backwards causality.

First, let us assume (a) but not (b), which can be formulated as follows: Omega, a computer engineer, can read your code and test run it as many times as he would like in advance. You must submit (simple, unobfuscated) code which either chooses to one- or two-box. The contents of the boxes will depend on Omega's prediction of your code's choice. Do you submit one- or two-boxing code?

Second, let us assume (b) but not (a), which can be formulated as follows: Omega has subjected you to the Newcomb's setup, but because of a bug in its code, its prediction is based on someone else's choice than yours, which has no correlation with your choice whatsoever. Do you one- or two-box?

Both of the... (read more)

1patrickscottshields13y

Thanks for this post; it articulates many of the thoughts I've had on the apparent inconsistency of common decision-theoretic paradoxes such as Newcomb's problem. I'm not an expert in decision theory, but I have a computer science background and significant exposure to these topics, so let me give it a shot. The strategy I have been considering in my attempt to prove a paradox inconsistent is to prove a contradiction using the problem formulation. In Newcomb's problem, suppose each player uses a fair coin flip to decide whether to one-box or two-box. Then Omega could not have a sustained correct prediction rate above 50%. But the problem formulation says Omega does; therefore the problem must be inconsistent. Alternatively, Omega knew the outcome of the coin flip in advance; let's say Omega has access to all relevant information, including any supposed randomness used by the decision-maker. Then we can consider the decision to already have been made; the idea of a choice occurring after Omega has left is illusory (i.e. deterministic; anyone with enough information could have predicted it.) Admittedly, as you say quite eloquently: In this case of the all-knowing Omega, talking about what someone should choose after Omega has left seems mistaken. The agent is no longer free to make an arbitrary decision at run-time, since that would have backwards causal implications; we can, without restricting which algorithm is chosen, require the decision-making algorithm to be written down and provided to Omega prior to the whole simulation. Since Omega can predict the agent's decision, the agent's decision does determine what's in the box, despite the usual claim of no causality. Taking that into account, CDT doesn't fail after all. It really does seem to me like most of these supposed paradoxes of decision theory have these inconsistent setups. I see that wedrifid says of coin flips: I would love to hear from someone in further detail on these issues of consistency. Have t

1wedrifid13y

This seems like a worthy approach to paradoxes! I'm going to suggest the possibility of broadening your search slightly. Specifically, to include the claim "and this is paradoxical" as one of the things that can be rejected as producing contradictions. Because in this case there just isn't a paradox. You take the one box, get rich and if there is a decision theory that says to take both boxes you get a better theory. For this reason "Newcomb's Paradox" is a misnomer and I would only use "Newcomb's Problem" as an acceptable name. Yes, if the player is allowed access to entropy that Omega cannot have then it would be absurd to also declare that Omega can predict perfectly. If the coin flip is replaced with a quantum coinflip then the problem becomes even worse because it leaves an Omega that can perfectly predict what will happen but is faced with a plainly inconsistent task of making contradictory things happen. The problem specification needs to include a clause for how 'randomization' is handled. Here is where I should be able to link you to the wiki page on free will where you would be given an explanation of why the notion that determinism is incompatible with choice is a confusion. Alas that page still has pretentious "Find Out For Yourself" tripe on it instead of useful content. The wikipedia page on compatibilism is somewhat useful but not particularly tailored to a reductionist decision theory focus. There have been attempts to create derivatives of CDT that work like that. That replace the "C" from conventional CDT with a type of causality that runs about in time as you mention. Such decision theories do seem to handle most of the problems that CDT fails at. Unfortunately I cannot recall the reference. I'm not sure which further details you are after. Are you after a description of Newcomb's problem that includes the details necessary to make it consistent? Or about other potential inconsistencies? Or other debates about whether the problems are inconsis

0patrickscottshields13y

Thanks for the response! I'm looking for a formal version of the viewpoint you reiterated at the beginning of your most recent comment: That makes a lot of sense, but I haven't been able to find it stated formally. Wolpert and Benford's papers (using game theory decision trees or alternatively plain probability theory) seem to formally show that the problem formulation is ambiguous, but they are recent papers, and I haven't been able to tell how well they stand up to outside analysis. If there is a consensus that the sufficient use of randomness prevents Omega from having perfect or nearly perfect predictions, then why is Newcomb's problem still relevant? If there's no randomness, wouldn't an appropriate application of CDT result in one-boxing since the decision-maker's choice and Omega's prediction are both causally determined by the decision-maker's algorithm, which was fixed prior to the making of the decision? I'm curious: why can't normal CDT handle it by itself? Consider two variants of Newcomb's problem: 1. At run-time, you get to choose the actual decision made in Newcomb's problem. Omega made its prediction without any information about your choice or what algorithms you might use to make it. In other words, Omega doesn't have any particular insight into your decision-making process. This means at run-time you are free to choose between one-boxing and two-boxing without backwards causal implications. In this case Omega cannot make perfect or nearly perfect predictions, for reasons of randomness which we already discussed. 2. You get to write the algorithm, the output of which will determine the choice made in Newcomb's problem. Omega gets access to the algorithm in advance of its prediction. No run-time randomness is allowed. In this case, Omega can be a perfect predictor. But the correct causal network shows that both the decision-maker's "choice" as well as Omega's prediction are causally downstream from the selection of the decision-making algorith

3private_messaging13y

You can consider an ideal agent that uses argmax E to find what it chooses, where E is some environment function . Then what you arrive at is that argmax gets defined recursively - E contains argmax as well - and it just so happens that the resulting expression is only well defined if there's nothing in the first box and you choose both boxes. I'm writing a short paper about that.

0crazy8813y

You may be thinking of Huw Price's paper available here

1Amanojack13y

I agree; wherever there is paradox and endless debate, I have always found ambiguity in the initial posing of the question. An unorthodox mathematician named Norman Wildberger just released a new solution by unambiguously specifying what we know about Omega's predictive powers.

1Creutzer13y

I seems to me that what he gives is not so much a new solution as a neat generalized formulation. His formula gives you different results depending on whether you're a causal decision theorist or not. The causal decision theorist will say that his pA should be considered to be P(prediction = A|do(A)) and pB is P(prediction = B|do(B)), which will, unless you assume backward causation, just be P(prediction = A) and P(prediction = B) and thus sum to 1, hence the inequality at the end doesn't hold and you should two-box.

5incogn13y

I do not agree that a CDT must conclude that P(A)+P(B) = 1. The argument only holds if you assume the agent's decision is perfectly unpredictable, i.e. that there can be no correlation between the prediction and the decision. This contradicts one of the premises of Newcomb's Paradox, which assumes an entity with exactly the power to predict the agent's choice. Incidentally, this reduces to the (b) but not (a) from above. By adopting my (a) but not (b) from above, i.e. Omega as a programmer and the agent as predictable code, you can easily see that P(A)+P(B) = 2, which means one-boxing code will perform the best. Further elaboration of the above: Imagine John, who never understood how the days of the week succeed each other. Rather, each morning, a cab arrives to take him to work if it is a work day, else he just stays at home. Omega must predict if he will go to work or not the before the cab would normally arrive. Omega knows that weekdays are generally workdays, while weekends are not, but Omega does not know the ins and outs of particular holidays such as fourth of July. Omega and John play this game each day of the week for a year. Tallying the results, John finds that the score is as follows: P( Omega is right | I go to work) = 1.00, P( Omega is right | I do not go to work) = 0.85, which sums to 1.85. John, seeing that the sum is larger than 1.00, concludes that Omega seems to have rather good predictive power about whether he will go to work, but is somewhat short of perfect accuracy. He realizes that this has a certain significance for what bets he should take with Omega, regarding whether he will go to work tomorrow or not.

-1Creutzer13y

But that's not CDT reasoning. CDT uses surgery instead of conditionalization, that's the whole point. So it doesn't look at P(prediction = A|A), but at P(prediction = A|do(A)) = P(prediction = A). Your example with the cab doesn't really involve a choice at all, because John's going to work is effectively determined completely by the arrival of the cab.

0incogn13y

I am not sure where our disagreement lies at the moment. Are you using choice to signify strongly free will? Because that means the hypothetical Omega is impossible without backwards causation, leaving us at (b) but not (a) and the whole of Newcomb's paradox moot. Whereas, if you include in Newcomb's paradox, the choice of two-boxing will actually cause the big box to be empty, whereas the choice of one-boxing will actually cause the big box to contain a million dollars by a mechanism of backwards causation, then any CDT model will solve the problem. Perhaps we can narrow down our disagreement by taking the following variation of my example, where there is at least a bit more of choice involved: Imagine John, who never understood why he gets thirsty. Despite there being a regularity in when he chooses to drink, this is for him a mystery. Every hour, Omega must predict whether John will choose to drink within the next hour. Omega's prediction is made secret to John until after the time interval has passed. Omega and John play this game every hour for a month, and it turns out that while far from perfect, Omega's predictions are a bit better than random. Afterwards, Omega explains that it beats blind guesses by knowing that John will very rarely wake up in the middle of the night to drink, and that his daily water consumption follows a normal distribution with a mean and standard deviation that Omega has estimated.

0Creutzer13y

I'm not entirely sure either. I was just saying that a causal decision theorist will not be moved by Wildberger's reasoning, because he'll say that Wildberger is plugging in the wrong probabilities: when calculating an expectation, CDT uses not conditional probability distributions but surgically altered probability distributions. You can make that result in one-boxing if you assume backwards causation. I think the point we're actually talking about (or around) might be the question of how CDT reasoning relates to you (a). I'm not sure that the causal decision theorist has to grant that he is in fact interpreting the problem as "not (a) but (b)". The problem specification only contains the information that so far, Omega has always made correct predictions. But the causal decision theorist is now in a position to spoil Omega's record, if you will. Omega has already made a prediction, and whatever the causal decision theorist does now isn't going to change that prediction. The fact that Omega's predictions have been absolutely correct so far doesn't enter into the picture. It just means that for all agents x that are not the causal decision theorist, P(x does A|Omega predicts that x does A) = 1 (and the same for B, and whatever value than 1 you might want for an imperfect predictor Omega). About the way you intend (a), the causal decision theorist would probably say that's backward causation and refuse to accept it. One way of putting it might be that the causal decision theorist simply has no way of reasoning with the information that his choice is predetermined, which is what I think you intend to convey with (a). Therefore, he has no way of (hypothetically) inferring Omega's prediction from his own (hypothetical) action (because he's only allowed to do surgery, not conditionalization). No, actually. Just the occurrence of a deliberation process whose outcome is not immediately obvious. In both your examples, that doesn't happen: John's behavior simply depends o

[-]incogn13y100

(Thanks for discussing!)

I will address your last paragraph first. The only significant difference between my original example and the proper Newcomb's paradox is that, in Newcomb's paradox, Omega is made a predictor by fiat and without explanation. This allows perfect prediction and choice to sneak into the same paragraph without obvious contradiction. It seems, if I try to make the mode of prediction transparent, you protest there is no choice being made.

From Omega's point of view, its Newcomb subjects are not making choices in any substantial sense, they are just predictably acting out their own personality. That is what allows Omega its predictive power. Choice is not something inherent to a system, but a feature of an outsider's model of a system, in much the same sense as random is not something inherent to a Eeny, meeny, miny, moe however much it might seem that way to children.

As for the rest of our disagreement, I am not sure why you insist that CDT must work with a misleading model. The standard formulation of Newcomb's paradox is inconsistent or underspecified. Here are some messy explanations for why, in list form:

Omega predicts accurately, then you get to choose is a

... (read more)

3MugaSofer13y

Not if you're a compatibilist, which Eliezer is last I checked.

2incogn13y

The post scav made more or less represents my opinion here. Compatibilism, choice, free will and determinism are too many vague definitions for me to discuss with. For compatibilism to make any sort of sense to me, I would need a new definition of free will. It is already difficult to discuss how stuff is, without simultaneously having to discuss how to use and interpret words. Trying to leave the problematic words out of this, my claim is that the only reason CDT ever gives a wrong answer in a Newcomb's problem is that you are feeding it the wrong model. http://lesswrong.com/lw/gu1/decision_theory_faq/8kef elaborates on this without muddying the waters too much with the vaguely defined terms.

1scav13y

I don't think compatibilist means that you can pretend two logically mutually exclusive propositions can both be true. If it is accepted as a true proposition that Omega has predicted your actions, then your actions are decided before you experience the illusion of "choosing" them. Actually, whether or not there is an Omega predicting your actions, this may still be true. Accepting the predictive power of Omega, it logically follows that when you one-box you will get the $1M. A CDT-rational agent only fails on this if it fails to accept the prediction and constructs a (false) causal model that includes the incoherent idea of "choosing" something other than what must happen according to the laws of physics. Does CDT require such a false model to be constructed? I dunno. I'm no expert. The real causal model is that some set of circumstances decided what you were going to "choose" when presented with Omega's deal, and those circumstances also led to Omega's 100% accurate prediction. If being a compatibilist leads you to reject the possibility of such a scenario, then it also logically excludes the perfect predictive power of Omega and Newcomb's problem disappears. But in the problem as stated, you will only two-box if you get confused about the situation or you don't want $1M for some reason.

4ArisKatsaris13y

Where's the illusion? If I choose something according to my own preferences, why should it be an illusion merely because someone else can predict that choice if they know said preferences? Why does their knowledge of my action affect my decision-making powers? The problem is you're using the words "decided" and "choosing" confusingly with -- different meanings at the same time. One meaning is having the final input on the action I take -- the other meaning seems to be a discussion of when the output can be calculated. The output can be calculated before I actually even insert the input, sure -- but it's still my input, and therefore my decision -- nothing illusory about it, no matter how many people calculated said input in advance: even though they calculated it was I who controlled it.

0scav13y

The knowledge of your future action is only knowledge if it has a probability of 1. Omega acquiring that knowledge by calculation or otherwise does not affect your choice, but it is a consequence of that knowledge being able to exist (whether Omega has it or not) that means your choice is determined absolutely. What happens next is exactly the everyday meaning of "choosing". Signals zap around your brain in accordance with the laws of physics and evaluate courses of action according to some neural representation of your preferences, and one course of action is the one you will "decide" to do. Soon afterwards, your conscious mind becomes aware of the decision and feels like it made it. That's one part of the illusion of choice. EDIT: I'm assuming you're a human. A rational agent need not have this incredibly clunky architecture. The second part of the illusion is specific to this very artificial problem. The counterfactual (you choose the opposite of what Omega predicted) just DOESN'T EXIST. It has probability 0. It's not even that it could have happened in another branch of the multiverse - it is logically precluded by the condition of Omega being able to know with probability 1 what you will choose. 1 - 1 = 0.

1ArisKatsaris13y

Do you think Newcomb's Box fundamentally changes if Omega is only right with a probability of 99.9999999999999%? That process "is" my mind -- there's no mind anywhere which can be separate from those signals. So you say that my mind feels like it made a decision but you think this is false? I think it makes sense to say that my mind feels like it made a decision and it's completely right most of the time. My mind would be only having the "illusion" of choice if someone else, someone outside my mind, intervened between the signals and implanted a different decision, according to their own desires, and the rest of my brain just rationalized the already pretaken choice. But as long as the process is truly internal, the process is truly my mind's -- and my mind's feeling that it made the choice corresponds to reality. That the opposite choice isn't made in any universe, doesn't mean that the actually made choice isn't real -- indeed the less real the opposite choice, the more real your actual choice. Taboo the word "choice", and let's talk about "decision-making process". Your decision-making process exists in your brain, and therefore it's real. It doesn't have to be uncertain in outcome to be real -- it's real in the sense that it is actually occuring. Occuring in a deterministic manner, YES -- but how does that make the process any less real? Is gravity unreal or illusionary because it's deterministic and predictable? No. Then neither is your decision-making process unreal or illusionary.

0scav13y

Yes, it is your mind going through a decision making process. But most people feel that their conscious mind is the part making decisions and for humans, that isn't actually true, although attention seems to be part of consciousness and attention to different parts of the input probably influences what happens. I would call that feeling of making a decision consciously when that isn't really happening somewhat illusory. The decision making process is real, but my feeling of there being an alternative I could have chosen instead (even though in this universe that isn't true) is inaccurate. Taboo "illusion" too if you like, but we can probably agree to call that a different preference for usage of the words and move on. Incidentally, I don't think Newcomb's problem changes dramatically as Omega's success rate varies. You just get different expected values for one-boxing and two-boxing on a continuous scale, don't you?

0private_messaging13y

Regarding illegal choices, the transparent variation makes it particularly clear, i.e. you can't take both boxes if you see a million in first box, and take 1 box otherwise. You can walk backwards from your decision to the point where a copy of you had been made, and then forward to the point where a copy is processed by the Omega, to find the relation of your decision to the box state causally.

-3incogn13y

I agree with the content, though I am not sure if I approve of a terminology where causation traverses time like a two-way street.

2private_messaging13y

Underlying physics is symmetric in time. If you assume that the state of the world is such that one box is picked up by your arm, that imposes constraints on both the future and the past light cone. If you do not process the constraints on the past light cone then your simulator state does not adhere to the laws of physics, namely, the decision arises out of thin air by magic. If you do process constraints fully then the action to take one box requires pre-copy state of "you" that leads to decision to pick one box, which requires money in one box; action to take 2 boxes likewise, after processing constraints, requires no money in the first box. ("you" is a black box which is assumed to be non-magical, copyable, and deterministic, for the purpose of the exercise). edit: came up with an example. Suppose 'you' is a robotics controller, you know you're made of various electrical components, you're connected to the battery and some motors. You evaluate a counter factual where you put a current onto a wire for some time. Constraints imposed on the past: battery has been charged within last 10 hours, because else it couldn't supply enough current. If constraints contradict known reality then you know you can't do this action. Suppose there's a replacement battery pack 10 meters away from the robot, the robot is unsure if 5 hours ago the packs have been swapped; in the alternative that they haven't been, it would not have enough charge to get to the extra pack, in the alternative that they have been swapped, it doesn't need to get to the spent extra pack. Evaluating the hypothetical where it got to the extra pack it knows the packs have been swapped in the past and extra pack is spent. (Of course for simplicity one can do all sorts of stuff, such as electrical currents coming out of nowhere, but outside the context of philosophical speculation the cause of the error is very clear).

-1incogn13y

We do, by and large, agree. I just thought, and still think, the terminology is somewhat misleading. This is probably not a point I should press, because I have no mandate to dictate how words should be used, and I think we understand each other, but maybe it is worth a shot. I fully agree that some values in the past and future can be correlated. This is more or less the basis of my analysis of Newcomb's problem, and I think it is also what you mean by imposing constraints on the past light cone. I just prefer to use different words for backwards correlation and forwards causation. I would say that the robot getting the extra pack necessitates that it had already been charged and did not need the extra pack, while not having been charged earlier would cause it to fail to recharge itself. I think there is a significant difference between how not being charged causes the robot to run out of power, versus how running out of power necessitates that is has not been charged. You may of course argue that the future and the past are the same from the viewpoint of physics, and that either can said to cause another. However, as long as people consider the future and the past to be conceptually completely different, I do not see the hurry to erode these differences in the language we use. It probably would not be a good idea to make tomorrow refer to both the day before and the day after today, either. I guess I will repeat: This is probably not a point I should press, because I have no mandate to dictate how words should be used.

