Downvoting people for not understanding that beliefs constrain expectation I’m okay with.
Beliefs are only sometimes about anticipation. LessWrong repeatedly makes huge errors when they interpret “belief” in such a naive fashion;—giving LessWrong a semi-Bayesian justification for this collective failure of hermeneutics is unwise. Maybe beliefs “should” be about anticipation, but LessWrong, like everybody else, can’t reliably separate descriptive and normative claims, which is exactly why this “beliefs constrain anticipation” thing is misleading. …There’s a neat level-crossing thingy in there.
Downvoting people for not agreeing with EY’s moral intuitions seems… mistaken.
EY thinking of meta-ethics as a “solved problem” is one of the most obvious signs that he’s very spotty when it comes to philosophy and can’t really be trusted to do AI theory.
EY thinking of meta-ethics as a “solved problem” is one of the most obvious signs that he’s very spotty when it comes to philosophy and can’t really be trusted to do AI theory.
He does? I know he doesn’t take it as seriously as other knowledge required for AI but I didn’t think he actually thought it was a ‘solved problem’.
From my favorite post and comments section on Less Wrong thus far:
Take metaethics, a solved problem: what are the odds that someone who still thought metaethics was a Deep Mystery could write an AI algorithm that could come up with a correct metaethics?
Yes, it looks like Eliezer is mistaken there (or speaking hyperbole).
I agree with:
what are the odds that someone who still thought metaethics was a Deep Mystery could write an AI algorithm that could come up with a correct metaethics?
… but would weaken the claim drastically to “Take metaethics, a clearly reducible problem with many technical details to be ironed out”. I suspect you would disagree with even that, given that you advocate meta-ethical sentiments that I would negatively label “Deeply Mysterious”. This places me approximately equidistant from your respective positions.
I only weakly advocate certain (not formally justified) ideas about meta-ethics, and remain deeply confused about certain meta-ethical questions that I wouldn’t characterize as mere technical details. One simple example: Eliezer equates reflective consistency (a la CEV) with alignment with the big blob of computation he calls “right”; I still don’t know what argument, technical or non-technical, could justify such an intuition, and I don’t know how Eliezer would make tradeoffs if the two did in fact have different referents. This strikes me as a significant problem in itself, and there are many more problems like it.
Are you sure Eliezer does equate reflective consistency with alignment with what-he-calls-”right”? Because my recollection is that he doesn’t claim either (1) that a reflectively consistent alien mind need have values at all like what he calls right, or (2) that any individual human being, if made reflectively consistent, would necessarily end up with values much like what he calls right.
(Unless I’m awfully confused, denial of (1) is an important element in his thinking.)
I think he is defining “right” to mean something along the lines of “in line with the CEV of present-day humanity”. Maybe that’s a sensible way to use the word, maybe not (for what it’s worth, I incline towards “not”) but it isn’t the same thing as identifying “right” with “reflectively consistent”, and it doesn’t lead to a risk of confusion if the two turn out to have different referents (because they can’t).
But the key notion is the idea that what we name by ‘right’ is a fixed question, or perhaps a fixed framework. We can encounter moral arguments that modify our terminal values, and even encounter moral arguments that modify what we count as a moral argument; nonetheless, it all grows out of a particular starting point. We do not experience ourselves as embodying the question “What will I decide to do?” which would be a Type 2 calculator; anything we decided would thereby become right. We experience ourselves as asking the embodied question: “What will save my friends, and my people, from getting hurt? How can we all have more fun? …” where the ”...” is around a thousand other things.
So ‘I should X’ does not mean that I would attempt to X were I fully informed.
Aghhhh this is so confusing. Now I’m left thinking both you and Wei Dai have furnished quotes supporting my position, User:thomblake has interpreted your quote as supporting his position, and neither User:thomblake nor User:gjm have replied to Wei Dai’s quote so I don’t know if they’d interpret it as evidence of their position too! I guess I’ll just assume I’m wrong in the meantime.
Now two people have said the exact opposite things both of which disagree with me. :( Now I don’t know how to update. I plan on re-reading the relevant stuff anyway.
If you mean me and thomblake, I don’t see how we’re saying exact opposite things, or even slightly opposite things. We do both disagree with you, though.
I guess I can interpret User:thomblake two ways, but apparently my preferred way isn’t correct. Let me rephrase what you said from memory. It was like, “right is defined as the output of something like CEV, but that doesn’t mean that individuals won’t upon reflection differ substantially”. User:thomblake seemed to be saying “Eliezer doesn’t try to equate those two or define one as the other”, not “Eliezer defines right as CEV, he doesn’t equate it with CEV”. But you think User:thomblake intended the latter? Also, have I fairly characterized your position?
I don’t know whether thomblake intended the latter, but he certainly didn’t say the former. I think you said “Eliezer said A and B”, thomblake said “No he didn’t”, and you are now saying he meant “Eliezer said neither A nor B”. I suggest that he said, or at least implied, something rather like A, and would fiercely repudiate B.
Eliezer defines right as CEV, he doesn’t equate it with CEV
I definitely meant the latter, and I might be persuaded of the former.