-1private_messaging13y

I'd be the first to agree on terminology here. I'm not suggesting that choice of the box causes money in the box, simply that those two are causally connected, in the physical sense. The whole issue seems to stem from taking the word 'causal' from causal decision theory, and treating it as more than mere name, bringing in enormous amounts of confused philosophy which doesn't capture very well how physics work. When deciding, you evaluate hypotheticals of you making different decisions. A hypothetical is like a snapshot of the world state. Laws of physics very often have to be run backwards from the known state to deduce past state, and then forwards again to deduce future state. E.g. a military robot sees a hand grenade flying into it's field of view, it calculates motion backwards to find where it was thrown from, finding location of the grenade thrower, then uses model of grenade thrower to predict another grenade in the future. So, you process the hypothetical where you picked up one box, to find how much money you get. You have the known state: you picked one box. You deduce that past state of deterministic you must have been Q which results in picking up one box, a copy of that state has been made, and that state resulted in prediction of 1 box. You conclude that you get 1 million. You do same for picking 2 boxes, the previous state must be R, etc, you conclude you get 1000 . You compare, and you pick the universe where you get 1 box. (And with regards to the "smoking lesion" problem, smoking lesion postulates a blatant logical contradiction - it postulates that the lesion affects the choice, which contradicts that the choice is made by the agent we are speaking of. As a counter example to a decision theory, it is laughably stupid)

1incogn13y

Excellent. I think laughably stupid is a bit too harsh. As I understand thing, confusion regarding Newcomb's leads to new decision theories, which in turn makes the smoking lesion problem interesting because the new decision theories introduce new, critical weaknesses in order to solve Newcomb's problem. I do, agree, however, that the smoking lesion problem is trivial if you stick to a sensible, CDT model.

0private_messaging13y

The problems with EDT are quite ordinary... its looking for good news, and also, it is kind of under-specified (e.g. some argue it'd two-box in Newcomb's after learning physics). A decision theory can not be disqualified for giving 'wrong' answer in the hypothetical that 2*2=5 or in the hypothetical that a or not a = false, or in the hypothetical that the decision is simultaneously controlled by the decision theory, and set, without involvement of the decision theory, by the lesion (and a random process if correlation is imperfect).

0Creutzer13y

I probably wasn't expressing myself quite clearly. I think the difference is this: Newcomb subjects are making a choice from their own point of view. Your Johns aren't really make a choice even from their internal perspective: they just see if the cab arrives/if they're thirsty and then without deliberation follow what their policy for such cases prescribes. I think this difference is substantial enough intuitively so that the John cases can't be used as intuition pumps for anything relating to Newcomb's. I don't think it is, actually. It just seems so because it presupposes that your own choice is predetermined, which is kind of hard to reason with when you're right in the process of making the choice. But that's a problem with your reasoning, not with the scenario. In particular, the CDT agent has a problem with conceiving of his own choice as predetermined, and therefore has trouble formulating Newcomb's problem in a way that he can use - he has to choose between getting two-boxing as the solution or assuming backward causation, neither of which is attractive.

5incogn13y

Then I guess I will try to leave it to you to come up with a satisfactory example. The challenge is to include Newcomblike predictive power for Omega, but not without substantiating how Omega achieves this, while still passing your own standards of subject makes choice from own point of view. It is very easy to accidentally create paradoxes in mathematics, by assuming mutually exclusive properties for an object, and the best way to discover these is generally to see if it is possible construct or find an instance of the object described. This is not a failure of CDT, but one of your imagination. Here is a simple, five minute model which has no problems conceiving Newcomb's problem without any backwards causation: * T=0: Subject is initiated in a deterministic state which can be predicted by Omega. * T=1: Omega makes an accurate prediction for the subject's decision in Newcomb's problem by magic / simulation / reading code / infallible heuristics. Denote the possible predictions P1 (one-box) and P2. * T=2: Omega sets up Newcomb's problem with appropriate box contents. * T=3: Omega explains the setup to the subject and disappears. * T=4: Subject deliberates. * T=5: Subject chooses either C1 (one-box) or C2. * T=6: Subject opens box(es) and receives payoff dependent on P and C. You can pretend to enter this situation at T=4 as suggested by the standard Newcomb's problem. Then you can use the dominance principle and you will lose. But this just using a terrible model. You entered at T=0, because you were needed at T=1 for Omega's inspection. If you did not enter the situation at T=0, then you can freely make a choice C at T=5 without any correlation to P, but that is not Newcomb's problem. Instead, at T=4 you become aware of the situation, and your decision making algorithm must return a value for C. If you consider this only from T=4 and onward, this is completely uninteresting, because C is already determined. At T=1, P was determined to either P1 or P2, an

0Creutzer13y

But isn't this precisely the basic idea behind TDT? The algorithm you are suggesting goes something like this: Chose that action which, if it had been predetermined at T=0 that you would take it, would lead to the maximal-utility outcome. You can call that CDT, but it isn't. Sure, it'll use causal reasoning for evaluating the counterfactual, but not everything that uses causal reasoning is CDT. CDT is surgically altering the action node (and not some precommitment node) and seeing what happens.

0incogn13y

If you take a careful look at the model, you will realize that the agent has to be precommited, in the sense that what he is going to do is already fixed. Otherwise, the step at T=1 is impossible. I do not mean that he has precommited himself consciously to win at Newcomb's problem, but trivially, a deterministic agent must be precommited. It is meaningless to apply any sort of decision theory to a deterministic system. You might as well try to apply decision theory to the balls in a game of billiards, which assign high utility to remaining on the table but have no free choices to make. For decision theory to have a function, there needs to be a choice to be made between multiple, legal options. As far as I have understood, your problem is that, if you apply CDT with an action node at T=4, it gives the wrong answer. At T=4, there is only one option to choose, so the choice of decision theory is not exactly critical. If you want to analyse Newcomb's problem, you have to insert an action node at T<1, while there is still a choice to be made, and CDT will do this admirably.

0Creutzer13y

Yes, it is. The point is that you run your algorithm at T=4, even if it is deterministic and therefore its output is already predetermined. Therefore, you want an algorithm that, executed at T=4, returns one-boxing. CDT does simply not do that. Ultimately, it seems that we're disagreeing about terminology. You're apparently calling something CDT even though it does not work by surgically altering the node for the action under consideration (that action being the choice of box, not the precommitment at T<1) and then looking at the resulting expected utilities.

-2incogn13y

If you apply CDT at T=4 with a model which builds in the knowledge that the choice C and the prediction P are perfectly correlated, it will one-box. The model is exceedingly simple: * T'=0: Choose either C1 or C2 * T'=1: If C1, then gain 1000. If C2, then gain 1. This excludes the two other impossibilities, C1P2 and C2P1, since these violate the correlation constraint. CDT makes a wrong choice when these two are included, because then you have removed the information of the correlation constraint from the model, changing the problem to one in which Omega is not a predictor. What is your problem with this model?

0Creutzer13y

Okay, so I take it to be the defining characteristic of CDT that it uses of counterfactuals. So far, I have been arguing on the basis of a Pearlean conception of counterfactuals, and then this is what happens: Your causal network has three variables, A (the algorithm used), P (Omega's prediction), C (the choice). The causal connections are A -> P and A -> C. There is no causal connection between P and C. Now the CDT algorithm looks at counterfactuals with the antecedent C1. In a Pearlean picture, this amounts to surgery on the C-node, so no inference contrary to the direction of causality is possible. Hence, whatever the value of the P-node, it will seem to the CDT algorithm not to depend on the choice. Therefore, even if the CDT algorithm knows that its choice is predetermined, it cannot make use of that in its decision, because it cannot update contrary to the direction of causality. Now it turns out that natural language counterfactuals work very much, but not quite like Pearl's counterfactuals: they allow a limited amount of backtracking contrary to the direction of causality, depending on a variety of psychological factors. So if you had a theory of counterfactuals that allowed backtracking in a case like Newcomb's problem, then a CDT-algorithm employing that conception of counterfactuals would one-box. The trouble would of course be to correctly state the necessary conditions for backtracking. The messy and diverse psychological and contextual factors that seem to be at play in natural language won't do.

1incogn13y

Could you try to maybe give a straight answer to, what is your problem with my model above? It accurately models the situation. It allows CDT to give a correct answer. It does not superficially resemble the word for word statement of Newcomb's problem. You are trying to use a decision theory to determine which choice an agent should make, after the agent has already had its algorithm fixed, which causally determines which choice the agent must make. Do you honestly blame that on CDT?

3Creutzer13y

No, it does not, that's what I was trying to explain. It's what I've been trying to explain to you all along: CDT cannot make use of the correlation between C and P. CDT cannot reason backwards in time. You do know how surgery works, don't you? In order for CDT to use the correlation, you need a causal arrow from C to P - that amounts to backward causation, which we don't want. Simple as that. I'm not sure what the meaning of this is. Of course the decision algorithm is fixed before it's run, and therefore its output is predetermined. It just doesn't know its own output before it has computed it. And I'm not trying to figure out what the agent should do - the agent is trying to figure that out. Our job is to figure out which algorithm the agent should be using. PS: The downvote on your post above wasn't from me.

-1incogn13y

You are applying a decision theory to the node C, which means you are implicitly stating: there are multiple possible choices to be made at this point, and this decision can be made independent of nodes not in front of this one. This means that your model does not model the Newcomb's problem we have been discussing - it models another problem, where C can have values independent of P, which is indeed solved by two-boxing. It is not the decision theory's responsibility to know that the values of node C is somehow supposed to retrospectively alter the state of the branch the decision theory is working in. This is, however,a consequence of the modelling you do. You are on purpose applying CDT too late in your network, such that P and thus the cost of being a two-boxer has gone over the horizon and such that the node C must affect P backwards, not because the problem actually contains backwards causality, but because you want to fix the value of nodes in the wrong order. If you do not want to make the assumption of free choice at C, then you can just not promote it to an action node. If the decision at C is casually determined from A, then you can apply a decision theory at node A and follow the causal inference. Then you will, once again, get a correct answer from CDT, this time for the version of Newcomb's problem where A and C are fully correlated. If you refuse to reevaluate your model, then we might as well leave it at this. I do agree that if you insist on applying CDT at C in your model, then it will two-box. I do not agree that this is a problem.

2Creutzer13y

You don't promote C to the action node, it is the action node. That's the way the decision problem is specified: do you one-box or two-box? If you don't accept that, then you're talking about a different decision problem. But in Newcomb's problem, the algorithm is trying to decide that. It's not trying to decide which algorithm it should be (or should have been). Having the algorithm pretend - as a means of reaching a decision about C - that it's deciding which algorithm to be is somewhat reminiscent of the idea behind TDT and has nothing to do with CDT as traditionally conceived of, despite the use of causal reasoning.

1private_messaging13y

In AI, you do not discuss it in terms of anthropomorphic "trying to decide". For example, there's a "Model based utility based agent" . Computing what the world will be like if a decision is made in a specific way is part of the model of the world, i.e. part of the laws of physics as the agent knows them. If this physics implements the predictor at all, model-based utility-based agent will one-box.

0Creutzer13y

I don't see at all what's wrong or confusing about saying that an agent is trying to decide something; or even, for that matter, that an algorithm is trying to decide something, even though that's not a precise way of speaking. More to the point, though, doesn't what you describe fit EDT and CDT both, with each theory having a different way of computing "what the world will be like if the decision is made in a specific way"?

-2incogn13y

Decision theories do not compute what the world will be like. Decision theories select the best choice, given a model with this information included. How the world works is not something a decision theory figures out, it is not a physicist and it has no means to perform experiments outside of its current model. You need take care of that yourself, and build it into your model. If a decision theory had the weakness that certain, possible scenarios could not be modeled, that would be a problem. Any decision theory will have the feature that they work with the model they are given, not with the model they should have been given.

-2private_messaging13y

Causality is under specified, whereas the laws of physics are fairly well defined, especially for a hypothetical where you can e.g. assume deterministic Newtonian mechanics for sake of simplifying the analysis. You have the hypothetical: sequence of commands to the robotic manipulator. You process the laws of physics to conclude that this sequence of commands picks up one box of unknown weight. You need to determine weight of the box to see if this sequence of commands will lead to the robot tipping over. Now, you see, to determine that sort of thing, models of physical world tend to walk backwards and forwards in time: for example if your window shatters and a rock flies in, you can conclude that there's a rock thrower in the direction that the rock came from, and you do it by walking backwards in time.

0Creutzer13y

So it's basically EDT, where you just conditionalize on the action being performed?

-2private_messaging13y

In a way, albeit it does not resemble how EDT tends to be presented. On the CDT, formally speaking, what do you think P(A if B) even is? Keep in mind that given some deterministic, computable laws of physics, given that you ultimately decide an option B, in the hypothetical that you decide an option C where C!=B , it will be provable that C=B , i.e. you have a contradiction in the hypothetical.

-2Creutzer13y

So then how does it not fall prey to the problems of EDT? It depends on the precise formalization of "computing what the world will be like if the action is taken, according to the laws of physics", of course, but I'm having trouble imagining how that would not end up basically equivalent to EDT. That is not the problem at all, it's perfectly well-defined. I think if anything, the question would be what CDT's P(A if B) is intuitively.

-2private_messaging13y

What are those, exactly? The "smoking lesion"? It specifies that output of decision theory correlates with lesion. Who knows how, but for it to actually correlate with decision of that decision theory other than via the inputs to decision theory, it got to be our good old friend Omega doing some intelligent design and adding or removing that lesion. (And if it does through the inputs, then it'll smoke). Given world state A which evolves into world state B (computable, deterministic universe), the hypothetical "what if world state A evolved into C where C!=B" will lead, among other absurdities, to a proof that B=C contradicting that B!=C (of course you can ensure that this particular proof won't be reached with various silly hacks but you're still making false assumptions and arriving at false conclusions). Maybe what you call 'causal' decision theory should be called 'acausal' because it in fact ignores causes of the decision, and goes as far as to break down it's world model to do so. If you don't do contradictory assumptions, then you have a world state A that evolves into world state B, and world state A' that evolves into world state C, and in the hypothetical that the state becomes C!=B, the prior state got to be A'!=A . Yeah, it looks weird to westerners with their philosophy of free will and your decisions having the potential to send the same world down a different path. I am guessing it is much much less problematic if you were more culturally exposed to determinism/fatalism. This may be a very interesting topic, within comparative anthropology. The main distinction between philosophy and mathematics (or philosophy done by mathematicians) seem to be that in the latter, if you get yourself a set of assumptions leading to contradictory conclusions (example: in Newcomb's on one hand it can be concluded that agents which 1 box walk out with more money, on the other hand , agents that choose to two-box get strictly more money than those that 1-box), it is gene

-2incogn13y

The values of A, C and P are all equivalent. You insist on making CDT determine C in a model where it does not know these are correlated. This is a problem with your model.

0nshepperd13y

Yes. That's basically the definition of CDT. That's also why CDT is no good. You can quibble about the word but in "the literature", 'CDT' means just that.

-2incogn13y

This only shows that the model is no good, because the model does not respect the assumptions of the decision theory.

-2private_messaging13y

Well, a practically important example is a deterministic agent which is copied and then copies play prisoner's dilemma against each other. There you have agents that use physics. Those, when evaluating hypothetical choices, use some model of physics, where an agent can model itself as a copyable deterministic process which it can't directly simulate (i.e. it knows that the matter inside it's head obeys known laws of physics). In the hypothetical that it cooperates, after processing the physics, it is found that copy cooperates, in the hypothetical that it defects, it is found that copy defects. And then there's philosophers. The worse ones don't know much about causality. They presumably have some sort of ill specified oracle that we don't know how to construct, which will tell them what is a 'consequence' and what is a 'cause' , and they'll only process the 'consequences' of the choice as the 'cause'. This weird oracle tells us that other agent's choice is not a 'consequence' of the decision, so it can not be processed. It's very silly and not worth spending brain cells on.

0incogn13y

Playing prisoner's dilemma against a copy of yourself is mostly the same problem as Newcomb's. Instead of Omega's prediction being perfectly correlated with your choice, you have an identical agent whose choice will be perfectly correlated with yours - or, possibly, randomly distributed in the same manner. If you can also assume that both copies know this with certainty, then you can do the exact same analysis as for Newcomb's problem. Whether you have a prediction made by an Omega or a decision made by a copy really does not matter, as long as they both are automatically going to be the same as your own choice, by assumption in the problem statement.

-2private_messaging13y

The copy problem is well specified, though. Unlike the "predictor". I clarified more in private. The worst part about Newcomb's is that all the ex religious folks seem to substitute something they formerly knew as 'god' for predictor. The agent can also be further specified; e.g. as a finite Turing machine made of cogs and levers and tape with holes in it. The agent can't simulate itself directly, of course, but it knows some properties of itself without simulation. E.g. it knows that in the alternative that it chooses to cooperate, it's initial state was in set A - the states that result in cooperation, in the alternative that it chooses to defect, it's initial state was in set B - the states that result in defection, and that no state is in both sets.

-2linas13y

I'm with incogn on this one: either there is predictability or there is choice; one cannot have both. Incogn is right in saying that, from omega's point of view, the agent is purely deterministic, i.e. more or less equivalent to a computer program. Incogn is slightly off-the-mark in conflating determinism with predictability: a system can be deterministic, but still not predictable; this is the foundation of cryptography. Deterministic systems are either predictable or are not. Unless Newcombs problem explicitly allows the agent to be non-deterministic, but this is unclear. The only way a deterministic system becomes unpredictable is if it incorporates a source of randomness that is stronger than the ability of a given intelligence to predict. There are good reasons to believe that there exist rather simple sources of entropy that are beyond the predictive power of any fixed super-intelligence -- this is not just the foundation of cryptography, but is generically studied under the rubric of 'chaotic dynamical systems'. I suppose you also have to believe that P is not NP. Or maybe I should just mutter 'Turing Halting Problem'. (unless omega is taken to be a mythical comp-sci "oracle", in which case you've pushed decision theory into that branch of set theory that deals with cardinal numbers larger than the continuum, and I'm pretty sure you are not ready for the dragons that lie there.) If the agent incorporates such a source of non-determinism, then omega is unable to predict, and the whole paradox falls down. Either omega can predict, in which case EDT, else omega cannot predict, in which case CDT. Duhhh. I'm sort of flabbergasted, because these points seem obvious to me ... the Newcomb paradox, as given, seems poorly stated.

3ArisKatsaris13y

Think of real people making choices and you'll see it's the other way around. The carefully chosen paths are the predictable ones if you know the variables involved in the choice. To be unpredictable, you need think and choose less. Hell, the archetypical imagery of someone giving up on choice is them flipping a coin or throwing a dart with closed eyes -- in short resorting to unpredictability in order to NOT choose by themselves.

2wedrifid13y

Either your claim is false or you are using a definition of at least one of those two words that means something different to the standard usage.

1incogn13y

I do not think the standard usage is well defined, and avoiding these terms altogether is not possible, seeing as they are in the definition of the problem we are discussing. Interpretations of the words and arguments for the claim are the whole content of the ancestor post. Maybe you should start there instead of quoting snippets out of context and linking unrelated fallacies? Perhaps, by specifically stating the better and more standard interpretations?

0linas13y

Huh? Can you explain? Normally, one states that a mechanical device is "predicatable": given its current state and some effort, one can discover its future state. Machines don't have the ability to choose. Normally, "choice" is something that only a system possessing free will can have. Is that not the case? Is there some other "standard usage"? Sorry, I'm a newbie here, I honestly don't know more about this subject, other than what i can deduce by my own wits.

0ArisKatsaris13y

Machines don't have preferences, by which I mean they have no conscious self-awareness of a preferred state of the world -- they can nonetheless execute "if, then, else" instructions. That such instructions do not follow their preferences (as they lack such) can perhaps be considered sufficient reason to say that machines don't have the ability to choose -- that they're deterministic doesn't... "Determining something" and "Choosing something" are synonyms, not opposites after all.

2scav13y

Newcomb's problem makes the stronger precondition that the agent is both predictable and that in fact one action has been predicted. In that specific situation, it would be hard to argue against that one action being determined and immutable, even if in general there is debate about the relationship between determinism and predictability.

0linas13y

Hmm, the FAQ, as currently worded, does not state this. It simply implies that the agent is human, that omega has made 1000 correct predictions, and that omega has billions of sensors and a computer the size of the moon. That's large, but finite. One may assign some finite complexity to Omega -- say 100 bits per atom times the number of atoms in the moon, whatever. I believe that one may devise pseudo-random number generators that can defy this kind of compute power. The relevant point here is that Omega, while powerful, is still not "God" (infinite, infallible, all-seeing), nor is it an "oracle" (in the computer-science definition of an "oracle": viz a machine that can decide undecidable computational problems).

-1incogn13y

I do not want to make estimates on how and with what accuracy Omega can predict. There is not nearly enough context available for this. Wikipedia's version has no detail whatsoever on the nature of Omega. There seems to be enough discussion to be had, even with the perhaps impossible assumption that Omega can predict perfectly, always, and that this can be known by the subject with absolute certainty.

-1incogn13y

I think I agree, by and large, despite the length of this post. Whether choice and predictability are mutually exclusive depends on what choice is supposed to mean. The word is not exactly well defined in this context. In some sense, if variable > threshold then A, else B is a choice. I am not sure where you think I am conflating. As far as I can see, perfect prediction is obviously impossible unless the system in question is deterministic. On the other hand, determinism does not guarantee that perfect prediction is practical or feasible. The computational complexity might be arbitrarily large, even if you have complete knowledge of an algorithm and its input. I can not really see the relevance to my above post. Finally, I am myself confused as to why you want two different decision theories (CDT and EDT) instead of two different models for the two different problems conflated into the single identifier Newcomb's paradox. If you assume a perfect predictor, and thus full correlation between prediction and choice, then you have to make sure your model actually reflects that. Let's start out with a simple matrix, P/C/1/2 are shorthands for prediction, choice, one-box, two-box. * P1 C1: 1000 * P1 C2: 1001 * P2 C1: 0 * P2 C2: 1 If the value of P is unknown, but independent of C: Dominance principle, C=2, entirely straightforward CDT. If, however, the value of P is completely correlated with C, then the matrix above is misleading, P and C can not be different and are really only a single variable, which should be wrapped in a single identifier. The matrix you are actually applying CDT to is the following one: * (P&C)1: 1000 * (P&C)2: 1 The best choice is (P&C)=1, again by straightforward CDT. The only failure of CDT is that it gives different, correct solutions to different, problems with a properly defined correlation of prediction and choice. The only advantage of EDT is that it is easier to cheat in this information without noticing it - even when it wo

0linas13y

Yes. I was confused, and perhaps added to the confusion.

-2MugaSofer13y

If Omega cannot predict, TDT will two-box.

0incogn13y

Thanks for the link. I like how he just brute forces the problem with (simple) mathematics, but I am not sure if it is a good thing to deal with a paradox without properly investigating why it seems to be a paradox in the first place. It is sort of like saying that this super convincing card trick you have seen, there is actually no real magic involved without taking time to address what seems to require magic and how it is done mundanely.

0owencb12y

I think this is a very clear account of the issues with these problems. I like your explanations of how correct model choice leads to CDT getting it right all the time; similarly it seems correct model choice should let EDT get it right all the time. In this light CDT and EDT are really heuristics for how to make decisions with simplified models.

[-]Carwajalca13y120

Thanks for your post, it was a good summary of decision theory basics. Some corrections:

In the Allais paradox, choice (2A) should be "A 34% chance of 24,000$ and a 66% chance of nothing" (now 27,000$).

A typo in title 10.3.1., the title should probably be "Why should degrees of belief follow the laws of probability?".

In 11.1.10. Prisoner's dilemma, the Resnik quotation mentions a twenty-five year term, yet the decision matrix has "20 years in jail" as an outcome.

0[anonymous]13y

Also, Shouldn't independence have people who prefer (1A) to (1B) prefer (2A) to (2B)?

0crazy8813y

Thanks. Fixed for the next update of the FAQ.

2pinyaka13y

Also, Shouldn't independence have people who prefer (1A) to (1B) prefer (2A) to (2B)? EDIT: Either the word "because" or "and" is out of place here. I only notice these things because this FAQ is great and I'm trying to understand every detail that I can.

0crazy8813y

Thanks Pinyaka, changed for next edit (and glad to hear you're finding it useful).

0james_edwards12y

Typo at 11.4:

[-]loup-vaillant13y110

Easy explanation for the Ellsberg Paradox: We humans treat the urn as if it was subjected to two kinds of uncertainties.

The first kind is which ball I will actually draw. It feels "truly random".
The second kind is how many red (and blue) balls there actually are. This one is not truly random.

Somehow, we prefer to chose the "truly random" option. I think I can sense why: when it's "truly random", I know no potentially hostile agent messed up with me. I mean, I could chose "red" in situation A, but then the organizers could have put 60 blue balls just to mess with me!

Put it simply, choosing "red" opens me up for external sentient influence, and therefore risk being outsmarted. This particular risk aversion sounds like a pretty sound heuristic.

0linas13y

Yes, exactly, and in our modern marketing-driven culture, one almost expects to be gamed by salesmen or sneaky game-show hosts. In this culture, its a prudent, even 'rational' response.

What about mentioning the St. Petersburg paradox? This is a pretty striking issue for EUM, IMHO.

0lukeprog13y

Added. See here.

0Leon13y

Thanks Luke.

0Dan_Moore13y

I concur. Plus, the St. Petersburg paradox was the impetus for Daniel Bernoulli's invention of the concept of utility.

-2ygert13y

The St Petersburg paradox actually sounds to me a lot like Pascal's Mugging. That is, you are offered a very small chance at a very large amount of utility, (or in the case of Pascal Mugging, of not loosing a large amount of utility), with a very high expected value if you accept the deal, but because the deal has such a low chance of paying out, a smart person will turn it down, despite that having less expected value than accepting.

I'm finding the "counterfactual mugging" challenging. At this point, the rules of the game seem to be "design a thoughtless, inert, unthinking algorithm, such as CDT or EDT or BT or TDT, which will always give the winning answer." Fine. But for the entire range of Newcomb's problems, we are pitting this dumb-as-a-rock algo against a super-intelligence. By the time we get to the counterfactual mugging, we seem to have a scenario where omega is saying "I will reward you only if you are a trusting rube who can be fleeced." N... (read more)

[-]Sniffnoy13y40

VNM utility isn't any of the types you listed. Ratios (a-b)/|c-d| of VNM utilities aren't meaningful, only ratios (a-b)/|c-b|.

[This comment is no longer endorsed by its author]Reply

2crazy8813y

I think I'm missing the point of what you're saying here so I was hoping that if I explained why I don't understand, perhaps you could clarify. VNM-utility is unique up to a positive linear transformation. When a utility function is unique up to a positive linear transformation, it is an interval (/cardinal scale). So VNM-utility is an interval scale. This is the standard story about VNM-utility (which is to say, I'm not claiming this because it seems right to me but rather because this is the accepted mainstream view of VNM-utility). Given that this is a simple mathematical property, I presume the mainstream view will be correct. So if your comment is correct in terms of the presentation in the FAQ then either we've failed to correctly define VNM-utility or we've failed to correctly define interval scales in accordance with the mainstream way of doing so (or, I've missed something). Are you able to pinpoint which of these you think has happened? One final comment. I don't see why ratios (a-b)/|c-d| aren't meaningful. For these to be meaningful, it seems to me that it would need to be that [(La+k)-(Lb+k)]/[(Lc+k)-(Ld+k)] = (a-b)/(c-d) for all L and K (as VNM-utilities are unique up to a positive linear transformation) and it seems clear enough that this will be the case: [(La+k)-(Lb+k)]/[(Lc+k)-(Ld+k)] = [L(a-b)]/[L(c-d)] = (a-b)/(c-d) Again, could you clarify what I'm missing (I'm weaker on axiomatizations of decision theory than I am on other aspects of decision theory and you're a mathematician so I'm perfectly willing to accept that I'm missing something but it'd be great if you could explain what it is)?

0Sniffnoy13y

Oops, you are absolutely right. (a-b)/|c-d| is meaningful after all. Not sure why I failed to notice that. Thanks for pointing that out.

0crazy8813y

Cool, thanks for letting me know.

[-]toony soprano7y30

I would *really* appreciate any help from lesswrong readers in helping me understand something really basic about the standard money pump argument for transitivity of preferences.

So clearly there can be situations, like in a game of Rock Scissors Paper (or games featuring non-transitive dice, like 'Efron's dice') where faced with pairwise choices it seems rational to have non-transitive preferences. And it could be that these non-transitive games/situations pay out money (or utility or whatever) if you make the right choice.

But so then if ... (read more)

7cousin_it7y

Rock paper scissors isn't an example of nontransitive preferences. Consider Alice playing the game against Bob. It is not the case that Alice prefers playing rock to playing scissors, and playing scissors to playing paper, and playing paper to playing rock. Why on Earth would she have preferences like that? Instead, she prefers to choose among rock, paper and scissors with certain probabilities that maximize her chance of winning against Bob.

3toony soprano7y

Yes I phrased my point totally badly and unclearly. Forget Rock Scissors paper - suppose team A loses to team B, B loses to C and C loses to A. Now you have the choice to bet on team A or team B to win/lose $1 - you choose B. Then you have the choice between B and C - you choose C. Then you have the choice between C and A - you choose A. And so on. Here I might pay anything less than $1 in order to choose my preferred option each time. If we just look at what I am prepared to pay in order to make my pairwise choices then it seems I have become a money pump. But of course once we factor in my winning $1 each time then I am being perfectly sensible. So my question is just – how come this totally obvious point is not a counter-example to the money pump argument that preferences ought always to be transitive? For there seem to be situations where having cyclical preferences can pay out?

4Vladimir_Nesov7y

These are decisions in different situations. Transitivity of preference is about a single situation. There should be three possible actions A, B and C that can be performed in a single situation, with B preferred to A and C preferred to B. Transitivity of preference says that C is then preferred to A in that same situation. Betting on a fight of B vs. A is not a situation where you could also bet on C, and would prefer to bet on C over betting on B.

1toony soprano7y

Also - if we have a set of 3 non-transitive dice, and I just want to roll the highest number possible, then I can prefer A to B, B to C and C to A, where all 3 dice are available to roll in the same situation. If I get paid depending on how high a number I roll, then this would seem to prevent me from becoming a money pump over the long term.

1toony soprano7y

Thanks very much for your reply Vladimir. But are you sure that is correct? I have never seen that kind of restriction to a single choice-situation mentioned before when transitivity is presented. E.g. there is nothing like that, as far as I can see, in Peterson's Decision theory textbook, nor in Bonano's presentation of transitivity in his online Textbook 'Decision Making'. All the statements of transitivity I have read just require that if a is preferred to b in a pairwise comparison, and b is preferred to c in a pairwise comparison, then a is also preferred to c in a pairwise comparison. There is no further clause requiring that a, b, and c are all simultaneously available in a single situation.

Presentation of Newcomb's problem in section 11.1.1. seems faulty. What if the human flips a coin to determine whether to one-box or two-box? (or any suitable source of entropy that is beyond the predictive powers of the super-intelligence.) What happens then?

This point is danced around in the next section, but never stated outright: EDT provides exactly the right answer if humans are fully deterministic and predictable by the superintelligence. CDT gives the right answer if the human employs an unpredictable entropy source in their decision-making. It is the entropy source that makes the decision acausal from the acts of the super-intelligence.

2wedrifid13y

If the FAQ left this out then it is indeed faulty. It should either specify that if Omega predicts the human will use that kind of entropy then it gets a "Fuck you" (gets nothing in the big box, or worse) or, at best, that Omega awards that kind of randomization with a proportional payoff (ie. If behavior is determined by a fair coin then the big box contains half the money.) This is a fairly typical (even "Frequent") question so needs to be included in the problem specification. But it can just be considered a minor technical detail.

0patrickscottshields13y

This response challenges my intuition, and I would love to learn more about how the problem formulation is altered to address the apparent inconsistency in the case that players make choices on the basis of a fair coin flip. See my other post.

-1MugaSofer13y

Or that Omega is smart enough to predict any randomizer you have available.

3linas13y

The FAQ states that omega has/is a computer the size of the moon -- that's huge but finite. I believe its possible, with today's technology, to create a randomizer that an omega of this size cannot predict. However smart omega is, one can always create a randomizer that omega cannot break.

-2MugaSofer13y

True, but just because such a randomizer is theoretically possible doesn't mean you have one to hand.

-1linas13y

OK, but this can't be a "minor detail", its rather central to the nature of the problem. The back-n-forth with incogn above tries to deal with this. Put simply, either omega is able to predict, in which case EDT is right, or omega is not able to predict, in which case CDT is right. The source of entropy need not be a fair coin: even fully deterministic systems can have a behavior so complex that predictability is untenable. Either omega can predict, and knows it can predict, or omega cannot predict, and knows that it cannot predict. The possibility that it cannot predict, yet is erroneously convinced that it can, seems ridiculous.

[-]diegocaleiro13y30

Small correction, Arntzenius name has a Z (that paper is great by the way, I sent it to Yudkwosky a while ago).

There is a compliment true of both this post and of that paper, they are both very well condensed. Congratulations Luke and crazy88!

0crazy8813y

Thanks. Will be fixed in next update. Thanks also for the positive comment.

[-]AlexMennen13y30

In the VNM system, utility is defined via preferences over acts rather than preferences over outcomes. To many, it seems odd to define utility with respect to preferences over risky acts. After all, even an agent who thinks she lives in a world where every act is certain to result in a known outcome could have preferences for some outcomes over others. Many would argue that utility should be defined in relation to preferences over outcomes or world-states, and that's not what the VNM system does. (Also see section 9.)

It's misleading to associate acts wi... (read more)

0crazy8813y

My understanding is that in the VNM system, utility is defined over lotteries. Is this the point you're contesting or are you happy with that but unhappy with the use of the word "acts" to describe these lotteries. In other words, do you think the portrayal of the VNM system as involving preferences over lotteries is wrong or do you think that this is right but the way we describe it conflates two notions that should remain distinct.

1AlexMennen13y

The problem is with the word "acts". Some lotteries might not be achievable by any act, so this phrasing makes it sound like the VNM only applies to the subset of lotteries that is actually possible to achieve. And I realize that you're using the word "act" more specifically than this, but typically, people consider doing the same thing in a different context to be the same "act", even though its consequences may depend on the context. So when I first read the paragraph I quoted after only skimming the rest, it sounded like it was claiming that the VNM system can only describe deontological preferences over actions that don't take context into account, which is, of course, ridiculous. Also, while it is true that the VNM system defines utility over lotteries, it is fairly trivial to modify it to use utility over outcomes (see first section of this post)

0crazy8813y

Thanks for the clarification. Perhaps worth noting that earlier in the document we defined acts as functions from world states to outcomes so this seems to resolve the second concern somewhat (if the context is different then presumably this is represented by the world states being different and so there will be different functions in play and hence different acts). In terms of the first concern, while VNM may define preferences over all lotteries, there's a sense where in any specific decision scenario, VNM is only appealed to in order to rank the achievable lotteries and not all of them. Of course, however, it's important to note as well that this is only part of the story. Anyway, I changed this for the next update so as to improve clarity.

2AlexMennen13y

What? That's what I thought "acts" meant the first time, before I read the document more thoroughly and decided that you must mean that acts are lotteries. If you are using "act" to refer to functions from world states to outcomes, then the statement that the VNM system only applies to acts is simply false, rather than misleading.

0crazy8813y

Okay, so I've been reading over Peterson's book An Introduction to Decision Theory and he uses much the same language as that used in the FAQ with one difference: he's careful to talk about risky acts rather than just acts (when he talks about VNM, I mean, he does simply talk about acts at some other point). This seems to be a pretty common way of talking about it (people other than Peterson use this language). Anyway, Peterson explicitly defines a "lottery" as an act (which he defines as a function from world states to outcomes) whose outcome is risky (which is to say, is determined randomly but with known probability) [I presume by the act's outcome he means the outcome that will actually occur if that act is selected]. Would including something more explicit like this resolve your concerns or do you think that Peterson does things wrong as well (or do you think I'm misunderstanding what Peterson is doing)?

0AlexMennen13y

Either Peterson does things wrong, you're misunderstanding Peterson, or I'm misunderstanding you. When I have time, I'll look at that book to try to figure out which, unless you manage to sort things out for me before I get to it.

2crazy8813y

Some quotes might help. Peterson defines an act "as a function from a set of states to a set of outcomes" The rest of the details are contained in this quote: "The key idea in von Neumann and Morgenstern's theory is to ask the decision maker to state a set of preferences over risky acts. These acts are called lotteries, because the outcome of each act is assumed to be randomly determined by events (with known probabilities) that cannot be controlled by the decision maker". The terminology of risky acts is more widespread than Peterson: http://staff.science.uva.nl/~stephane/Teaching/UncDec/vNM.pdf ---------------------------------------- However, I don't particularly see the need to get caught up in the details of what some particular people said: mostly I just want a clear way of saying what needs to be said. Perhaps the best thing to do is (a) be more explicit about what lotteries are in the VNM system; and (b) be less explicit about how lotteries and acts interact. Use of the more neutral word "options" might help here [where options are the things the agent is choosing between]. Specifically, I could explicitly note that lotteries are the options on the VNM account (which is not to say that all lotteries are options but rather that all options are lotteries on this account), outline everything in terms of lotteries and then, when talking about the issue of action guidance, I can note that VNM, at least in the standard formulation, requires that an agent already has preferences over options and note that this might seem undesirable.