Though “define” still seems like the wrong word. More like, ” ‘right’ is defined as *point at big blob of poetry*, and I expect it will be correctly found via the process of CEV.”—but that’s still off-the-cuff.
Thanks much; I’ll keep your opinion in mind while re-reading the meta-ethics sequence/CEV/CFAI. I might be being unduly uncharitable to Eliezer as a reaction to noticing that I was unduly (objectively-unjustifiably) trusting him. (This would have been a year or two ago.) (I notice that many people seem to unjustifiably disparage Eliezer’s ideas, but then again I notice that many people seem to unjustifiably anti-disparage (praise, re-confirm, spread) Eliezer’s ideas;—so I might be biased.)
(Really freaking drunk, apologies for errors, e.g. poltiically unmotivated adulation/anti-adulation, or excessive self-divulgation. (E.g., I suspect “divulgation” isn’t a word.))
But yeah, I just find it odd that it’s a couple of steps removed from the obvious usage. I ask myself, “Why science specifically?” and “Why public awareness rather than making the public aware?”
One simple example: Eliezer equates reflective consistency (a la CEV) with alignment with the big blob of computation he calls “right”; I still don’t know what argument, technical or non-technical, could justify such an intuition, and I don’t know how Eliezer would make tradeoffs if the two did in fact have different referents.
If I understand you correctly then this particular example I don’t think I have a problem with, at least not when I assume the kind of disclaimers and limitations of scope that I would include if I were to attempt to formally specify such a thing.
This strikes me as a significant problem in itself, and there are many more problems like it.
I suspect I agree with some of your objections to various degrees.
I thought the upshot of Eliezer’s metaethics sequence was just that “right” is a fixed abstract computation, not that it’s (the output of) some particular computation that involves simulating really smart people. CEV is not even mentioned in the sequence (EDIT: whoops it is.).
(Indeed just saying that it’s a fixed abstract computation is at the right level of abstraction to qualify as metaethics; saying that it’s some particular computation would be more like just plain ethics. The upshot does feel kind of underwhelming and obvious. This might be because I just don’t remember how confusing the issue looked before I read those posts. It could also mean that Eliezer claiming that metaethics is a solved problem is not as questionable as it might seem. And it could also mean that metaethics being solved doesn’t consitute as massive progress as it might seem.)
The upshot does feel kind of underwhelming and obvious. This might be because I just don’t remember how confusing the issue looked before I read those posts.
BTW, I’ve had numerous “wow” moments with philosophical insights, some of which made me spend years considering their implications. For example:
Bayesian interpretation of probability
AI / intelligence explosion
Tegmark’s mathematical universe
anthropic principle / anthropic reasoning
free will as the ability to decide logical facts
I expect that a correct solution to metaethics would produce a similar “wow” reaction. That is, it would be obvious in retrospect, but in an overwhelming instead of underwhelming way.
Is the insight about free will and logical facts part of the sequences? or is it something you or others discuss in a post somewhere? I’d like to learn about it, but my searches failed.
I never wrote a post on it specifically, but it’s sort of implicit in my UDT post (see also this comment). Eliezer also has a free will sequence) which is somewhat similar/related but I’m not sure if he would agree with my formulation.
“What is it that you’re deciding when you make a decision?”
What is “you”? And what is “deciding”? Personally I haven’t been able to come to any redefinition of free will that makes more sense than this one.
I haven’t read the free will sequence. And I haven’t read up on decision theory because I wasn’t sure if my math education is good enough yet. But I doubt that if I was going to read it I would learn that you can salvage the notion of “deciding” from causality and logical facts. The best you can do is look at an agent and treat it is as a transformation. But then you’d still be left with the problem of identity.
(Agreed; I also think meta-ethics and ethics are tied into each other in a way that would require that a solution to meta-ethics would at least theoretically solve any ethical problems. Given that I can think of hundreds or thousands of object level ethical problems, and given that I don’t think my inability to answer at least some of them is purely due to boundedness, fallibility, self-delusion, or ignorance as such, I don’t think I have a solution to meta-ethics. (But I would characterize my belief in God as at least a belief that meta-ethics and ethical problems do at least have some unique (meta-level) solution. This might be optimistic bias, though.))
Wei Dai, have you read the Sermon on the Mount, particularly with superintelligences, Tegmark, (epistemic or moral) credit assignment, and decision theory in mind? If not I suggest it, if only for spiritual benefits. (I suggest the Douay-Rheims translation, but that might be due to a bias towards Catholics as opposed to Protestants.)
(Pretty damn drunk for the third day in a row, apologies for errors.)
Are you planning on starting a rationalist’s drinking club? A byob lesswrong meetup with one sober note-taker? You usually do things purposefully, even if they’re unusual purposes, so consistent drunkenness seems uncharacteristic unless it’s part of a plan.
(FWIW the “post-rationalist” label isn’t my invention, I think it mostly belongs to the somewhat separate Will Ryan / Nick Tarleton / Michael Vassar / Divia / &c. crowd; I agree with Nick and Vassar way more than I agree with the LessWrong gestalt, but I’m still off on my own plot of land. Jennifer Rodriguez-Mueller could be described similarly.)