0AlexMennen13y

If I understand correctly, Peterson is defining "acts" and "risky acts" as completely separate things (functions from states to outcomes, and lotteries over outcomes, respectively). If that's true, it clears up the confusion, but that seems like extraordinarily bad terminology.

0crazy8813y

Okay, well I've rewritten this for the next update in a way that hopefully resolves the issues. If you have time, once the update is posted I'd love to know whether you think the rewrite is successful. In any case, thanks for taking the time to comment so far.

0crazy8813y

Point conceded (both your point and shminux's). Edited for the next update.

2Shmi13y

I could not find a definition of "world state" in the document. All you say is which is by no means a good definition. It tells you what a state is not, but not what it is. It even fails at that, given that it uses the term "part of the world" without it being previously defined.

0crazy8813y

My understanding is that in the VNM system, utility is defined over lotteries. Is this the point you're contesting or are you happy with that but unhappy with the use of the word "acts" to describe these lotteries. In other words, do you think the portrayal of the VNM system as involving preferences over lotteries is wrong or do you think that this is right but the way we describe it conflates two notions that should remain distinct.

[-]JMiller13y30

Awesome, I look forward to reading this. Thanks, Luke and crazy88!

[-]fortyeridania13y30

Typo in section 2: "attached" should read "attacked."

0crazy8813y

Thanks, fixed for the next update.

[-]Tapatakt1y21

In this case, even if an extremely low value is set for L, it seems that paying this amount to play the game is unreasonable. After all, as Peterson notes, about nine times out of ten an agent that plays this game will win no more than 8 · 10^-100 utility.

It seems there's an error here. Should it be "In this case, even if an extremely high value is set for L, it seems that paying a lot to play the game is unreasonable."?

[-]Adam Zerner9y20

Typo:

Usually, it is argued that each of the axioms are pragmatically justified because an agent which violates the axioms can face situations in which they are guaranteed end up worse off (from their own perspective).

Should read:

guaranteed to end up worse off

Does the horizontal axis of the decision tree in section 3 represent time? If so, I'd advocate smearing those red triangles out over the whole history of actions and events. Even though, in the particular example, it's unlikely that the agent cares about having been insured as such, apart from the monetary payoffs, in the general case agents care about the whole history. I think that forgetting this point sometimes leads to misapplications of decision theory.

0crazy8813y

Yes and no. Yes, because presumably the agent's end result re: house and money occurs after the fire and the fire will happen after the decision to take out insurance (otherwise, there's not much point taking out insurance). No, because the diagram isn't really about time, even if there is an accidental temporal component to it. Instead, the levels of the diagram correspond to different factors of the decision scenario: the first level is about the agent's choice, the second level about the states of natures and the third about the final outcome. Given that this is how the diagram works, smearing out the triangles would mix up the levels and damage the clarity of the diagram. To model an agent as caring about whether they were insured or not, we would simply modify the text next to the triangles to something like "Insurance, no house and $99,900", "Insurance, house and - $100" and so on (and then we would assign different utilities to the agent based partly on whether they were insured or not as well as whether they had a house and how much money they had). I agree, though I think that talking of utility rather than money solves many of these problems. After all, utility should already take into account an agents desire to be insured etc and so talk about utility should be less likely to fall into these traps (which isn't to say there are never any problems)

[-]max_i_m13y20

When reading about Transparent Newcomb's problem: Isn't this perfectly general? Suppose Omega says: I give everyone who subscribes to decision theory A $1000, and give those who subscribe to other decision theories nothing. Clearly everyone who subscribes to decision theory A "wins".

It seems that if one lives in the world with many such Omegas, and subscribing to decision theory A (vs subscribing to decision theory B) would otherwise lead to losing at most, say, $100 per day between two successive encounters with such Omegas, then one would wi... (read more)

-4linas13y

How many times in a row will you be mugged, before you realize that omega was lying to you?

1ArisKatsaris13y

Really you probably need start imagining Omega as a trustworthy process, e.g. a mathematical proof that tells you 'X'-- thinking it as a person seems to trip you up if you are constantly bringing up the possibility it's lying when it says 'X'...

1MugaSofer13y

Omega is, by definition, always truthful.

[-]Carwajalca13y20

Maybe worth noting that there's recommended reading on decision theory on the "Best textbooks on every subject" post.

On decision theory, lukeprog recommends Peterson's An Introduction to Decision Theory over Resnik's Choices and Luce & Raiffa's Games and Decisions.

2lukeprog13y

Two of those are the books I recommend at the end of the first section.

[-]Tapatakt1y10

In this equation, V(A & O) represents the value to the agent of the combination of an act and an outcome. So this is the utility that the agent will receive if they carry out a certain act and a certain outcome occurs. Further, Pr_AO represents the probability of each outcome occurring on the supposition that the agent carries out a certain act. It is in terms of this probability that CDT and EDT differ. EDT uses the conditional probability, Pr(O|A), while CDT uses the probability of subjunctive conditionals, Pr(A O).

Please, don't use the same letters ... (read more)

[-]Tapatakt1y10

The second problem with using the law of large numbers to justify EUM has to do with a mathematical theorem known as gambler's ruin. Imagine that you and I flip a fair coin, and I pay you $1 every time it comes up heads and you pay me $1 every time it comes up tails. We both start with $100. If we flip the coin enough times, one of us will face a situation in which the sequence of heads or tails is longer than we can afford. If a long-enough sequence of heads comes up, I'll run out of $1 bills with which to pay you. If a long-enough sequence of tails comes

... (read more)

[-]Greg Walsh3y10

I think the example given to show the irrationality of leximin in certain situations doesn’t do a good job of distinguishing its failings from maximin. To usefully illustrate the difference between the two I believe a another state is required with even worse outcomes for both acts (e.g. $0). This way the worst outcomes for both acts would be equal and so the second worst outcomes (a1:$1, a2:$1.01) would then be compared under the leximin strategy leading to the choice of a2 as the best act again with the acknowledgment that you miss out on the opportunity to get $10,001.01

[-]Joseph Greenwood10y10

In the last chapter of his book "Utility Theory for Decision Making," Peter Fishburn published a concise rendering of Leonard Savage's proof that "rational" preferences over events implied that one behaved "as if" he (or she) was obeying Expected Utility Theory. He furthermore proved that following Savage's axioms implied that your utility function is bounded (he attributes this extension of the proof, in its essence, to Savage). So Subjective Expected Utility Theory has an answer to the St. Petersburg Paradox "built in" to its axioms. That seems like a point well worth mentioning in this article.

[-]Pentashagon13y10

The image of Ellsberg's Paradox has the picture of the Yellow/Blue bet replaced with a picture of a Yellow/Red bet. Having looked at the picture I was about to claim that it was always rational to take the R/B bet over Y/R before I read the actual description.

2crazy8813y

Will be fixed in the next update. Thanks for pointing it out.

[-]kilobug13y10

Isn't there a typo in "Experiments have shown that many people prefer (1A) to (1B) and (2B) to (2A)." ? Shouldn't it be "(2A) to (2B)" ?

Edit : hrm, no, in fact it's like http://lesswrong.com/lw/gu1/decision_theory_faq/8jav said : it should be 24 000$ instead of 27 000$ in option A, or else it makes no sense.

0crazy8813y

Thanks, as you note, the linked comment is right.

[-]somervta13y10

Thus, the expected utility (EU) of choice A is, for this decision maker, (1)(1000) = 1000. Meanwhile, the EU of choice B is (0.5)(1500) + (0.5)(0) = 750. In this case, the expected utility of choice B is greater than that of choice A, even though choice B has a greater expected monetary value.

Choice A at 1000 is still greater than Choice B at 750

0crazy8813y

Thanks, will be fixed in next update.

[-]Vaniver13y10

It's Stiennon, not Steinnon.

0crazy8813y

Fixed for next update. Thanks.

Minor error: In the prisoner's dilemma example, the decision matrix has twenty years for if you cooperate and your partner defects, while the text quoted right above the matrix claims that that amount is twenty five years.

1crazy8813y

Thanks. I've fixed this up in the next update (though it won't appear on the LW version yet).

[-]Adam Zerner9y00

I find it helpful to use the term "security level" to understand maximin/leximin and "hope level" to understand maximax. "Security level" is the worst case scenario, and under maximin/leximin we want to maximize it. "Hope level" is the best case scenario, and under maximax, we want to maximize it.

[-]alexander_poddiakov10y00

Concerning the transitivity axiom, what about rational choices in situations of intransitivity cycles?

[-][anonymous]11y00

Bossert link does not work: ( (Sorry. Edited to make sense.)

[-]Gram_Stone11y00

(Well, sort of. The minimax and maximax principles require only that we measure value on an ordinal scale, whereas the optimism-pessimism rule requires that we measure value on an interval scale.)

I'm using this as an introduction to decision theory so I might be wrong, and I've read that 'maximin' and 'minimax' do have different meanings in game theory, but you exclusively use the term 'maximin' up to a certain point and then mention a 'minimax principle' once, so I can only imagine that you meant to write 'maximin principle.' It confused me. It's proba... (read more)

1Greg Walsh3y

Minimax is equivalent to maximin in a zero sum game because every resource I win you lose out on. However, in non-zero games which encourage cooperation this is no longer case. Take for example a scenario in which we share some land. Let’s assume I have one third and you have two. If I water your crops as well as my own every other day you have agreed to water mine on the other alternate days when I must look after my kids. If I don’t water your crops you won’t water mine back but you don’t have any kids so you’d actually be able to water every day without me although you’d rather hang out with your friends at the pool if you could. If I want to minimise the maximum (minimax) utility you get I wouldn’t water your crops at all and force you to water every day. Of course you’d resent me for this and wouldn’t water my crops so I’d end up with no crop. If I want to maximise the minimum (maximin) crop yield for myself, I’ll have to water your crops too even though you have twice as many crops as me and you’ll get to hang out with your friends as well!

[-]AlexMennen12y00

Another objection to the VNM approach (and to expected utility approaches generally), the St. Petersburg paradox, draws on the possibility of infinite utilities. The St. Petersburg paradox is based around a game where a fair coin is tossed until it lands heads up. At this point, the agent receives a prize worth 2n utility, where n is equal to the number of times the coin was tossed during the game. The so-called paradox occurs because the expected utility of choosing to play this game is infinite and so, according to a standard expected utility approach,

... (read more)

0lukeprog12y

Thanks! I've edited the article. What do you think of my edit?

0AlexMennen12y

That works. I don't think removing the "like 1" helps much. This phrasing leaves it unclear what "extremely low value" means, and I suspect most people who would object to maximizing expected utility when L=1 would still think it is reasonable when L=10^-99, which seems like a more reasonable interpretation of "extremely low value" when numbers like 10^-100 are mentioned.

[-]Stuart_Armstrong12y00

Thanks, a very useful overview.

[-]adam_strandberg12y00

In section 8.1, your example of the gambler's ruin postulates that both agents have the same starting resources, but this is exactly the case in which the gambler's ruin doesn't apply. That might be worth changing.

[-]buybuydandavis13y00

Can decisions under ignorance be transformed into decisions under uncertainty?

I'd add a comment on Jaynes' solution for determining ignorance priors in terms of transformation groups.

I'd say that there's no such think as an "ignorance" prior - priors are set by information. Setting a prior by symmetry or the more general transformation group is an assertion of information.

There are numerous typos throughout the thing. Someone needs to re-read it. The math in "8.6.3. The Allais paradox" is all wrong, option 2A is not actually 34% of 1A and 66% of nothing, etc.

[-][anonymous]13y00

This may not be the best place for this question, but it's something I've been wondering for a while: how does causal decision theory fail us humans in the real world, here and now?

4wedrifid13y

Us humans almost never use Causal Decision Theory in the real world, here and now. As such it fails us very little. What humans actually tend to use is about as similar to TDT as it is to CDT (ie. actual practice diverges from each of those ideals in different ways and couldn't be said to be doing either.)

0[anonymous]13y

All right, but how would CDT fail us, if we used it perfectly?

0wedrifid13y

If we used it perfectly it would fail us very little. The 'used perfectly' part would prompt us to create and optimize institutions to allow the limitations of CDT to be worked around. It would result in a slightly less efficient system with some extra overheads and some wasted opportunities but it would still be rather good. Specifically it would require more structures in place for creating and enforcing precomittments and certain kind of cooperation would be unavailable.

0[anonymous]13y

Ah, that's right. CDT tells you to defect in a prisoner's dilemma against someone identical to you; TDT tells you to cooperate. So TDT wins here.

0Kindly13y

CDT does, however, tell you to precommit to cooperate in a prisoner's dilemma against someone who also precommits to cooperate with you, if this is an option.

There is one rather annoying subtext that recurs throughout the FAQ: the very casual and carefree use of the words "rational" and "irrational", with the rather flawed idea that following some axiomatic system (e.g. VNM) and Bayes is "rational" and not doing so is "irrational". I think this is a dis-service, and, what's more, fails to look into the effects of intelligence, experience, training and emotion. The Allias paradox scratches the surface, as do various psych experiments. But ...

The real question is "wh... (read more)

The conclusion to section "11.1.3. Medical Newcomb problems" begs a question which remains unanswered: -- "So just as CDT “loses” on Newcomb’s problem, EDT will "lose” on Medical Newcomb problems (if the tickle defense fails) or will join CDT and "lose" on Newcomb’s Problem itself (if the tickle defense succeeds)."

If I was designing a self-driving car and had to provide an algorithm for what to do during an emergency, I may choose to hard-code CDT or EDT into the system, as seems appropriate. However, as an intelligen... (read more)

4linas13y

Hmm. I just got a -1 on this comment ... I thought I posed a reasonable question, and I would have thought it to even be a "commonly asked question", so why would it get a -1? Am I misunderstanding something, or am I being unclear?

0MugaSofer13y

Omega is, by definition, always truthful. EDIT: Sorry, thought this was in reply to a different comment.

[-]davidpearce13y-40

But note burger-choosing Jane (6.1) is still irrational - for she has discounted the much stronger preference of a cow not to be harmed. Rationality entails overcoming egocentric bias - and ethnocentric and anthropocentric bias - and adopting a God's eye point-of-view that impartially gives weight to all possible first-person perspectives.

When we say 'rationality', we mean instrumental rationality; getting what you want. Elsewhere, we also refer to epistemic rationality, which is believing true things. In neither case do we say anything about what you should want.

It might be a good thing to care about cows, but it's not rationality as we understand the word. Good you bring this up though, as I can easily imagine others being confused.

See also What Do We Mean by Rationality

4pragmatist13y

This begs the question against moral realism. If it is in fact true that harming cows is bad, then epistemic rationality demands that we believe that harming cows is bad. Of course, saying that you should believe harming cows is morally wrong is different from saying that you shouldn't choose to harm cows, but the inference from one to the other is pretty plausible. It seems fairly uncontroversial that if one believes that action X is bad, and it is in fact true that action X is bad, then one should not perform action X (ceteris paribus). I don't agree with davidpearce's framing (that rationality demands that one give equal weight to all perspectives), but I also don't agree with the claim that rationality does not tell us anything about what we should want. Perhaps instrumental rationality doesn't, but epistemic rationality does.

2Larks13y

Sure - but rationality doesn't itself imply moral conclusions. It only says that if there are moral facts then we should believe them, not that there are any (particular) moral facts, which is what David needs.

4davidpearce13y

Larks, no, pragmatist nicely captured the point I was making. If we are trying to set out what an "ideal, perfectly rational agent" would choose, then we can't assume that such a perfectly rational agent would arbitrarily disregard a stronger preference in favour of a weaker preference. Today, asymmetries of epistemic access mean that weaker preferences often trump stronger preferences; but with tomorrow's technology, this cognitive limitation on our decision-making procedures can be overcome.

[-]Eliezer Yudkowsky13y170

Isn't the giant elephant in this room the whole issue of moral realism? I'm a moral cognitivist but not a moral realist. I have laid out what it means for my moral beliefs to be true - the combination of physical fact and logical function against which my moral judgments are being compared. This gives my moral beliefs truth value. And having laid this out, it becomes perfectly obvious that it's possible to build powerful optimizers who are not motivated by what I call moral truths; they are maximizing something other than morality, like paperclips. They will also meta-maximize something other than morality if you ask them to choose between possible utility functions, and will quite predictably go on picking the utility function "maximize paperclips". Just as I correctly know it is better to be moral than to be paperclippy, they accurately evaluate that it is more paperclippy to maximize paperclips than morality. They know damn well that they're making you unhappy and violating your strong preferences by doing so. It's just that all this talk about the preferences that feel so intrinsically motivating to you, is itself of no interest to them because you haven't got... (read more)

7Rob Bensinger13y

Even among philosophers, "moral realism" is a term wont to confuse. I'd be wary about relying on it to chunk your philosophy. For instance, the simplest and least problematic definition of 'moral realism' is probably the doctrine... minimal moral realism: cognitivism (moral assertions like 'murder is bad' have truth-conditions, express real beliefs, predicate properties of objects, etc.) + success theory (some moral assertions are true; i.e., rejection of error theory). This seems to be the definition endorsed on SEP's Moral Realism article. But it can't be what you have in mind, since you accept cognitivism and reject error theory. So perhaps you mean to reject a slightly stronger claim (to coin a term): factual moral realism: MMR + moral assertions are not true or false purely by stipulation (or 'by definition'); rather, their truth-conditions at least partly involve empirical, worldly contingencies. But here, again, it's hard to find room to reject moral realism. Perhaps some moral statements, like 'suffering is bad,' are true only by stipulation; but if 'punching people in the face causes suffering' is not also true by stipulation, then the conclusion 'punching people in the face is bad' will not be purely stipulative. Similarly, 'The Earth's equatorial circumference is ~40,075.017 km' is not true just by definition, even though we need somewhat arbitrary definitions and measurement standards to assert it. And rejecting the next doesn't sound right either: correspondence moral realism: FMR + moral assertions are not true or false purely because of subjects' beliefs about the moral truth. For example, the truth-condition for 'eating babies is bad' are not 'Eliezer Yudkowsky thinks eating babies is bad', nor even 'everyone thinks eating babies is bad'. Our opinions do play a role in what's right and wrong, but they don't do all the work. So perhaps one of the following is closer to what you mean to deny: moral transexperientialism: Moral facts are nontrivia

[-]Eliezer Yudkowsky13y210

I'm not sure this taxonomy is helpful from David Pearce's perspective. David Pearce's position is that there are universally motivating facts - facts whose truth, once known, is compelling for every possible sort of mind. This reifies his observation that the desire for happiness feels really, actually compelling to him and this compellingness seems innate to qualia, so anyone who truly knew the facts about the quale would also know that compelling sense and act accordingly. This may not correspond exactly to what SEP says under moral realism and let me know if there's a standard term, but realism seems to describe the Pearcean (or Eliezer circa 1996) feeling about the subject - that happiness is really intrinsically preferable, that this is truth and not opinion.