I’m pretty sure the term “rationalist’s drinking club” wouldn’t be used ingenuously as a self-description. I have noticed the justifiable use of “post-rationalist” and distance from the LW gestalt, though. I think if there were a site centered around a sequence written by Steve Rayhawk with the kind of insights into other people’s minds he regularly writes out here, with Sark and a few others as heavy contributors, that would be a “more agenty less wrong” Will would endorse. I’d actually like to see that, too.
For a human this is a much huger blob of a computation that looks like, “Did everyone survive? How many people are happy? Are people in control of their own lives? …” Humans have complex emotions, have many values—the thousand shards of desire, the godshatter of natural selection. I would say, by the way, that the huge blob of a computation is not just my present terminal values (which I don’t really have—I am not a consistent expected utility maximizers); the huge blob of a computation includes the specification of those moral arguments, those justifications, that would sway me if I heard them. So that I can regard my present values, as an approximation to the ideal morality that I would have if I heard all the arguments, to whatever extent such an extrapolation is coherent. [link in the original]
ETA: Just in case you’re right and Eliezer somehow meant for that paragraph not to be part of his metaethics, and that his actual metaethics is just “morality is a fixed abstract computation”, then I’d ask, “If morality is a fixed abstract computation, then it seems that rationality must also be a fixed abstract computation. But don’t you think a complete “solved” metaethics should explain how morality differs from rationality?”
“If morality is a fixed abstract computation, then it seems that rationality must also be a fixed abstract computation. But don’t you think a complete “solved” metaethics should explain how morality differs from rationality?”
Rationality computation outputs statements about the world, morality evaluates them. Rationality is universal and objective, so it is unique as an abstract computation, not just fixed. Morality is arbitrary.
If we assume some kind of mathematical realism (which seems to be necessary for “abstract computation” and “uniqueness” to have any meaning) then there exist objectively true statements and computations that generate them. At some point there are Goedelian problems, but at least all of the computations agree on the primitive-recursive truths, which are therefore universal, objective, unique, and true.
Any rational agent (optimization process) in any world with some regularities would exploit these regularities, which means use math. A reflective self-optimizing rational agent would arrive to the same math as us, because the math is unique.
Of course, all these points are made by a fallible human brain and so may be wrong.
But there is nothing even like that for morality. In fact, when a moral statement seems universal under sufficient reflection, it stops being a moral statement and becomes simply rational, like cooperating in the Prisoner’s Dilemma when playing against the right opponents.
But there is nothing even like that for morality. In fact, when a moral statement seems universal under sufficient reflection, it stops being a moral statement and becomes simply rational, like cooperating in the Prisoner’s Dilemma when playing against the right opponents.
What is the distinction you are making between rationality and morality, then? What makes you think the former won’t be swallowed up by the latter (or vice versa!) in the limit of infinite reflection?
(Sorta drunk, apologies for conflating conflation of rationality and morality with lack of conflation of rationality and morality, probabilistically-shouldly.)
ETA: I don’t understand how my comments can be so awesome when I’m obviously so freakin’ drunk. ;P . Maybe I should get drunk all the freakin’ time. Or study Latin all the freakin’ time, or read the Bible all the freakin’ time, or ponder how often people are obviously wrong when they use the phrase “all the freakin’ time” (let alone “freakin[‘]”) (especially when they use the phrase “all the freakin’ time” all the freakin’ time, naturally-because-reflexively)....
What is the distinction you are making between rationality and morality, then? What makes you think the former won’t be swallowed up by the latter (or vice versa!) in the limit of infinite reflection?
That was the distinction—one is universal, another arbitrary, in the limit of infinite reflection. I suppose, “there is nothing arbitrary” is a valid (consistent) position, but I don’t see any evidence for it.
Interesting! You seem to be a moral realist (cognitivist, whatever) and an a-theist. (I suspect this is the typical LessWrong position, even if the typical LessWronger isn’t as coherent as you.) I’ll take note that I should pester you and/or take care to pay attention to your opinions (comments) more in the future. Also, I thank you for showing me what the reasoning process would be that would lead one to that position. (And I think that position has a very good chance of being correct—in the absence of justifiably-ignorable inside-view (non-communicable) evidence I myself hold.)
(It’s probably obvious that I’m pretty damn drunk. (Interesting that alcohol can be just as effective as LSD or cannabis. (Still not as effective as nitrous oxide or DMT.)))
Any rational agent (optimization process) in any world with some regularities would exploit these regularities, which means use math. A reflective self-optimizing rational agent would arrive to the same math as us, because the math is unique.
Assuming it started with the same laws of inference and axioms. Also I was mostly thinking of statements about the world, e.g., physics.
Assuming it started with the same laws of inference and axioms
Or equivalent ones. But no matter where it started, it won’t arrive at different primitive-recursive truths, at least according to my brain’s current understanding.
Also I was mostly thinking of statements about the world, e.g., physics.
Is there significant difference? Wherever there are regularities in physics, there’s math (=study of regularities). Where no regularities exist, there’s no rationality.
I think the poor things are already dead. More generally, I am aware of that post, but is it relevant? The possible mind design space is of course huge and contains lots of irrational minds, but here I am arguing about universality of rationality.
But rationality is defined by external criteria—it’s about how to win (=achieve intended goals). Morality doesn’t have any such criteria. Thus, “rational minds” is a natural category. “Moral minds” is not.
Take metaethics, a solved problem: what are the odds that someone who still thought metaethics was a Deep Mystery could write an AI algorithm that could come up with a correct metaethics? I tried that, you know, and in retrospect it didn’t work.
Can you give examples of beliefs that aren’t about anticipation?
Beliefs about things that are outside our future light cone possibly qualify, to the extent that the beliefs don’t relate to things that leave historical footprints. If you’ll pardon an extreme and trite case, I would have a belief that the guy who flew the relativistic rocket out of my light cone did not cease to exist as he passed out of that cone and also did not get eaten by a giant space monster ten minutes after. My anticipations are not constrained by beliefs about either of those possibilities.
In both cases my inability to constrain my anticipated experiences speaks to my limited ability to experience and not a limitation of the universe. The same principles of ‘belief’ apply even though it has incidentally fallen out of the scope which I am able to influence or verify even in principle.
Beliefs that aren’t easily testable also tend to be the kind of beliefs that have a lot of political associations, and thus tend not to act like beliefs as such so much as policies. Also, even falsified beliefs tend to be summarily replaced with new untested/not-intended-to-be-tested beliefs, e.g. “communism is good” with “correctly implemented communism is good”, or “whites and blacks have equal average IQ” with “whites and blacks would have equal average IQ if they’d had the same cultural privileges/disadvantages”. (Apologies for the necessary political examples. Please don’t use this as an opportunity to talk about communism or race.)
Many “beliefs” that aren’t politically relevant—which excludes most scientific “knowledge” and much knowledge of your self, the people you know, what you want to do with your life, et cetera—are better characterized as knowledge, and not beliefs as such. The answers to questions like “do I have one hand, two hands, or three hands?” or “how do I get back to my house from my workplace?” aren’t generally beliefs so much as knowledge, and in my opinion “knowledge” is not only epistemologically but cognitively-neurologically a more accurate description, though I don’t really know enough about memory encoding to really back up that claim (though the difference is introspectively apparent). Either way, I still think that given our knowledge of the non-fundamental-ness of Bayes, we shouldn’t try too hard to stretch Bayes-ness to fit decision problems or cognitive algorithms that Bayes wasn’t meant to describe or solve, even if it’s technically possible to do so.
Also, even falsified beliefs tend to be summarily replaced with new untested/not-intended-to-be-tested beliefs, e.g. “communism is good” with “correctly implemented communism is good”, or “whites and blacks have equal average IQ” with “whites and blacks would have equal average IQ if they’d had the same cultural privileges/disadvantages”.
I believe the common to term for that mistake is “no true Scotsman”.
Beliefs about things that are outside our future light cone possibly qualify, to the extent that the beliefs don’t relate to things that leave historical footprints. If you’ll pardon an extreme and trite case, I would have a belief that the guy who flew the relativistic rocket out of my light cone did not cease to exist as he passed out of that cone and also did not get eaten by a giant space monster ten minutes after. My anticipations are not constrained by beliefs about either of those possibilities.
What do we lose by saying that doesn’t count as a belief? Some consistency when we describe how our minds manipulate anticipations (because we don’t separate out ones we can measure and ones we can’t, but reality does separate those, and our terminology fits reality)? Something else?
I’m not clear on the relevance of caring to beliefs. I would prefer that those I care about not be tortured, but once they’re out of my future light cone whatever happens to them is a sunk cost- I don’t see what I (or they) get from my preferring or believing things about them.
Oops, I just realized that in my hypothetical scenario by someone being tortured outside your light cone, I meant someone being tortured somewhere your two future light cones don’t intersect.
Indeed; being outside of my future light cone just means whatever I do has no impact on them. But now not only can I not impact them, but they’re also dead to me (as they, or any information they emit, won’t exist in my future). I still don’t see what impact caring about them has.
Right, but for my actions to have an effect on them, they have to be in my future light cone at the time of action. It sounds like you’re interested in events in my future light cone but will not be in any of the past light cones centered at my future intervals- like, for example, things that I can set in motion now which will not come to fruition until after I’m dead, or the person I care about pondering whether or not to jump into a black hole. Those things are worth caring about so long as they’re in my future light cone, and it’s meaningful to have beliefs about them to the degree that they could be in my past light cone in the future.
The best illustration I’ve seen thus far is this one.
(Side note: I desire few things more than a community where people automatically and regularly engage in analyses like the one linked to. Such a community would actually be significantly less wrong than any community thus far seen on Earth. When LessWrong tries to engage in causal analyses of why others believe what they believe it’s usually really bad: proffered explanations are variations on “memetic selection pressures”, “confirmation bias”, or other fully general “explanations”/rationalizations. I think this in itself is a damning critique of LessWrong, and I think some of the attitude that promotes such ignorance of the causes of others’ beliefs is apparent in posts like “Our Phyg Is Not Exclusive Enough”.)