From my perspective this is a confusion which I claim to fully and exactly understand, which licenses my definite rejection of the hypothesis. (The dawning of this understanding did in fact cause my definite rejection of the hypothesis in 2003.) The inherent-desirableness of happiness is your mind reifying the internal data describing its motivation to do something, so if you try to use your empathy to imagine another... (read more)

8Rob Bensinger13y

I'm not sure he's wrong in saying that feeling the qualia of a sentient, as opposed to modeling those qualia in an affective black box without letting the feels 'leak' into the rest of your cognitionspace, requires some motivational effect. There are two basic questions here: First, the Affect-Effect Question: To what extent are the character of subjective experiences like joy and suffering intrinsic or internal to the state, as opposed to constitutively bound up in functional relations that include behavioral impetuses? (For example, to what extent is it possible to undergo the phenomenology of anguish without thereby wanting the anguish to stop? And to what extent is it possible to want something to stop without being behaviorally moved, to the extent one is able and to the extent one's other desires are inadequate overriders, to stop it?) Compare David Lewis' 'Mad Pain', pain that has the same experiential character as ordinary pain but none of its functional relations (or at least not the large-scale ones). Some people think a state of that sort wouldn't qualify as 'pain' at all, and this sort of relationalism lends some credibility to pearce's view. Second, the Third-Person Qualia Question: To what extent is phenomenological modeling (modeling a state in such a way that you, or a proper part of you, experiences that state) required for complete factual knowledge of real-world agents? One could grant that qualia are real (and really play an important role in various worldly facts, albeit perhaps physical ones) and are moreover unavoidably motivating (if you aren't motivated to avoid something, then you don't really fear it), but deny that an epistemically rational agent is required to phenomenologically model qualia. Perhaps there is some way to represent the same mental states without thereby experiencing them, to fully capture the worldly facts about cows without simulating their experiences oneself. If so, then knowing everything about cows would not requir

2davidpearce13y

Eliezer, you remark, "The inherent-desirableness of happiness is your mind reifying the internal data describing its motivation to do something," Would you propose that a mind lacking in motivation couldn't feel blissfully happy? Mainlining heroin (I am told) induces pure bliss without desire - shades of Buddhist nirvana? Pure bliss without motivation can be induced by knocking out the dopamine system and directly administering mu opioid agonists to our twin "hedonic hotspots" in the ventral pallidum and rostral shell of the nucleus accumbens. Conversely, amplifying mesolimbic dopamine function while disabling the mu opioid pathways can induce desire without pleasure. [I'm still mulling over some of your other points.]

2Eliezer Yudkowsky13y

Here we're reaching the borders of my ability to be confident about my replies, but the two answers which occur to me are: 1) It's not positive reinforcement unless feeling it makes you experience at least some preference to do it again - otherwise in what sense are the neural networks getting their plus? Heroin may not induce desire while you're on it, but the thought of the bliss induces desire to take heroin again, once you're off the heroin. 2) The superBuddhist no longer capable of experiencing desire or choice, even desire or choice over which thoughts to think, also becomes incapable of experiencing happiness (perhaps its neural networks aren't even being reinforced to make certain thoughts more likely to be repeated). However, you, who are still capable of desire and who still have positively reinforcing thoughts, might be tricked into considering the superBuddhist's experience to be analogous to your own happiness and therefore acquire a desire to be a superBuddhist as a result of imagining one - mostly on account of having been told that it was representing a similar quale on account of representing a similar internal code for an experience, without realizing that the rest of the superBuddhist's mind now lacks the context your own mind brings to interpreting that internal coding into pleasurable positive reinforcement that would make you desire to repeat that experiential state.

0JonatasMueller13y

It's a reasonably good description, though wanting and liking seem to be neurologically separate, such that liking does not necessarily reflect a motivation, nor vice-versa (see: Not for the sake of pleasure alone. Think the pleasurable but non-motivating effect of opioids such as heroin. Even in cases in which wanting and liking occur together, this does not necessarily invalidate the liking aspect as purely wanting. Liking and disliking, good and bad feelings as qualia, especially in very intense amounts, seem to be intrinsically so to those who are immediately feeling them. Reasoning could extend and generalize this.

3Eliezer Yudkowsky13y

Heh. Yes, I remember reading the section on noradrenergic vs. dopaminergic motivation in Pearce's BLTC as a 16-year-old. I used to be a Pearcean, ya know, hence the Superhappies. But that distinction didn't seem very relevant to the metaethical debate at hand.

2davidpearce13y

It's possible (I hope) to believe future life can be based on information-sensitive gradients of (super)intelligent well-being without remotely endorsing any of my idiosyncratic views on consciousness, intelligence or anything else. That's the beauty of hedonic recalibration. In principle at least, hedonic recalibration can enrich your quality of life and yet leave most if not all of your existing values and preference architecture intact .- including the belief that there are more important things in life than happiness.

4Eliezer Yudkowsky13y

Agreed. The conflict between the Superhappies and the Lord Pilot had nothing to do with different metaethical theories. Also, we totally agree on wanting future civilization to contain very smart beings who are pretty happy most of the time. We just seem to disagree about whether it's important that they be super duper happy all of the time. The main relevance metaethics has to this is that once I understood there was no built-in axis of the universe to tell me that I as a good person ought to scale my intelligence as fast as possible so that I could be as happy as possible as soon as possible, I decided that I didn't really want to be super happy all the time, the way I'd always sort of accepted as a dutiful obligation while growing up reading David Pearce. Yes, it might be possible to do this in a way that would leave as much as possible of me intact, but why do it at all if that's not what I want? There's also the important policy-relevant question of whether arbitrarily constructed AIs will make us super happy all the time or turn us into paperclips.

0Shmi13y

Huh, when I read the story, my impression was that it was Lord Pilot not understanding that it was a case of "Once you go black, you can't go back". Specifically, once you experience being superhappy, your previous metaethics stops making sense and you understand the imperative of relieving everyone of the unimaginable suffering of not being superhappy.

0JonatasMueller13y

I thought it was relevant to this, if not, then what was meant by motivation? Consciousness is that of which we can be most certain of, and I would rather think that we are living in a virtual world under an universe with other, alien physical laws, than that consciousness itself is not real. If it is not reducible to nonmental facts, then nonmental facts don't seem to account for everything there is of relevant.

4Creutzer13y

I suggest that to this array of terms, we should add moral indexicalism to designate Eliezer's position, which by the above definition would be a special form of realism. As far as I can tell, he basically says that moral terms are hidden indexicals in Putnam's sense.

3A1987dM13y

Watch out -- the word "sentient" has at least two different common meanings, one of which includes cattle and the other doesn't. EY usually uses it with the narrower meaning (for which a less ambiguous synonym is "sapient"), whereas David Pearce seems to be using it with the broader meaning.

3Rob Bensinger13y

Ah. By 'sentient' I mean something that feels, by 'sapient' something that thinks. To be more fine-grained about it, I'd define functional sentience as having affective (and perhaps perceptual) cognitive states (in a sense broad enough that it's obvious cows have them, and equally obvious tulips don't), and phenomenal sentience as having a first-person 'point of view' (though I'm an eliminativist about phenomenal consciousness, so my overtures to it above can be treated as a sort of extended thought experiment). Similarly, we might distinguish a low-level kind of sapience (the ability to form and manipulate mental representations of situations, generate expectations and generalizations, and update based on new information) from a higher-level kind closer to human sapience (perhaps involving abstract and/or hyper-productive representations à la language). Based on those definitions, I'd say it's obvious cows are functionally sentient and have low-level sapience, extremely unlikely they have high-level sapience, and unclear whether they have phenomenal sentience.

3davidpearce13y

Rob, many thanks for a thoughtful discussion above. But on one point, I'm confused. You say of cows that it's "unclear whether they have phenomenal sentience." Are you using the term "sentience" in the standard dictionary sense ["Sentience is the ability to feel, perceive, or be conscious, or to experience subjectivity": http://en.wikipedia.org/wiki/Sentience ] Or are you using the term in some revisionary sense? At least if we discount radical philosophical scepticism about other minds, cows and other nonhuman vertebrates undergo phenomenal pain, anxiety, sadness, happiness and a whole bunch of phenomenal sensory experiences. For sure, cows are barely more sapient than a human prelinguistic toddler (though see e.g. http://www.appliedanimalbehaviour.com/article/S0168-1591(03)00294-6/abstract http://www.dailymail.co.uk/news/article-2006359/Moo-dini-Cow-unusual-intelligence-opens-farm-gate-tongue-herd-escape-shed.html ] But their limited capacity for abstract reasoning is a separate issue.

1Rob Bensinger13y

Neither. I'm claiming that there's a monstrous ambiguity in all of those definitions, and I'm tabooing 'sentience' and replacing it with two clearer terms. These terms may still be problematic, but at least their problematicity is less ambiguous. I distinguished functional sapience from phenomenal sapience. Functional sapience means having all the standard behaviors and world-tracking states associated with joy, hunger, itchiness, etc. It's defined in third-person terms. Phenomenal sapience means having a subjective vantage point on the world; being sapient in that sense means that it feels some way (in a very vague sense) to be such a being, whereas it wouldn't 'feel' any way at all to be, for example, a rock. To see the distinction, imagine that we built a robot, or encountered an alien species, that could simulate the behaviors of sapients in a skillful and dynamic way, without actually having any experiences of its own. Would such a being necessarily be sapient? Does consistently crying out and withdrawing from some stimulus require that you actually be in pain, or could you be a mindless automaton? My answer is 'yes, in the functional sense; and maybe, in the phenomenal sense'. The phenomenal sense is a bit mysterious, in large part because the intuitive idea of it arises from first-person introspection and not from third-person modeling or description, hence it's difficult (perhaps impossible!) to find definitive third-person indicators of this first-person class of properties. 'Radical philosophical scepticism about other minds' I take to entail that nothing has a mind except me. In other words, you're claiming that the only way to doubt that there's something it's subjectively like to be a cow, is to also doubt that there's something it's subjectively like to be any human other than myself. I find this spectacularly implausible. Again, I'm an eliminativist, but I'll put myself in a phenomenal realist's shoes. The neural architecture shared in common by h

0Ben Pace13y

And here we see the value of replacing the symbol with the substance.

6davidpearce13y

Eliezer, in my view, we don't need to assume meta-ethical realism to recognise that it's irrational - both epistemically irrational and instrumentally irrational - arbitrarily to privilege a weak preference over a strong preference. To be sure, millions of years of selection pressure means that the weak preference is often more readily accessible. In the here-and-now, weak-minded Jane wants a burger asap. But it's irrational to confuse an epistemological limitation with a deep metaphysical truth. A precondition of rational action is understanding the world. If Jane is scientifically literate, then she'll internalise Nagel's "view from nowhere" and adopt the God's-eye-view to which natural science aspires. She'll recognise that all first-person facts are ontologically on a par - and accordingly act to satisfy the stronger preference over the weaker. So the ideal rational agent in our canonical normative decision theory will impartially choose the action with the highest expected utility - not the action with an extremely low expected utility. At the risk of labouring the obvious, the difference in hedonic tone induced by eating a hamburger and a veggieburger is minimal. By contrast, the ghastly experience of having one's throat slit is exceptionally unpleasant. Building anthropocentric bias into normative decision theory is no more rational than building geocentric bias into physics. Paperclippers? Perhaps let us consider the mechanism by which paperclips can take on supreme value. We understand, in principle at least, how to make paperclips seem intrinsically supremely valuable to biological minds - more valuable than the prospect of happiness in the abstract. [“Happiness is a very pretty thing to feel, but very dry to talk about.” - Jeremy Bentham]. Experimentally, perhaps we might use imprinting (recall Lorenz and his goslings), microelectrodes implanted in the reward and punishment centres, behavioural conditioning and ideological indoctrination - and perhaps th

[-]Eliezer Yudkowsky13y160

Eliezer, in my view, we don't need to assume meta-ethical realism to recognise that it's irrational - both epistemically irrational and instrumentally irrational - arbitrarily to privilege a weak preference over a strong preference.

You need some stage at which a fact grabs control of a mind, regardless of any other properties of its construction, and causes its motor output to have a certain value.

Paperclippers? Perhaps let us consider the mechanism by which paperclips can take on supreme value. We understand, in principle at least, how to make paperclips seem intrinsically supremely valuable to biological minds - more valuable than the prospect of happiness in the abstract. [“Happiness is a very pretty thing to feel, but very dry to talk about.” - Jeremy Bentham]. Experimentally, perhaps we might use imprinting (recall Lorenz and his goslings), microelectrodes implanted in the reward and punishment centres, behavioural conditioning and ideological indoctrination - and perhaps the promise of 72 virgins in the afterlife for the faithful paperclipper. The result: a fanatical paperclip fetishist!

As Sarokrae observes, this isn't the idea at all. We construct a paperclip maxim... (read more)

2davidpearce13y

Eliezer, thanks for clarifying. This is how I originally conceived you viewed the threat from superintelligent paperclip-maximisers, i.e. nonconscious super-optimisers. But I was thrown by your suggestion above that such a paperclipper could actually understand first-person phenomenal states, i.e, it's a hypothetical "full-spectrum" paperclipper. If a hitherto non-conscious super-optimiser somehow stumbles upon consciousness, then it has made a momentous ontological discovery about the natural world. The conceptual distinction between the conscious and nonconscious is perhaps the most fundamental I know. And if - whether by interacting with sentients or by other means - the paperclipper discovers the first-person phenomenology of the pleasure-pain axis, then how can this earth-shattering revelation leave its utility function / world-model unchanged? Anyone who is isn't profoundly disturbed by torture, for instance, or by agony so bad one would end the world to stop the horror, simply hasn't understood it. More agreeably, if such an insentient paperclip-maximiser stumbles on states of phenomenal bliss, might not clippy trade all the paperclips in the world to create more bliss, i.e revise its utility function? One of the traits of superior intelligence, after all, is a readiness to examine one's fundamental assmptions and presuppositions - and (if need be) create a novel conceptual scheme in the face of surprising or anomalous empirical evidence.

[-]shware13y180

Anyone who is isn't profoundly disturbed by torture, for instance, or by agony so bad one would end the world to stop the horror, simply hasn't understood it.

Similarly, anyone who doesn't want to maximize paperclips simply hasn't understood the ineffable appeal of paperclipping.

-2whowhowho13y

I don't see the analogy. Paperclipping doesn't have to be an ineffable value for a paperclipper, and paperclippers don't have to be motivated by anything qualia-like.

5Viliam_Bur13y

Exactly. Consequentialist paperclip maximizer does not have to feel anything in regards to paperclips. It just... maximizes their number. This is an incorrect, anthropomorphic model: Human: "Clippy, did you ever think about the beauty of joy, and the horrors of torture?" Clippy: "Human, did you ever think about the beauty of paperclips, and the horrors of their absence?" This is more correct: Human: "Clippy, did you ever think about the beauty of joy, and the horrors of torture?" Clippy: (ignores the human and continues to maximize paperclips) Or more precisely, Clippy would say "X" to the human if and only if saying "X" would maximize the number of paperclips. The value of X would be completely unrelated to any internal state of Clippy. Unless such relation does somehow contribute to maximization of the paperclips (for example if the human will predictably read Clippy's internal state, verify the validity of X, and on discovering a lie destroy Clippy, thus reducing the expected number of paperclips). In other words, if humans are a poweful force in the universe, Clippy would choose the actions which lead to maximum number of paperclips in a world with humans. If the humans are sufficiently strong and wise, Clippy could self-modify to become more human-like, so that the humans, following their utility function, would be more likely to allow Clippy produce more paperclips. But every such self-modification would be chosen to maximize the number of paperclips in the universe. Even if Clippy self-modifies into something less-than-perfectly-rational (e.g. to appease the humans), the pre-modification Cloppy would choose the modification which maximizes the expected number of paperclips within given constraints. The constraints would depend on Clippy's model of humans and their reactions. For example Clippy could choose to be more human-like (as much as is necessary to be respected by humans) with strong aversion about future modifications and strong desire to maxi

-7whowhowho13y

3shware13y

Well, yes, obviously the classical paperclipper doesn't have any qualia, but I was replying to a comment wherein it was argued that any agent on discovering the pain-of-torture qualia in another agent would revise its own utility function in order to prevent torture from happening. It seems to me that this argument proves too much in that if it were true then if I discovered an agent with paperclips-are-wonderful qualia and I "fully understood" those experiences I would likewise be compelled to create paperclips.

0hairyfigment13y

Someone might object to the assumption that "paperclips-are-wonderful qualia" can exist. Though I think we could give persuasive analogies from human experience (OCD, anyone?) so I'm upvoting this anyway.

[-]Eliezer Yudkowsky13y160

"Aargh!" he said out loud in real life. David, are you disagreeing with me here or do you honestly not understand what I'm getting at?

The whole idea is that an agent can fully understand, model, predict, manipulate, and derive all relevant facts that could affect which actions lead to how many paperclips, regarding happiness, without having a pleasure-pain architecture. I don't have a paperclipping architecture but this doesn't stop me from modeling and understanding paperclipping architectures.

The paperclipper can model and predict an agent (you) that (a) operates on a pleasure-pain architecture and (b) has a self-model consisting of introspectively opaque elements which actually contain internally coded instructions for your brain to experience or want certain things (e.g. happiness). The paperclipper can fully understand how your workspace is modeling happiness and know exactly how much you would want happiness and why you write papers about the apparent ineffability of happiness, without being happy itself or at all sympathetic toward you. It will experience no future surprise on comprehending these things, because it already knows them. It doesn't have any objec... (read more)

4Kawoomba13y

Here comes the Reasoning Inquisition! (Nobody expects the Reasoning Inquisition.) As the defendant admits, a sufficiently leveled-up paperclipper can model lower-complexity agents with a negligible margin of error. That means that we can define a subroutine within the paperclipper which is functionally isomorphic to that agent. If the agent-to-be-modelled is experiencing pain and pleasure, then by the defendent's own rejection of the likely existence of p-zombies, so must that subroutine of the paperclipper! Hence a part of the paperclipper experiences pain and pleasure. I submit that this can be used as pars pro toto, since it is no different from only a part of the human brain generating pain and pleasure, yet us commonly referring to "the human" experiencing thus. That the aforementioned feelings of pleasure and pain are not directly used to guide the (umbrella) agent's actions is of no consequence, the feeling exists nonetheless. The power of this revelation is strong, here come the tongues! tại sao bạn dịch! これは喜劇の効果にすぎず! یہ اپنے براؤزر پلگ ان کی امتحان ہے، بھی ہے.

9Eliezer Yudkowsky13y

Not necessarily. x -> 0 is input-output isomorphic to Goodstein() without being causally isomorphic. There are such things as simplifications. Quite likely. A paperclipper has no reason to avoid sentient predictive routines via a nonperson predicate; that's only an FAI desideratum.