I agree that that post is the sort of thing that I want more of on LW.
It seems to me like Steve_Rayhawk’s comment is all about anticipation- I hold position X because I anticipate it will have Y impact on the future. But I think I see the disconnect you’re talking about- the position one takes on global warming is based on anticipations one has about politics, not the climate, but it’s necessary (and/or reduces cognitive dissonance) to state the political position in terms of anticipations one has about the climate.
I don’t think public stated beliefs have to be about anticipation- but I do think that private beliefs have to be (should be?) about anticipation. I also think I’m much more sympathetic to the view that rationalizations can use the “beliefs are anticipation” argument as a weapon without finding the true anticipations in question (like Steve_Rayhawk did), but I don’t think that implies that “beliefs are anticipation” is naive or incorrect. Separating out positions, identities, and beliefs seems more helpful than overloading the world beliefs.
it’s necessary (and/or reduces cognitive dissonance) to state the political position in terms of anticipations one has about the climate.
I don’t think public stated beliefs have to be about anticipation
You seem to be modeling the AGW disputant’s decision policy as if he is internally representing, in a way that would be introspectively clear to him, his belief about AGW and his public stance about AGW as explicitly distinguished nodes;—as opposed to having “actual belief about AGW” as a latent node that isn’t introspectively accessible. That’s surely the case sometimes, but I don’t think that’s usually the case. Given the non-distinguishability of beliefs and preferences (and the theoretical non-unique-decomposability (is there a standard economic term for that?) of decision policies) I’m not sure it’s wise to use “belief” to refer to only the (in many cases unidentifiable) “actual anticipation” part of decision policies, either for others or ourselves, especially when we don’t have enough time to be abnormally reflective about the causes and purposes of others’/our “beliefs”.
(Areas where such caution isn’t as necessary are e.g. decision science modeling of simple rational agents, or largescale economic models. But if you want to model actual people’s policies in complex situations then the naive Bayesian approach (e.g. with influence diagrams) doesn’t work or is way too cumbersome. Does your experience differ from mine? You have a lot more modeling experience than I do. Also I get the impression that Steve disagrees with me at least a little bit, and his opinion is worth a lot more than mine.)
Another more theoretical reason I encourage caution about the “belief as anticipation” idea is that I don’t think it correctly characterizes the nature of belief in light of recent ideas in decision theory. To me, beliefs seem to be about coordination, where your choice of belief (e.g. expecting a squared rather than a cubed modulus Born rule) is determined by the innate preference (drilled into you by ecological contingencies and natural selection) to coordinate your actions with the actions and decision policies of the agents around you, and where your utility function is about self-coordination (e.g. for purposes of dynamic consistency). The ‘pure’ “anticipation” aspect of beliefs only seems relevant in certain cases, e.g. when you don’t have “anthropic” uncertainty (e.g. uncertainty about the extent to which your contexts are ambiently determined by your decision policy). Unfortunately people like me always have a substantial amount of “anthropic” uncertainty, and it’s mostly only in counterfactual/toy problems where I can use the naive Bayesian approach to epistemology.
(Note that taking the general decision theoretic perspective doesn’t lead to wacky quantum-suicide-like implications, otherwise I would be a lot more skeptical about the prudence of partially ditching the Bayesian boat.)
You seem to be modeling the AGW disputant’s decision policy as if he is internally representing, in a way that would be introspectively clear to him, his belief about AGW and his public stance about AGW as explicitly distinguished nodes;—as opposed to having “actual belief about AGW” as a latent node that isn’t introspectively accessible.
I’m describing it that way but I don’t think the introspection is necessary- it’s just easier to talk about as if he had full access to his mind. (Private beliefs don’t have to be beliefs that the mind’s narrator has access to, and oftentimes are kept out of its reach for security purposes!)
But if you want to model actual people’s policies in complex situations then the naive Bayesian approach (e.g. with influence diagrams) doesn’t work or is way too cumbersome. Does your experience differ from mine?
I don’t think I’ve seen any Bayesian modeling of that sort of thing, but I haven’t gone looking for it.
Bayes nets in general are difficult for people, rather than computers, to manipulate, and so it’s hard to decide what makes them too cumbersome. (Bayes nets in industrial use, like for fault diagnostics, tend to have hundreds if not thousands of nodes, but you wouldn’t have a person traverse them unaided.)
If you wanted to code a narrow AI that determined someone’s mood by, say, webcam footage of them, I think putting your perception data into a Bayes net would be a common approach.
Political positions / psychology seem tough. I could see someone do belief-mapping and correlation in a useful way, but I don’t see analysis on the level of Steve_Rayhawk’s post coming out of a computer-run Bayes net anytime soon, and I don’t think drawing out a Bayes net would help significantly with that sort of analysis. Possible but unlikely- we’ve got pretty sophisticated dedicated hardware for very similar things.
Another more theoretical reason I encourage caution about the “belief as anticipation” idea is that I don’t think it correctly characterizes the nature of belief in light of recent ideas in decision theory. To me, beliefs seem to be about coordination
Hmm. I’m going to need to sleep on this, but this sort of coordination still smells to me like anticipation.
(A general comment: this conversation has moved me towards thinking that it’s useful for the LW norm to be tabooing “belief” and using “anticipation” instead when appropriate, rather than trying to equate the two terms. I don’t know if you’re advocating for tabooing “belief”, though.)
(Complement to my other reply: You might not have seen this comment, where I suggest “knowledge” as a better descriptor than “belief” in most mundane settings. (Also I suspect that people’s uses of the words “think” versus “believe” are correlated with introspectively distinct kinds of uncertainty.))
Don’t my beliefs about primordial cows constrain my anticipation of the fossil record and development of contemporary species?
I think “most people’s beliefs” fit the anticipation framework- so long as you express them in a compartmentalized fashion, and my understanding of the point of the ‘belief=anticipation’ approach is that it helps resist compartmentalization, which is generally positive.
Beliefs are only sometimes about anticipation. LessWrong repeatedly makes huge errors when they interpret “belief” in such a naive fashion;—giving LessWrong a semi-Bayesian justification for this collective failure of hermeneutics is unwise. Maybe beliefs “should” be about anticipation, but LessWrong, like everybody else, can’t reliably separate descriptive and normative claims, which is exactly why this “beliefs constrain anticipation” thing is misleading. …There’s a neat level-crossing thingy in there.
EY thinking of meta-ethics as a “solved problem” is one of the most obvious signs that he’s very spotty when it comes to philosophy and can’t really be trusted to do AI theory.
(Apologies if I come across as curmudgeonly.)
He does? I know he doesn’t take it as seriously as other knowledge required for AI but I didn’t think he actually thought it was a ‘solved problem’.
From my favorite post and comments section on Less Wrong thus far:
Yes, it looks like Eliezer is mistaken there (or speaking hyperbole).
I agree with:
… but would weaken the claim drastically to “Take metaethics, a clearly reducible problem with many technical details to be ironed out”. I suspect you would disagree with even that, given that you advocate meta-ethical sentiments that I would negatively label “Deeply Mysterious”. This places me approximately equidistant from your respective positions.
I only weakly advocate certain (not formally justified) ideas about meta-ethics, and remain deeply confused about certain meta-ethical questions that I wouldn’t characterize as mere technical details. One simple example: Eliezer equates reflective consistency (a la CEV) with alignment with the big blob of computation he calls “right”; I still don’t know what argument, technical or non-technical, could justify such an intuition, and I don’t know how Eliezer would make tradeoffs if the two did in fact have different referents. This strikes me as a significant problem in itself, and there are many more problems like it.
(Mildly inebriated, apologies for errors.)
Are you sure Eliezer does equate reflective consistency with alignment with what-he-calls-”right”? Because my recollection is that he doesn’t claim either (1) that a reflectively consistent alien mind need have values at all like what he calls right, or (2) that any individual human being, if made reflectively consistent, would necessarily end up with values much like what he calls right.
(Unless I’m awfully confused, denial of (1) is an important element in his thinking.)
I think he is defining “right” to mean something along the lines of “in line with the CEV of present-day humanity”. Maybe that’s a sensible way to use the word, maybe not (for what it’s worth, I incline towards “not”) but it isn’t the same thing as identifying “right” with “reflectively consistent”, and it doesn’t lead to a risk of confusion if the two turn out to have different referents (because they can’t).
He most certainly does not.
Relevant quote from Morality as Fixed Computation:
Thanks—I hope you’re providing that as evidence for my point.
Sort of. It certainly means he doesn’t define morality as extrapolated volition. (But maybe “equate” meant something looser than that?)
Aghhhh this is so confusing. Now I’m left thinking both you and Wei Dai have furnished quotes supporting my position, User:thomblake has interpreted your quote as supporting his position, and neither User:thomblake nor User:gjm have replied to Wei Dai’s quote so I don’t know if they’d interpret it as evidence of their position too! I guess I’ll just assume I’m wrong in the meantime.
Now two people have said the exact opposite things both of which disagree with me. :( Now I don’t know how to update. I plan on re-reading the relevant stuff anyway.
If you mean me and thomblake, I don’t see how we’re saying exact opposite things, or even slightly opposite things. We do both disagree with you, though.
I guess I can interpret User:thomblake two ways, but apparently my preferred way isn’t correct. Let me rephrase what you said from memory. It was like, “right is defined as the output of something like CEV, but that doesn’t mean that individuals won’t upon reflection differ substantially”. User:thomblake seemed to be saying “Eliezer doesn’t try to equate those two or define one as the other”, not “Eliezer defines right as CEV, he doesn’t equate it with CEV”. But you think User:thomblake intended the latter? Also, have I fairly characterized your position?