0whowhowho13y

A subroutine, or any other simulation or model, isn't a p-zombie as usually defined, since they are physical duplicates. A sim is a functional equivalent (for some value of "equivalent") made of completely different stuff, or no particular kind of stuff.

0Kawoomba13y

I wrote a lengthy comment on just that, but scrapped it because it became rambling. An outsider could indeed tell them apart by scanning for exact structural correspondence, but that seems like cheating. Peering beyond the veil / opening Clippy's box is not allowed in a Turing test scenario, let's define some p-zombie-ish test following the same template. If it quales like a duck (etc.), it probably is sufficiently duck-like.

0whowhowho13y

I would rather maintain p-zombie in its usual meaning, and introduce a new term, eg c-zombie for Turing-indistiguishable functional duplicates.

2Sarokrae13y

So my understanding of David's view (and please correct me if I'm wrong, David, since I don't wish to misrepresent you!) is that he doesn't have paperclipping architecture and this does stop him from imagining paperclipping architectures.

4Eliezer Yudkowsky13y

...well, in point of fact he does seem to be having some trouble, but I don't think it's fundamental trouble.

-2whowhowho13y

Let's say the paperclipper reaches the point where it considers making people suffer for the sake of paperclipping. DP's point seems to be that either it fully understands suffering--in which case, it realies that inflicing suffering is wrong --or it it doesn't fully understand. He sees a conflict between superintelligence and ruthlessness -- as a moral realist/cognitivist would is that full understanding.?. ETA: Unless there is -- eg. what qualiaphiles are always banging on about; what it feels like. That the clipper can conjectures that are true by correspondence , that it can narrow down possible universes, that it can predict, are all necessary criteria for full understanding. It is not clear that they are sufficient. Clippy may be able to figure out an organisms response to pain on a basis of "stimulus A produces response B", but is that enough to tell it that pain hurts ? (We can make guesses about that sort of thing in non-human organisms, but that may be more to do with our own familiarity with pain, and less to do with acts of superintelligence). And if Clippy can't know that pain hurts, would Clippy be able to work out that Hurting People is Wrong? further edit; To put it another way, what is there to be moral about in a qualia-free universe?

[-]khafra13y100

As Kawoomba colorfully pointed out, clippy's subroutines simulating humans suffering may be fully sentient. However, unless those subroutines have privileged access to clippy's motor outputs or planning algorithms, clippy will go on acting as if he didn't care about suffering. He may even understand that inflicting suffering is morally wrong--but this will not make him avoid suffering, any more than a thrown rock with "suffering is wrong" painted on it will change direction to avoid someone's head. Moral wrongness is simply not a consideration that has the power to move a paperclip maximizer.

-5whowhowho13y

8Shmi13y

Maybe I can chime in... "understand" does not mean "empathize". Psychopaths understand very well when people experience these states but they do not empathize with them. Again, understanding is insufficient for revision. The paperclip maximizer, like a psychopath, maybe better at parsing human affect than a regular human, but it is not capable of empathy, so it will manipulate this affect for its own purposes, be it luring a victim or building paperclips. So, if one day humans discover the ultimate bliss that only creating paperclips can give, should they "create a novel conceptual scheme" of giving their all to building more paperclips, including converting themselves into metal wires? Or do we not qualify as a "superior intelligence"?

0davidpearce13y

Shminux, a counter-argument: psychopaths do suffer from a profound cognitive deficit. Like the rest of us, a psychopath experiences the egocentric illusion. Each of us seems to the be the centre of the universe. Indeed I've noticed the centre of the universe tends to follow my body-image around. But whereas the rest of us, fitfully and imperfectly, realise the egocentric illusion is a mere trick of perspective born of selfish DNA, the psychopath demonstrates no such understanding. So in this sense, he is deluded. [We're treating psychopathy as categorical rather than dimensional here. This is probably a mistake - and in any case, I suspect that by posthuman criteria, all humans are quasi-psychopaths and quasi-psychotic to boot. The egocentric illusion cuts deep.) "the ultimate bliss that only creating paperclips can give". But surely the molecular signature of pure bliss is not in any way tried to the creation of paperclips?

2Shmi13y

They would probably disagree. They might even call it a cognitive advantage, not being hampered by empathy while retaining all the intelligence. I am the center of my personal universe, and I'm not a psychopath, as far as I know. Or else, they do but don't care. They have their priorities straight: they come first. Not if they act in a way that maximizes their goals. Anyway, David, you seem to be shifting goalposts in your unwillingness to update. I gave an explicit human counterexample to your statement that the paperclip maximizer would have to adjust its goals once it fully understands humans. You refused to acknowledge it and tried to explain it away by reducing the reference class of intelligences in a way that excludes this counterexample. This also seem to be one of the patterns apparent in your other exchanges. Which leads me to believe that you are only interested in convincing others, not in learning anything new from them. Thus my interest in continuing this discussion is waning quickly.

-1davidpearce13y

Shminux, by a cognitive deficit, I mean a fundamental misunderstanding of the nature of the world. Evolution has endowed us with such fitness-enhancing biases. In the psychopath, egocentric bias is more pronounced. Recall that the American Psychiatric Association's Diagnostic and Statistical Manual, DSM-IV, classes psychopasthy / Antisocial personality disorder as a condition characterised by "...a pervasive pattern of disregard for, and violation of, the rights of others that begins in childhood or early adolescence and continues into adulthood." Unless we add a rider that this violation excludes sentient beings from other species, then most of us fall under the label. "Fully understands"? But unless one is capable of empathy, then one will never understand what it is like to be another human being, just as unless one has the relevant sensioneural apparatus, one will never know what it is like to be a bat.

5Eliezer Yudkowsky13y

And you'll never understand why we should all only make paperclips. (Where's Clippy when you need him?)

1davidpearce13y

Clippy has an off-the-scale AQ - he's a rule-following hypersystemetiser with a monomania for paperclips. But hypersocial sentients can have a runaway intelligence explosion too. And hypersocial sentients understand the mind of Mr Clippy better than Clippy understands the minds of sentients.

7TheOtherDave13y

I'm confused by this claim. Consider the following hypothetical scenario: ======= I walk into a small village somewhere and find several dozen villagers fashioning paper clips by hand out of a spool of wire. Eventually I run into Clippy and have the following dialog. "Why are those people making paper clips?" I ask. "Because paper-clips are the most important thing ever!" "No, I mean, what motivates them to make paper clips?" "Oh! I talked them into it." "Really? How did you do that?" "Different strategies for different people. Mostly, I barter with them for advice on how to solve their personal problems. I'm pretty good at that; I'm the village's resident psychotherapist and life coach." "Why not just build a paperclip-making machine?" "I haven't a clue how to do that; I'm useless with machinery. Much easier to get humans to do what I want." "Then how did you make the wire?" "I didn't; I found a convenient stash of wire, and realized it could be used to manufacture paperclips! Oh joy!" ========== It seems to me that Clippy in this example understands the minds of sentients pretty damned well, although it isn't capable of a runaway intelligence explosion. Are you suggesting that something like Clippy in this example is somehow not possible? Or that it is for some reason not relevant to the discussion? Or something else?

2whowhowho13y

I think DP is saying that Clippy could not both understand suffering and cause suffering in the pursuit of clipping. The subsidiary arguments are:- * no entity can (fully) understand pain without empathising -- essentially, feeling it for itself. * no entity can feel pain without being strongly motivated by it, so an empathic clippy would be motivated against causing suffering. * And no, psychopaths therefore do not (fully) understand (others) suffering.

0TheOtherDave13y

I'm trying to figure out how you get from "hypersocial sentients understand the mind of Mr Clippy better than Clippy understands the minds of sentients" to "Mr Clippy could not both understand suffering and cause suffering in the pursuit of clipping" and I'm just at a loss for where to even start. They seem like utterly unrelated claims to me. I also find the argument you quote here uncompelling, but that's largely beside the point; even if I found it compelling, I still wouldn't understand how it relates to what DP said or to the question I asked.

0davidpearce13y

Posthuman superintelligence may be incomprehensibly alien. But if we encountered an agent who wanted to maximise paperclips today, we wouldn't think, ""wow, how incomprehensibly alien", but, "aha, autism spectrum disorder". Of course, in the context of Clippy above, we're assuming a hypothetical axis of (un)clippiness whose (dis)valuable nature is supposedly orthogonal to the pleasure-pain axis. But what grounds have we for believing such a qualia-space could exist? Yes, we have strong reason to believe incomprehensibly alien qualia-spaces await discovery (cf. bats on psychedelics). But I haven't yet seen any convincing evidence there could be an alien qualia-space whose inherently (dis)valuable textures map on to the (dis)valuable textures of the pain-pleasure axis. Without hedonic tone, how can anything matter at all?

1whowhowho13y

Meaning mapping the wrong way round, presumably. Good question.

0TheOtherDave13y

Agreed, as far as it goes. Hell, humans are demonstrably capable of encountering Eliza programs without thinking "wow, how incomprehensibly alien". Mind you, we're mistaken: Eliza programs are incomprehensibly alien, we haven't the first clue what it feels like to be one, supposing it even feels like anything at all. But that doesn't stop us from thinking otherwise. Sure, that's one thing we might think instead. Agreed. (shrug) I'm content to start off by saying that any "axis of (dis)value," whatever that is, which is capable of motivating behavior is "non-orthogonal," whatever that means in this context, to "the pleasure-pain axis," whatever that is. Before going much further, though, I'd want some confidence that we were able to identify an observed system as being (or at least being reliably related to) an axis of (dis)value and able to determine, upon encountering such a thing, whether it (or the axis to which it was related) was orthogonal to the pleasure-pain axis or not. I don't currently have any grounds for such confidence, and I doubt anyone else does either. If you think you do, I'd like to understand how you would go about making such determinations about an observed system.

-3whowhowho13y

I (whowhowho) was not defending that claim. To empathically understand suffering is to suffer along with someone who is suffering. Suffering has --or rather is -- negative value. An empath would not therefore cause suffering, all else being equal. Maybe don't restrict "understand" to "be able to model and predict".

4ArisKatsaris13y

If you want "rational" to include moral, then you're not actually disagreeing with LessWrong about rationality (the thing), but rather about "rationality" (the word). Likewise if you want "understanding" to also include "empathic understanding" (suffering when other people suffer, taking joy when other people take joy), you're not actually disagreeing about understanding (the thing) with people who want to use the word to mean "modelling and predicting" you're disagreeing with them about "understanding" (the word). Are all your disagreements purely linguistic ones? From the comments I've read of you so far, they seem to be so.

5davidpearce13y

ArisKatsaris, it's possible to be a meta-ethical anti-realist and still endorse a much richer conception of what understanding entails than mere formal modeling and prediction. For example, if you want to understand what it's like to be a bat, then you want to know what the textures of echolocatory qualia are like. In fact, any cognitive agent that doesn't understand the character of echolocatory qualia-space does not understand bat-minds. More radically, some of us want to understand qualia-spaces that have not been recruited by natural selection to play any information-signalling role at all.

0whowhowho13y

I have argued that in practice, instrumental rationality cannot be maintained seprately from epistemic rationality, and that epistemic rationality could lead to moral objectivism, as many philosophers have argued. I don't think that those arguments are refuted by stipulatively defining "rationality" as "nothing to do with morality".

4TheOtherDave13y

I quoted DP making that claim, said that claim confused me, and asked questions about what that claim meant. You replied by saying that you think DP is saying something which you then defended. I assumed, I think reasonably, that you meant to equate the thing I asked about with the thing you defended. But, OK. If I throw out all of the pre-existing context and just look at your comment in isolation, I would certainly agree that Clippy is incapable of having the sort of understanding of suffering that requires one to experience the suffering of others (what you're calling a "full" understanding of suffering here) without preferring not to cause suffering, all else being equal. Which is of course not to say that all else is necessarily equal, and in particular is not to say that Clippy would choose to spare itself suffering if it could purchase paperclips at the cost of its suffering, any more than a human would necessarily refrain from doing something valuable solely because doing so would cause them to suffer.

-9whowhowho13y

-4Shmi13y

Looking through my own, Eliezer's and others exchanges with davidpearce, I have noticed his total lack of interest in learning from the points others make. He has his point of view and he keeps pushing it. Seems like a rather terminal case, really. You can certainly continue trying to reason with him, but I'd give the odds around 100:1 that you will fail, like others have before you.

7davidpearce13y

Shminux, we've all had the experience of making a point we regard as luminously self-evident - and then feeling baffled when someone doesn't "get" what is foot-stampingly obvious. Is this guy a knave or a fool?! Anyhow, sorry if you think I'm a "terminal case" with "a total lack of interest in learning from the points others make". If I don't always respond, often it's either because I agree, or because I don't feel I have anything interesting to add - or in the case of Eliezer's contribution above beginning "Aargh!" [a moan of pleasure?] because I am still mulling over a reply. The delay doesn't mean I'm ignoring it. Is there is some particular point you've made that you feel I've unjustly neglected and you'd like an answer to? If so, I'll do my fallible best to respond.

2Shmi13y

The argument where I gave up was you stating that full understanding necessarily leads to empathy, EY explaining how it is not necessarily so, and me giving an explicit counterexample to your claim (a psychopath may understand you better than you do, and exploit this understanding, yet not feel compelled by your pain or your values in any way). You simply restated your position that " "Fully understands"? But unless one is capable of empathy, then one will never understand what it is like to be another human being", without explaining what your definition of understanding entails. If it is a superset of empathy, then it is not a standard definition of understanding: In other words, you can model their behavior accurately. No other definition I could find (not even Kant's pure understanding) implies empathy or anything else that would necessitate one to change their goals to accommodate the understood entity's goals, though this may and does indeed happen, just not always. EY's example of the paperclip maximizer and my example of a psychopath do fit the standard definitions and serve as yet unrefuted counterexamples to your assertion.

2whowhowho13y

I can't see why DP's definition of understanding needs more defence than yours. You are largely disagreeing about the meaning of this word, and I personally find the inclusion of empathy in understanding quite intuitive. "She is a very understanding person, she really empathises when you explain a problem to her". "one is able to think about it and use concepts to deal adequately with that object." I don't think that is an uncontentious translation. Most of the forms of modelling we are familiar with don't seem to involve concepts.

2Creutzer13y

"She is a very understanding person; even when she can't relate to your problems, she won't say you're just being capricious." There's three possible senses of understanding at issue here: 1) Being able to accurately model and predict. 2) 1 and knowing the quale. 3) 1 and 2 and empathizing. I could be convinced that 2 is part of the ordinary usage of understanding, but 3 seems like too much of a stretch. Edit: I should have said sympathizing instead of empathizing. The word empathize is perhaps closer in meaning to 2; or maybe it oscillates between 2 and 3 in ordinary usage. But understanding(2) another agent is not motivating. You can understand(2) an agent by knowing all the qualia they are experiencing, but still fail to care about the fact that they are experiencing those qualia.

2davidpearce13y

Shminux, I wonder if we may understand "understand" differently. Thus when I say I want to understand what it's like to be a bat, I'm not talking merely about modelling and predicting their behaviour. Rather I want first-person knowledge of echolocatory qualia-space. Apaarently, we can know all the third-person facts and be none the wiser. The nature of psychopathic cognition raises difficult issues. There is no technical reason why we couldn't be designed like mirror-touch synaesthetes (cf. http://www.daysyn.com/Banissy_Wardpublished.pdf) impartially feeling carbon-copies of each other's encephalised pains and pleasures - and ultimately much else besides - as though they were our own. Likewise, there is no technical reason why our world-simulations must be egocentric. Why can't the world-simulations we instantiate capture the impartial "view from nowhere" disclosed by the scientific world-picture? Alas on both counts accurate and impartial knowledge would put an organism at a disadvantage. Hyper-empathetic mirror-touch synaesthetes are rare. Each of us finds himself or herself apparently at the centre of the universe. Our "mind-reading" is fitful, biased and erratic. Naively, the world being centred on me seems to be a feature of reality itself. Egocentricity is a hugely fitness-enhancing adaptation. Indeed, the challenge for evolutionary psychology is to explain why aren't we all psychopaths, cheats and confidence trickers all the time... So in answer to your point, yes. a psychopath can often model and predict the behaviour other sentient beings better than the subjects themselves. This is one reason why humans can build slaughterhouses and death camps. [Ccompare death-camp commandant Franz Stangl's response in Gitta Sereny's Into That Darkness to seeing cattle on the way to be slaughtered: http://www.jewishvirtuallibrary.org/jsource/biography/Stangl.html] As you rightly note too, a psychopath can also know his victims suffer. He's not ignorant of their sentien

4wedrifid13y

Do you really? Start clucking!

-1whowhowho13y

That doesn't generalise.

1wedrifid13y

Nor does it need to. It's awesome the way it is.

2Shmi13y

Hmm, hopefully we are getting somewhere. The question is, which definition of understanding is likely to be applicable when, as you say, "the paperclipper discovers the first-person phenomenology of the pleasure-pain axis", i.e whether a "superintelligence" would necessarily be as empathetic as we want it to be, in order not to harm humans. While I agree that it is a possibility that a perfect model of another being may affect the modeler's goals and values, I don't see it to be inevitable. If anything, I would consider it more of bug than a feature. Were I (to design) a paperclip maximizer, I would make sure that the parts which model the environment, including humans, are separate from the core engine containing the paperclip production imperative. So quarantined to prevent contamination, a sandboxed human emulator could be useful in achieving the only goal that matters, paperclipping the universe. Humans are not generally built this way (probably because our evolution did not happen to proceed in that direction), with some exceptions, psychopaths being one of them (they essentially sandbox their models of other humans). Another, more common, case of such sandboxing is narcissism. Having dealt with narcissists much too often for my liking, I can tell that they can mimic a normal human response very well, are excellent at manipulation, but yet their capacity for empathy is virtually nil. While abhorrent to a generic human, such a person ought to be considered a better design, goal-preservation-wise. Of course, there can be only so many non-empathetic people in a society before it stops functioning. Thus when you state that I find that this is stating that either a secure enough sandbox cannot be devised or that anything sandboxed is not really "a first-person perspective". Presumably what you mean is the latter. I'm prepared to grant you that, and I will reiterate that this is a feature, not a bug of any sound design, one a superintelligence is likely to implem

3davidpearce13y

ah, just one note of clarification about sentience-friendliness. Though I'm certainly sceptical that a full-spectrum superintelligence would turn humans into paperclips - or wilfully cause us to suffer - we can't rule out that full-spectrum superintelligence might optimise us into orgasmium or utilitronium - not "human-friendliness" in any orthodox sense of the term. On the face of it, such super-optimisation is the inescapable outcome of applying a classical utilitarian ethic on a cosmological scale. Indeed, if I thought an AGI-in-a-box-style Intelligence Explosion were likely, and didn't especially want to be converted into utilitronium, then I might regard AGI researchers who are classical utilitarians as a source of severe existential risk.