I don’t know whether thomblake intended the latter, but he certainly didn’t say the former. I think you said “Eliezer said A and B”, thomblake said “No he didn’t”, and you are now saying he meant “Eliezer said neither A nor B”. I suggest that he said, or at least implied, something rather like A, and would fiercely repudiate B.
I definitely meant the latter, and I might be persuaded of the former.
Though “define” still seems like the wrong word. More like, ” ‘right’ is defined as *point at big blob of poetry*, and I expect it will be correctly found via the process of CEV.”—but that’s still off-the-cuff.
Thanks much; I’ll keep your opinion in mind while re-reading the meta-ethics sequence/CEV/CFAI. I might be being unduly uncharitable to Eliezer as a reaction to noticing that I was unduly (objectively-unjustifiably) trusting him. (This would have been a year or two ago.) (I notice that many people seem to unjustifiably disparage Eliezer’s ideas, but then again I notice that many people seem to unjustifiably anti-disparage (praise, re-confirm, spread) Eliezer’s ideas;—so I might be biased.)
(Really freaking drunk, apologies for errors, e.g. poltiically unmotivated adulation/anti-adulation, or excessive self-divulgation. (E.g., I suspect “divulgation” isn’t a word.))
Not to worry, it means “The act of divulging” or else “public awareness of science” (oddly).
I mean, it’s not so odd. di-vulgar-tion; the result of making public (something).
Well,
divulge
divulgate
divulgation
But yeah, I just find it odd that it’s a couple of steps removed from the obvious usage. I ask myself, “Why science specifically?” and “Why public awareness rather than making the public aware?”
If I understand you correctly then this particular example I don’t think I have a problem with, at least not when I assume the kind of disclaimers and limitations of scope that I would include if I were to attempt to formally specify such a thing.
I suspect I agree with some of your objections to various degrees.
Part of my concern about Eliezer trying to build FAI also stems from his treatment of metaethics. Here’s a caricature of how his solution looks to me:
Alice: Hey, what is the value of X?
Bob: Hmm, I don’t know. Actually I’m not even sure what it means to answer that question. What’s the definition of X?
Alice: I don’t know how to define it either.
Bob: Ok… I don’t know how to answer your question, but what if we simulate a bunch of really smart people and ask them what the value of X is?
Alice: Great idea! But what about the definition of X? I feel like we ought to be able to at least answer that now...
Bob: Oh that’s easy. Let’s just define it as the output of that computation I just mentioned.
I thought the upshot of Eliezer’s metaethics sequence was just that “right” is a fixed abstract computation, not that it’s (the output of) some particular computation that involves simulating really smart people. CEV is not even mentioned in the sequence (EDIT: whoops it is.).
(Indeed just saying that it’s a fixed abstract computation is at the right level of abstraction to qualify as metaethics; saying that it’s some particular computation would be more like just plain ethics. The upshot does feel kind of underwhelming and obvious. This might be because I just don’t remember how confusing the issue looked before I read those posts. It could also mean that Eliezer claiming that metaethics is a solved problem is not as questionable as it might seem. And it could also mean that metaethics being solved doesn’t consitute as massive progress as it might seem.)
BTW, I’ve had numerous “wow” moments with philosophical insights, some of which made me spend years considering their implications. For example:
Bayesian interpretation of probability
AI / intelligence explosion
Tegmark’s mathematical universe
anthropic principle / anthropic reasoning
free will as the ability to decide logical facts
I expect that a correct solution to metaethics would produce a similar “wow” reaction. That is, it would be obvious in retrospect, but in an overwhelming instead of underwhelming way.
Is the insight about free will and logical facts part of the sequences? or is it something you or others discuss in a post somewhere? I’d like to learn about it, but my searches failed.
I never wrote a post on it specifically, but it’s sort of implicit in my UDT post (see also this comment). Eliezer also has a free will sequence) which is somewhat similar/related but I’m not sure if he would agree with my formulation.
What is “you”? And what is “deciding”? Personally I haven’t been able to come to any redefinition of free will that makes more sense than this one.
I haven’t read the free will sequence. And I haven’t read up on decision theory because I wasn’t sure if my math education is good enough yet. But I doubt that if I was going to read it I would learn that you can salvage the notion of “deciding” from causality and logical facts. The best you can do is look at an agent and treat it is as a transformation. But then you’d still be left with the problem of identity.
(Agreed; I also think meta-ethics and ethics are tied into each other in a way that would require that a solution to meta-ethics would at least theoretically solve any ethical problems. Given that I can think of hundreds or thousands of object level ethical problems, and given that I don’t think my inability to answer at least some of them is purely due to boundedness, fallibility, self-delusion, or ignorance as such, I don’t think I have a solution to meta-ethics. (But I would characterize my belief in God as at least a belief that meta-ethics and ethical problems do at least have some unique (meta-level) solution. This might be optimistic bias, though.))
Wei Dai, have you read the Sermon on the Mount, particularly with superintelligences, Tegmark, (epistemic or moral) credit assignment, and decision theory in mind? If not I suggest it, if only for spiritual benefits. (I suggest the Douay-Rheims translation, but that might be due to a bias towards Catholics as opposed to Protestants.)