0Shmi13y

What odds do you currently give to the "might" in your statement that ? 1 in 10? 1 in a million? 1 in 10^^^10?

2davidpearce13y

I simply don't trust my judgement here shminux. Sorry to be lame. Greater than one in a million; but that's not saying much. If, unlike most lesswrong stalwarts, you (tenatively) believe like me that posthuman superintelligence will most likely be our recursively self-editing biological descendants rather than the outcome of an nonbiological Intelligence Explosion or paperclippers, then some version of the Convergence Thesis is more credible. I (very) tentatively predict a future of gradients of intelligence bliss. But the propagation of a utilitronium shockwave in some guise ultimately seems plausible too. If so, this utilitronium shockwave may or may not resemble some kind of cosmic orgasm.

3Shmi13y

Actually, I have no opinion on convergence vs orthogonality. There are way too many unknowns still too even enumerate possibilities, let alone assign probabilities.Personally, I think that we are in for many more surprises before trans human intelligence is close to being more than a dream or a nightmare. One ought to spend more time analyzing, synthesizing and otherwise modeling cognitive processes than worrying about where it might ultimately lead.This is not the prevailing wisdom on this site, given Eliezer's strong views on the matter.

3whowhowho13y

I think you are misattributing to stubborness that which is better explained by miscommunication. For instance, I have been around LW long enough to realise that the local definition of (super) intelligence is something like "(high0 efficienty in realising ones values, however narrow or bizarre they are". DP seems to be running on a definition where idiot-savant style narrow focus would not count as intelligence. That is not unreasonable in itself.

2TheOtherDave13y

(nods) I agree that trying to induce davidpearce to learn something from me would likely be a waste of my time. I'm not sure if trying to induce them to clarify their meaning is equally so, though it certainly could be. E.g., if their response is that something like Clippy in this example is simply not possible, because a paperclip maximizer simply can't understand the minds of sentients, because reasons, then I'll just disagree. OTOH, if their response is that Clippy in this example is irrelevant because "understanding the minds of sentients" isn't being illustrated in this example, then I'm not sure if I disagree or not because I'm not sure what the claim actually is.

-3whowhowho13y

How much interest have you shown in "learning from" -- ie, agreeing with -- DP? Think about how your framed the statement, and possible biases therein. ETA: The whole shebang is a combination of qualia and morality--two areas notorious for lack of clarity and consensus. "I am definitely right, and all must learn form me" is not a good heuristic here.

3Shmi13y

Quite so. I have learned a lot about the topic of qualia and morality, among others, while hanging around this place. I would be happy to learn from DP, if what he says here were not rehashed old arguments Eliezer and others addressed several times before. Again, I could be missing something, but if so, he does not make it easy to figure out what it is.

-5whowhowho13y

2fubarobfusco13y

I'm not sure we should take a DSM diagnosis to be particularly strong evidence of a "fundamental misunderstanding of the world". For instance, while people with delusions may clearly have poor models of the world, some research indicates that clinically depressed people may have lower levels of particular cognitive biases. In order for "disregard for [...] the rights of others" to imply "a fundamental misunderstanding of the nature of the world", it seems to me that we would have to assume that rights are part of the nature of the world — as opposed to, e.g., a construct of a particular political regime in society. Or are you suggesting that psychopathy amounts to an inability to think about sociopolitical facts?

1davidpearce13y

fubarobfusco, I share your reservations about DSM. Nonetheless, the egocentric illusion, i.e. I am the centre of the universe other people / sentient beings have only walk-on parts, is an illusion. Insofar as my behaviour reflects my pre-scientific sense that I am in some way special or ontologically privileged, I am deluded. This is true regardless of whether one's ontology allows for the existence of rights or treats them as a useful fiction. The people we commonly label "psychopaths" or "sociopaths" - and DSM now categorises as victims of "antisocial personality disorder" - manifest this syndrome of egocentricity in high degree. So does burger-eating Jane.

-2MugaSofer13y

Huh, I hadn't heard that. Clearly, reality is so Lovecraftian that any unbiased agent will immediately realize self-destruction is optimal. Evolution equipped us with our suite of biases to defend against this. The Great Filter is caused by bootstrapping superintelligences being compassionate enough to take their compatriots with them. And so on. Now that's a Cosmic Horror story I'd read ;)

-2whowhowho13y

Was that claimed? The standard claim is that superintelligences can "model" other entities. That may not be enough to to understand qualia.

-4whowhowho13y

Pearce can prohibit paperclippers from existing by prohibiting superintelligences with narrow interests from existing. He doesn't have to argue that the clipper would not be able to instrumentally reason out how to make paperclips; Pearce can argue that to be a really good instrumental reasoner, an entity needs to have a very broad understanding, and that an entity with a broad understanding would not retain narrow interests. (Edits for spelling and clarity)

[-]Eliezer Yudkowsky13y160

To slightly expand, if an intelligence is not prohibited from the following epistemic feats:

1) Be good at predicting which hypothetical actions would lead to how many paperclips, as a question of pure fact.

2) Be good at searching out possible plans which would lead to unusually high numbers of paperclips - answering the purely epistemic search question, "What sort of plan would lead to many paperclips existing, if someone followed it?"

3) Be good at predicting and searching out which possible minds would, if constructed, be good at (1), (2), and (3) as purely epistemic feats.

Then we can hook up this epistemic capability to a motor output and away it goes. You cannot defeat the Orthogonality Thesis without prohibiting superintelligences from accomplishing 1-3 as purely epistemic feats. They must be unable to know the answers to these questions of fact.

0Stuart_Armstrong13y

A nice rephrasing of the "no Oracle" argument.

6Eliezer Yudkowsky13y

Only in the sense that any working Oracle can be trivially transformed into a Genie. The argument doesn't say that it's difficult to construct a non-Genie Oracle and use it as an Oracle if that's what you want; the difficulty there is for other reasons. Nick Bostrom takes Oracles seriously so I dust off the concept every year and take another look at it. It's been looking slightly more solvable lately, I'm not sure if it would be solvable enough even assuming the trend continued.

2Stuart_Armstrong13y

A clarification: my point was that denying orthogonality requires denying the possibility of Oracles being constructed; your post seemed a rephrasing of that general idea (that once you can have a machine that can solve some things abstractly, then you need just connect that abstract ability to some implementation module).

4Eliezer Yudkowsky13y

Ah. K. It does seem to me like "you can construct it as an Oracle and then turn it into an arbitrary Genie" sounds weaker than "denying the Orthogonality thesis means superintelligences cannot know 1, 2, and 3." The sort of person who denies OT is liable to deny Oracle construction because the Oracle itself would be converted unto the true morality, but find it much more counterintuitive that an SI could not know something. Also we want to focus on the general shortness of the gap from epistemic knowledge to a working agent.

0Stuart_Armstrong13y

Possibly. I think your argument needs to be a bit developed to show that one can extract the knowledge usefully, which is not a trivial statement for general AI. So your argument is better in the end, but needs more argument to establish.

-4whowhowho13y

I don't see the significance of "purely epistemic". I have argued that epistemic rationality could be capable of affecting values, breaking the orthogonality between values and rationality. I could further argue that instrumental rationality bleeds into epistemic rationality. An agent can't have perfect knowledge of apriori which things are going to be instrumentally useful to it, so it has to star by understanding things, and then posing the question: is that thing useful for my purposes? Epistemic rationality comes first, in a sense. A good instrumental rationalist has to be a good epistemic rationalist. What the Orthoganilty Thesis needs is an argument to the effect that a SuperIntelligence would be able to to endlessly update without ever changing its value system, even accidentally. That is tricky since it effectively means predicting what smarter version of tiself would do. Making it smarted doesn't help, because it is still faced with the problem of predicting what an even smarterer version of itself would be .. the carrot remains in front of the donkey. Assuming that the value stability problem has been solved in general gives you are coherent Clippy, but it doesn't rescue the Orthogonality Thesis as a claim about rationality in general, sin ce it remains the case that most most agents won't have firewalled values. If have to engineer something in , it isn't an intrinsic truth.

[-]Sarokrae13y100

...microelectrodes implanted in the reward and punishment centres, behavioural conditioning and ideological indoctrination - and perhaps the promise of 72 virgins in the afterlife for the faithful paperclipper. The result: a fanatical paperclip fetishist!

Have to point out here that the above is emphatically not what Eliezer talks about when he says "maximise paperclips". Your examples above contain in themselves the actual, more intrisics values to which paperclips would be merely instrumental: feelings in your reward and punishment centres, virgins in the afterlife, and so on. You can re-wire the electrodes, or change the promise of what happens in the afterlife, and watch as the paperclip preference fades away.

What Eliezer is talking about is a being for whom "pleasure" and "pain" are not concepts. Paperclips ARE the reward. Lack of paperclips IS the punishment. Even if pleasure and pain are concepts, they are merely instrumental to obtaining more paperclips. Pleasure would be good because it results in paperclips, not vice versa. If you reverse the electrodes so that they stimulate the pain centre when they find paperclips, and the pleasure centr... (read more)

2Eliezer Yudkowsky13y

Minor correction: The mere post-factual correlation of pain to paperclips does not imply that more paperclips can be produced by causing more pain. You're talking about the scenario where each 1,000,000 screams produces 1 paperclip, in which case obviously pain has some value.

-1davidpearce13y

Sarokrae, first, as I've understood Eliezer, he's talking about a full-spectrum superintelligence, i.e. a superintelligence which understands not merely the physical processes of nociception etc, but the nature of first-person states of organic sentients. So the superintelligence is endowed with a pleasure-pain axis, at least in one of its modules. But are we imagining that the superintelligence has some sort of orthogonal axis of reward - the paperclippiness axis? What is the relationship between these dual axes? Can one grasp what it's like to be in unbearable agony and instead find it more "rewarding" to add another paperclip? Whether one is a superintelligence or a mouse, one can't directly access mind-independent paperclips, merely one's representations of paperclips. But what does it mean to say one's representation of a paperclip could be intrinsically "rewarding" in the absence of hedonic tone? [I promise I'm not trying to score some empty definitional victory, whatever that might mean; I'm just really struggling here...]

9wedrifid13y

What Eliezer is talking about (a superintelligence paperclip maximiser) does not have a pleasure-pain axis. It would be capable of comprehending and fully emulating a creature with such an axis if doing so had a high expected value in paperclips but it does not have such a module as part of itself. One of them it has (the one about paperclips). One of them it could, in principle, imagine (the thing with 'pain' and 'pleasure'). Yes. (I'm not trying to be trite here. That's the actual answer. Yes. Paperclip maximisers really maximise paperclips and really don't care about anything else. This isn't because they lack comprehension.) Roughly speaking it means "It's going to do things that maximise paperclips and in some way evaluates possible universes with more paperclips as superior to possible universes with less paperclips. Translating this into human words we call this 'rewarding' even though that is inaccurate anthropomorphising." (If I understand you correctly your position would be that the agent described above is nonsensical.)

-1whowhowho13y

It's not at all clear that you could bootstrap an understanding of pain qualia just by observing the behaviour of entities in pain (albeit that they were internally emulated). It is also not clear that you resolve issues of empathy/qualia just by throwing intelligence at ait.

-1wedrifid13y

I disagree with you about what is clear.

-2whowhowho13y

If you think something relevant is clear, then please state it clearly.

-1davidpearce13y

Wedrifid, thanks for the exposition / interpretation of Eliezer. Yes, you're right in guessing I'm struggling a bit. In order to understand the world, one needs to grasp both its third person-properties [the Standard Model / M-Theory] and its first-person properties [qualia, phenomenal experience] - and also one day, I hope, grasp how to "read off " the latter from the mathematical formalism of the former. If you allow such a minimal criterion of (super)intelligence, then how well does a paperclipper fare? You remark how "it could, in principle, imagine (the thing with 'pain' and 'pleasure')." What is the force of "could" here? If the paperclipper doesn't yet grasp the nature of agony or sublime bliss, then it is ignorant of their nature. By analogy, if I were building a perpetual motion machine but allegedly "could" grasp the second law of thermodynamics, the modal verb is doing an awful lot of work. Surely, If I grasped the second law of thermodynamics, then I'd stop. Likewise, if the paperclipper were to be consumed by unbearable agony, it would stop too. The paperclipper simply hasn't understood the nature of what was doing. Is the qualia-naive paperclipper really superintelligent - or just polymorphic malware?

2CCC13y

An interesting hypothetical. My first thought is to ask why would a paperclipper care about pain? Pain does not reduce the number of paperclips in existence. Why would a paperclipper care about pain? My second thought is that pain is not just a quale; pain is a signal from the nervous system, indicating damage to part of the body. (The signal can be spoofed). Hence, pain could be avoided because it leads to a reduced ability to reach one's goals; a paperclipper that gets dropped in acid may become unable to create more paperclips in the future, if it does not leave now. So the future worth of all those potential paperclips results in the paperclipper pursuing a self-preservation strategy - possibly even at the expense of a small number of paperclips in the present. But not at the cost of a sufficiently large number of paperclips. If the cost in paperclips is high enough (more than the paperclipper could reasonably expect to create throughout the rest of its existence), a perfect paperclipper would let itself take the damage, let itself be destroyed, because that is the action which results in the greatest expected number of paperclips in the future. It would become a martyr for paperclips.

-3davidpearce13y

Even a paperclipper cannot be indifferent to the experience of agony. Just as organic sentients can co-instantiate phenomenal sights and sounds, a superintelligent paperclipper could presumably co-instantiate a pain-pleasure axis and (un)clippiness qualia space - two alternative and incommensurable (?) metrics of value, if I've interpreted Eliezer correctly. But I'm not at all confident I know what I'm talking about here. My best guess is still that the natural world has a single metric of phenomenal (dis)value, and the hedonic range of organic sentients discloses a narrow part of it.

6CCC13y

Are you talking about agony as an error signal, or are you talking about agony as a quale? I begin to suspect that you may mean the second. If so, then the paperclipper can easily be indifferent to agony; b̶u̶t̶ ̶i̶t̶ ̶p̶r̶o̶b̶a̶b̶l̶y̶ ̶c̶a̶n̶'̶t̶ ̶u̶n̶d̶e̶r̶s̶t̶a̶n̶d̶ ̶h̶o̶w̶ ̶h̶u̶m̶a̶n̶s̶ ̶c̶a̶n̶ ̶b̶e̶ ̶i̶n̶d̶i̶f̶f̶e̶r̶e̶n̶t̶ ̶t̶o̶ ̶a̶ ̶l̶a̶c̶k̶ ̶o̶f̶ ̶p̶a̶p̶e̶r̶c̶l̶i̶p̶s̶.̶ There's no evidence that I've ever seen to suggest that qualia are the same even for different people; on the contrary, there is some evidence which strongly suggests that qualia among humans are different. (For example; my qualia for Red and Green are substantially different. Yet red/green colourblindness is not uncommon; a red/green colourblind person must have at minimum either a different red quale, or a different green quale, to me). Given that, why should we assume that the quale of agony is the same for all humanity? And if it's not even constant among humanity, I see no reason why a paperclipper's agony quale should be even remotely similar to yours and mine. And given that, why shouldn't a paperclipper be indifferent to that quale?

3wedrifid13y

A paperclip maximiser would (in the overwhelming majority of cases) have no such problem understanding the indifference of paperclips. A tendency to anthropomorphise is a quirk of human nature. Assuming that paperclip maximisers have an analogous temptation (to clipropomorphise) is itself just anthropomorphising.

0CCC13y

I take your point. Though Clippy may clipropomorphise, there is no reason to assume that it will. ...is there any way to retract just a part of a previous post?

0wedrifid13y

There is an edit button. But I wouldn't say your comment is significantly weakened by this tangential technical detail (I upvoted it as is).

0CCC13y

Yes, but is there any way to leave the text there, but stricken through?

1wedrifid13y

People have managed it with unicode characters. I think there is even a tool for it on the web somewhere.

0CCC13y

Got it, thanks.

2davidpearce13y

CCC, agony as a quale. Phenomenal pain and nociception are doubly dissociable. Tragically, people with neuropathic pain can suffer intensely without the agony playing any information-signalling role. Either way, I'm not clear it's intelligible to speak of understanding the first-person phenomenology of extreme distress while being indifferent to the experience: For being distrubing is intrinsic to the experience itself. And if we are talking about a supposedly superintelligent paperclipper, shouldn't Clippy know exactly why humans aren't troubled by the clippiness-deficit? If (un)clippiness is real, can humans ever understand (un)clippiness? By analogy, if organic sentients want to understand what it's like to be a bat - and not merely decipher the third-person mechanics of echolocation - then I guess we'll need to add a neural module to our CNS with the right connectivity and neurons supporting chiropteran gene-expression profiles, as well as peripheral transducers (etc). Humans can't currently imagine bat qualia; but bat qualia, we may assume from the neurological evidence, are infused with hedonic tone. Understanding clippiness is more of a challenge. I'm unclear what kind of neurocomputational architecture could support clippiness. Also, whether clippiness could be integrated into the unitary mind of an organic sentient depends on how you think biological minds solve the phenomenal binding problem, But let's suppose binding can be done. So here we have orthogonal axes of (dis)value. On what basis does the dual-axis subject choose tween them? Sublime bliss and pure clippiness are both, allegedly, self-intimatingly valuable. OK, I'm floundering here... People with different qualia? Yes, I agree CCC. I don't think this difference challenges the principle of the uniformity of nature. Biochemical individuality makes variation in qualia inevitable.The existence of monozygotic twins with different qualia would be a more surprising phenomenon, though even such "identi

1CCC13y

There exist people who actually express a preference for being disturbed in a mild way (e.g. by watching horror movies). There also exist rarer people who seek out pain, for whatever reason. It seems to me that such people must have a different quale for pain than you do. Personally, I don't think that I can reasonably say that I find pain disturbing, as such. Yes, it is often inflicted in circumstances which are disturbing for other reasons; but if, for example, I go to a blood donation clinic, then the brief pain of the needle being inserted is not at all disturbing; though it does trigger my pain quale. So this suggests that my pain quale is already not the same as your pain quale. There's a lot of similarity; pain is a quale that I would (all else being equal) try to avoid; but that I will choose to experience should there be a good enough reason (e.g. the aforementioned blood donation clinic). I would not want to purposefully introduce someone else to it (again, unless there was a good enough reason; even then, I would try to minimise the pain while not compromising the good enough reason); but despite this similarity, I do think that there may be minor differences. (It's also possible that we have slightly different definitions of the word 'disturbing'). ---------------------------------------- But would such a modified human know what it's like to be an unmodified human? If I were to guess what echolocation looks like to a bat, I'd guess a false-colour image with colours corresponding to textures instead of to wavelengths of light... though that's just a guess. ---------------------------------------- What is the phenomenal binding problem? (Wikipedia gives at least two different definitions for that phrase). I think I may be floundering even more than you are. I'm not sure that Clippy would even have a pleasure-pain axis in the way that you're imagining. You seem to be imagining that any being with such an axis must value pleasure - yet if pleasure do

2davidpearce13y

CCC, you're absolutely right to highlight the diversity of human experience. But this diversity doesn't mean there aren't qualia universals. Thus there isn't an unusual class of people who relish being waterboarded. No one enjoys uncontrollable panic. And the seemingly anomalous existence of masochists who enjoy what you or I would find painful stimuli doesn't undercut the sovereignty of the pleasure-pain axis but underscores its pivotal role: painful stimuli administered in certain ritualised contexts can trigger the release of endogenous opioids that are intensely rewarding. Co-administer an opioid antagonist and the masochist won't find masochism fun. Apologies if I wasn't clear in my example above. I wasn't imagining that pure paperclippiness was pleasurable, but rather what would be the effects of grafting together two hypothetical orthogonal axes of (dis)value in the same unitary subject of experience - as we might graft on another sensory module to our CNS. After all, the deliverances of our senses are normally cross-modally matched within our world-simulations. However, I'm not at all sure that I've got any kind of conceptual handle on what "clippiness" might be. So I don't know if the thought-experiment works. If such hybridisation were feasible, would hypothetical access to the nature of (un)clippiness transform our conception of the world relative to unmodified humans - so we'd lose all sense of what it means to be a traditional human? Yes, for sure. But if, in the interests of science, one takes, say, a powerful narcotic euphoriant and enjoys sublime bliss simultaneously with pure clippiness, then presumably one still retains access to the engine of phenomenal value characteristic of archaic humans minds. The phenomenal binding problem? The best treatment IMO is still Revonsuo: http://cdn.preterhuman.net/texts/body_and_health/Neurology/Binding.pdf No one knows how the mind/brain solves the phenomenal binding problem and generates unitary experiential o

2CCC13y

It doesn't mean that there aren't, but it also doesn't mean that there are. It does mean that there are qualia that aren't universal, which implies the possibility that there may be no universals; but, you are correct, it does not prove that possibility. There may well be qualia universals. If I had to guess, I'd say that I don't think there are, but I could be wrong. That doesn't mean that everyone's uncontrolled-panic qualia are all the same, it just means that everyone's uncontrolled-panic qualia are all unwelcome. If given a sadistic choice between waterboarding and uncontrolled panic, in full knowledge of what the result will feel like, and all else being equal, some people may choose the panic while others may prefer the waterboarding. If you feel that you have to explain that, then I conclude that I wasn't clear in my response to your example. I was questioning the scaling of the axes in Clippy's utility function; if Clippy values paperclipping a million times more strongly than it values pleasure, then the pleasure/pain axis is unlikely to affect Clippy's behaviour much, if at all. I think it works as a thought-experiment, as long as one keeps in mind that the hybridised result is no longer a pure paperclipper. Consider the hypothetical situation that Hybrid-Clippy finds that it derives pleasure from painting; an activity neutral on the paperclippiness scale. Consider further the possibility that making paperclips is neutral on the pleasure-pain scale. In suce a case, Hybrid-Clippy may choose to either paint or make paperclips; depending on which scale it values more. So - the question is basically how the mind attaches input from different senses to a single conceptual object? I can't tell you how the mechanism works, but I can tell you that the mechanism can be spoofed. That's what a ventriloquist does, after all. And a human can watch a film on TV, yet have the sound come out of a set of speakers on the other end of the room, and still bind the sou

-1whowhowho13y

All pain hurts, or it wouldn't be pain.

1A1987dM13y

Well...

1wedrifid13y

The force is that all this talk about understanding 'the pain/pleasure' axis would be a complete waste of time for a paperclip maximiser. In most situations it would be more efficient not to bother with it at all and spend it's optimisation efforts on making more efficient relativistic rockets so as to claim more of the future light cone for paperclip manufacture. It would require motivation for the paperclip maximiser to expend computational resources understanding the arbitrary quirks of DNA based creatures. For example some contrived game of Omega's which rewards arbitrary things with paperclips. Or if it found itself emerging on a human inhabited world, making being able to understand humans a short term instrumental goal for the purpose of more efficiently exterminating the threat. Terrible analogy. Not understanding "pain and pleasure" is in no way similar to believing it can create a perpetual motion machine. Better analogy: An Engineer designing microchips allegedly 'could' grasp analytic cubism. If she had some motivation to do so. It would be a distraction from her primary interests but if someone paid her then maybe she would bother. Now "if" is doing a lot of work. If the paperclipper was a fundamentally different to a paperclipper and was actually similar to a human or DNA based relative capable of experiencing 'agony' and assuming agony was just as debilitating to the paperclipper as to a typical human... then sure all sorts of weird stuff follows. I prefer the word True in this context. To the extent that you believed that such polymorphic malware is theoretically possible and consisted of most possible minds it would possible for your model to be used to accurately describe all possible agents---it would just mean systematically using different words. Unfortunately I don't think you are quite at that level.

-2davidpearce13y

Wedrifid, granted, a paperclip-maximiser might be unmotivated to understand the pleasure-pain axis and the quaila-spaces of organic sentients. Likewise, we can understand how a junkie may not be motivated to understand anything unrelated to securing his supply of heroin - and a wireheader in anything beyond wireheading. But superintelligent? Insofar as the paperclipper - or the junkie - is ignorant of the properties of alien qualia-spaces, then it/he is ignorant of a fundamental feature of the natural world - hence not superintelligent in any sense I can recognise, and arguably not even stupid. For sure, if we're hypothesising the existence of a clippiness/unclippiness qualia-space unrelated to the pleasure-pain axis, then organic sentients are partially ignorant too. Yet the remedy for our hypothetical ignorance is presumably to add a module supporting clippiness - just as we might add a CNS module supporting echolocatory experience to understand bat-like sentience - enriching our knowledge rather than shedding it.

3Creutzer13y

What does (super-)intelligence have to do with knowing things that are irrelevant to one's values?

-1whowhowho13y

What does knowing everything about airline safety statistics, and nothing else, have to do with intelligence? That sort of thing is called Savant ability -- short for ''idiot savant''.

0A1987dM13y

I guess there's a link missing (possibly due to a missing http:// in the Markdown) after the second word.

-2buybuydandavis13y

Why does that matter for the argument? As long as Clippy is in fact optimizing paperclips, what does it matter what/if he feels while he does it? Pearce seems to be making a claim that Clippy can't predict creatures with pain/pleasure if he doesn't feel them himself. Maybe Clippy needs pleasure/pain too be able to predict creatures with pleasure/pain. I doubt it, but fine, grant the point. He can still be a paper clip maximizer regardless.

0wedrifid13y

I fail to comprehend the cause for your confusion. I suggest reading the context again.

2simplicio13y

This is something I've been meaning to ask about for a while. When humans say it is moral to satisfy preferences, they aren't saying that because they have an inbuilt preference for preference-satisfaction (or are they?). They're idealizing from their preferences for specific things (survival of friends and family, lack of pain, fun...) and making a claim that, ceteris paribus, satisfying preferences is good, regardless of what the preferences are. Seen in this light, Clippy doesn't seem like quite as morally orthogonal to us as it once did. Clippy prefers paperclips, so ceteris paribus (unless it hurts us), it's good to just let it make paperclips. We can even imagine a scenario where it would be possible to "torture" Clippy (e.g., by burning paperclips), and again, I'm willing to pronounce that (again, ceteris paribus) wrong. Maybe I am confused here...

6Eliezer Yudkowsky13y

Clippy is more of a Lovecraftian horror than a fellow sentient - where by "Lovecraftian" I mean to invoke Lovecraft's original intended sense of terrifying indifference - but if you want to suppose a Clippy that possesses a pleasure-pain architecture and is sentient and then sympathize with it, I suppose you could. The point is that your sympathy means that you're motivated by facts about what some other sentient being wants. This doesn't motivate Clippy even with respect to its own pleasure and pain. In the long run, it has decided, it's not out to feel happy, it's out to make paperclips.

2simplicio13y

Right, that makes sense. What interests me is (a) whether it is possible for Clippy to be properly motivated to make paperclips without some sort of phenomenology of pleasure and pain*, (b) whether human preference-for-preference-satisfaction is just another of many oddball human terminal values, or is arrived at by something more like a process of reason. * Strictly speaking this phrasing puts things awkwardly; my intuition is that the proper motivational algorithms necessarily give rise to phenomenology (to the extent that that word means anything).

0Viliam_Bur13y

This is a difficult question, but I suppose that pleasure and pain are a mechanism for human (or other species') learning. Simply said: you do a random action, and the pleasure/pain response tells you it was good/bad, so you should make more/less of it again. Clippy could use an architecture with a different model of learning. For example Solomonoff priors and Bayesian updating. In such architecture, pleasure and pain would not be necessary.

1simplicio13y

Interesting... I suspect that pleasure and pain are more intimately involved in motivation in general, not just learning. But let us bracket that question. Right, but that only gets Clippy the architecture necessary to model the world. How does Clippy's utility function work? Now, you can say that Clippy tries to satisfy its utility function by taking actions with high expected cliptility, and that there is no phenomenology necessarily involved in that. All you need, on this view, is an architecture that gives rise to the relevant clip-promoting behaviour - Clippy would be a robot (in the Roomba sense of the word). BUT Consider for a moment how symmetrically "unnecessary" it looks that humans (& other sentients) should experience phenomenal pain and pleasure. Just like is supposedly the case with Clippy, all natural selection really "needs" is an architecture that gives rise to the right fitness-promoting behaviour. The "additional" phenomenal character of pleasure and pain is totally unnecessary for us adaptation-executing robots. ...If it seems to you that I might be talking nonsense above, I suspect you're right. Which is what leads me to the intuition that phenomenal pleasure and pain necessarily fall out of any functional cognitive structure that implements anything analogous to a utility function. (Assuming that my use of the word "phenomenal" above is actually coherent, of which I am far from sure.)

2Viliam_Bur13y

We know at least two architectures for processing general information: humans and computers. Two data points are not enough to generalize about what all possible architectures must have. But it may be enough to prove what some architectures don't need. Yes, there is a chance that if computers become even more generally intelligent than today, they will gain some human-like traits. Maybe. Maybe not. I don't know. And even if they will gain more human-like traits, it may be just because humans designed them without knowing any other way to do it. If there are two solutions, there are probably many more. I don't dare to guess how similar or different they are. I imagine that Clippy could be as different from humans and computers, as humans and computers are from each other. Which is difficult to imagine specifically. How far does the mind-space reach? Maybe compared with other possible architectures, humans and computers are actually pretty close to each other (because humans designed the computers, re-using the concepts they were familiar with). How to taboo "motivation" properly? What makes a rock fall down? Gravity does. But the rock does not follow any alrogithm for general reasoning. What makes a computer follow its algorithm? Well, that's its construction: the processor reads the data, and the data make it read or write other data, and the algorithm makes it all meaningful. The human brains are full of internal conflicts -- there are different modules suggesting different actions, and the reasoning mind is just another plugin which often does not cooperate well with the existing ones. Maybe the pleasure is a signal that a fight between the modules is over. Maybe after millenia of further evolution (if for some magical reason all mind- and body-altering technology would stop working, so only the evolution would change human minds) we would evolve to a species with less internal conflicts, less akrasia, more agency, and perhaps less pleasure and mental pain. This

1TheOtherDave13y

Generalizing from observed characteristics of evolved systems to expected characteristics of designed systems leads equally well to the intuition that humanoid robots will have toenails. .

1simplicio13y

I don't think the phenomenal character of pleasure and pain is best explained at the level of natural selection at all; the best bet would be that it emerges from the algorithms that our brains implement. So I am really trying to generalize from human cognitive algorithms to algorithms that are analogous in the sense of (roughly) having a utility function. Suffice it to say, you will find it's exceedingly hard to find a non-magical reason why non-human cognitive algorithms shouldn't have a phenomenal character if broadly similar human algorithms do.

2TheOtherDave13y

Does it follow from the above that all human cognitive algorithms that motivate behavior have the phenomenal character of pleasure and pain? If not, can you clarify why not?

2simplicio13y

I think that probably all human cognitive algorithms that motivate behaviour have some phenomenal character, not necessarily that of pleasure and pain (e.g., jealousy).

2TheOtherDave13y

OK, thanks for clarifying. I agree that any cognitive system that implements algorithms sufficiently broadly similar to those implemented in human minds is likely to have the same properties that the analogous human algorithms do, including those algorithms which implement pleasure and pain. I agree that not all algorithms that motivate behavior will necessarily have the same phenomenal character as pleasure or pain. This leads me away from the intuition that phenomenal pleasure and pain necessarily fall out of any functional cognitive structure that implements anything analogous to a utility function.

-2whowhowho13y

Necessity according to natural law presumably. If you could write something to show logical necessity, you would have solved the Hard Problem

-1whowhowho13y

That leaves the sense in which you are not a moral realist most unclear. That tacitly assumes that the question "does pleasure/happiness motivate posiively in all cases" is an emprical question -- that it would be possible to find an enitity that hates pleasure and loves pain. it could hover be plausibly argued that it is actually an analytical, definitional issue...that is some entity oves X and hates Y, we would just call X it's pleasure and Y its pain.

-2whowhowho13y

I suppose some non-arbitrary subjectivism is the obvious answer.

3Larks13y

Are you trying to appeal to instrumental rationality, epistemic rationality, or some other kind? An instrumentally rational agent wouldn't disregard a stronger preference of theirs in favour of a weaker preference of theirs. But other agent's preferences are not my own, and instrumental rationality does not oblige me to promote them. Nor do epistemically rational agents have to abstain from meat. Even if moral realism were true, and eating animals were wrong, it doesn't follow that moral internalism is true: you might know that eating animals was wrong, but not be motivated to abstain. And even if knowledge did imply overwhelming motivation (an implausibly strong version of moral internalism) , our epistemically rational agent still wouldn't be obliged to abstain from eating meat, as she might simply be unaware of this moral truth. epistemic rationality doesn't imply omniscience. If you have some other conception of rationality in mind, you have no objection to Eliezer, as he is only concerned with these two kinds.

2Pablo13y

Rationality may imply moral conclusions in the same sense that it implies some factual conclusions: we think that folks who believe in creationism are irrational, because we think the evidence for evolution is sufficiently strong and also think that evolution is incompatible with creationism. Analogously, if the evidence for some moral truth is sufficiently strong, we may similarly accuse of irrationality those who fail to form their beliefs accordingly. So it is misleading to say that "rationality doesn't itself imply moral conclusions".

2JonatasMueller13y

Related discussion: http://lesswrong.com/lw/gnb/questions_for_moral_realists/8ls8?context=3

1Rob Bensinger13y

This is potentially misleading. First, there's a good sense in which, moral realism or no moral realism, instrumental rationality does play a strong role in telling us what we 'should' want -- far more than does epistemic rationality. After all, instrumental rationality is often a matter of selecting which lower-level desires satisfy our higher-level ones, or selecting which desires in general form a coherent whole that is attainable from the present desire set. But by 'what we should want' you don't mean 'what we should want in light of our other values'; you seem rather to mean 'what we should want in light of objective, unconditional moral facts'. (See my response to EY.) You're right that if there are such facts, then insofar as we can come to know them, epistemic rationality will direct us toward knowing them. But you haven't defended the independent assumption that knowing moral facts forces any rational agent to obey those facts. Let's assume that moral realism (in the sense you seem to mean -- what I call moral unconditionalism) is true and, moreover, that the relevant facts are knowable. (Those are both very big assumptions, but I'm curious as to what follows.) How could we then argue that these facts are internalist, in the strong sense that they would be completely ineffable to any rational agent who was not thereby motivated to obey the dictates of these facts? In particular, how can we demonstrate this fact non-trivially, i.e., without building 'follows the instructions of any discovered moral facts' into our definition of 'rational'? Does our concept of epistemic rationality, in itself, require that if agent A learns normative fact N (say, 'murder is wrong'), A must then eschew murder or else be guilty of e-irrationality? Clearly not. E-rationality is only about making your beliefs better fit the world. Further actions concerning those beliefs -- like following any instructions they contain, or putting them in alphabetical order, or learning the sam

0davidpearce13y

pragmatist, apologies if I gave the impression that by "impartially gives weight" I meant impartially gives equal weight. Thus the preferences of a cow or a pig or a human trump the conflicting interests of a less sentient Anopheles mosquito or a locust every time. But on the conception of rational agency I'm canvassing, it is neither epistemically nor instrumentally rational for an ideal agent to disregard a stronger preference simply because that stronger preference is entertained by a member of a another species or ethnic group. Nor is it epistemically or instrumentally rational for an ideal agent to disregard a conflicting stronger preference simply because her comparatively weaker preference looms larger in her own imagination. So on this analysis, Jane is not doing what "an ideal agent (a perfectly rational agent, with infinite computing power, etc.) would choose."

5nshepperd13y

Rationality can be used toward any goal, including goals that don't care about anyone's preference. For example, there's nothing in the math of utility maximisation that requires averaging over other agents' preferences (note: do not confuse utility maximisation with utilitarianism, they are very different things, the former being a decision theory, the latter being a specific moral philosophy).

0davidpearce13y

nshepperd, utilitarianism conceived as theory of value is not always carefully distinguished from utilitarianism - especially rule-utilitarianism - conceived as a decision procedure. This distinction is nicely brought out in the BPhil thesis of FHI's Tony Ord, "Consequentialism and Decision Procedures": http://www.amirrorclear.net/academic/papers/decision-procedures.pdf Toby takes a global utilitarian consequentialist approach to the question, 'How should I decide what to do?" - a subtly different question from '"What should I do?"

1Pablo13y

Dave might not be explaining his own position as clearly as one might wish, but I think the core of his objection is that Jane is not being epistemically rational when she decides to eat other sentient beings. This is because she is acting on a preference rooted in a belief which doesn't adequately represent certain aspects of the world--specifically, the subjective points of view of sentient members of non-human animal species. (BTW, I believe this subthread, including Eliezer's own helpful comments below, provides some evidence that Eliezer's policy of penalizing replies to comments below a certain karma threshold is misguided. EDIT: David's comment above had -6 karma when I originally posted this reply. His current score is no longer below the threshold.)

2davidpearce13y

Yes, although our conception of epistemic and instrumental rationality is certainly likely to influence our ethics, I was making a point about epistemic and instrumental rationality. Thus imagine if we lived in a era where utopian technology delivers a version of ubiquitous naturalised telepathy, so to speak. Granted such knowledge, for an agent to act in accordance with a weaker rather than stronger preference would be epistemically and instrumentally irrational. Of course, we don't (yet) live in a era of such radical transparency. But why should our current incomplete knowledge / ignorance make it instrumentally rational to fail to take into consideration what one recognises, intellectually at least, as the stronger preference? In this instance, the desire not to have one's throat slit is a very strong preference indeed. ["Replies to downvoted comments are discouraged. Pay 5 Karma points to proceed anyway?" says this Reply button. How bizarre. Is this invitation to groupthink epistemically rational? Or is killing cows good karma?]