(Pretty damn drunk for the third day in a row, apologies for errors.)
Are you planning on starting a rationalist’s drinking club? A byob lesswrong meetup with one sober note-taker? You usually do things purposefully, even if they’re unusual purposes, so consistent drunkenness seems uncharacteristic unless it’s part of a plan.
Will_Newsome isn’t a rationalist. (He has described himself as a ‘post-rationalist’, which seems as good a term as any.)
(FWIW the “post-rationalist” label isn’t my invention, I think it mostly belongs to the somewhat separate Will Ryan / Nick Tarleton / Michael Vassar / Divia / &c. crowd; I agree with Nick and Vassar way more than I agree with the LessWrong gestalt, but I’m still off on my own plot of land. Jennifer Rodriguez-Mueller could be described similarly.)
I’m pretty sure the term “rationalist’s drinking club” wouldn’t be used ingenuously as a self-description. I have noticed the justifiable use of “post-rationalist” and distance from the LW gestalt, though. I think if there were a site centered around a sequence written by Steve Rayhawk with the kind of insights into other people’s minds he regularly writes out here, with Sark and a few others as heavy contributors, that would be a “more agenty less wrong” Will would endorse. I’d actually like to see that, too.
In vino veritas et sanitas!
It’s mentioned here:
ETA: Just in case you’re right and Eliezer somehow meant for that paragraph not to be part of his metaethics, and that his actual metaethics is just “morality is a fixed abstract computation”, then I’d ask, “If morality is a fixed abstract computation, then it seems that rationality must also be a fixed abstract computation. But don’t you think a complete “solved” metaethics should explain how morality differs from rationality?”
Rationality computation outputs statements about the world, morality evaluates them. Rationality is universal and objective, so it is unique as an abstract computation, not just fixed. Morality is arbitrary.
How so? Every argument I’ve heard for why morality is arbitrary applies just as well to rationality.
If we assume some kind of mathematical realism (which seems to be necessary for “abstract computation” and “uniqueness” to have any meaning) then there exist objectively true statements and computations that generate them. At some point there are Goedelian problems, but at least all of the computations agree on the primitive-recursive truths, which are therefore universal, objective, unique, and true.
Any rational agent (optimization process) in any world with some regularities would exploit these regularities, which means use math. A reflective self-optimizing rational agent would arrive to the same math as us, because the math is unique.
Of course, all these points are made by a fallible human brain and so may be wrong.
But there is nothing even like that for morality. In fact, when a moral statement seems universal under sufficient reflection, it stops being a moral statement and becomes simply rational, like cooperating in the Prisoner’s Dilemma when playing against the right opponents.
What is the distinction you are making between rationality and morality, then? What makes you think the former won’t be swallowed up by the latter (or vice versa!) in the limit of infinite reflection?
(Sorta drunk, apologies for conflating conflation of rationality and morality with lack of conflation of rationality and morality, probabilistically-shouldly.)
ETA: I don’t understand how my comments can be so awesome when I’m obviously so freakin’ drunk. ;P . Maybe I should get drunk all the freakin’ time. Or study Latin all the freakin’ time, or read the Bible all the freakin’ time, or ponder how often people are obviously wrong when they use the phrase “all the freakin’ time” (let alone “freakin[‘]”) (especially when they use the phrase “all the freakin’ time” all the freakin’ time, naturally-because-reflexively)....
That was the distinction—one is universal, another arbitrary, in the limit of infinite reflection. I suppose, “there is nothing arbitrary” is a valid (consistent) position, but I don’t see any evidence for it.
Interesting! You seem to be a moral realist (cognitivist, whatever) and an a-theist. (I suspect this is the typical LessWrong position, even if the typical LessWronger isn’t as coherent as you.) I’ll take note that I should pester you and/or take care to pay attention to your opinions (comments) more in the future. Also, I thank you for showing me what the reasoning process would be that would lead one to that position. (And I think that position has a very good chance of being correct—in the absence of justifiably-ignorable inside-view (non-communicable) evidence I myself hold.)
(It’s probably obvious that I’m pretty damn drunk. (Interesting that alcohol can be just as effective as LSD or cannabis. (Still not as effective as nitrous oxide or DMT.)))
Cognitivist yes, moral realist, no. IIUC, it’s EY’s position (“morality is a computation”), so naturally it’s the typical LessWrong position.
Universally valid statements must have universally-available evidence, no?
Really nothing like LSD, which makes it impossible to write anything at all, at least for me.
Assuming it started with the same laws of inference and axioms. Also I was mostly thinking of statements about the world, e.g., physics.
Or equivalent ones. But no matter where it started, it won’t arrive at different primitive-recursive truths, at least according to my brain’s current understanding.
Is there significant difference? Wherever there are regularities in physics, there’s math (=study of regularities). Where no regularities exist, there’s no rationality.
What about the poor beings with an anti-iductive prior? More generally read this post by Eliezer.
I think the poor things are already dead. More generally, I am aware of that post, but is it relevant? The possible mind design space is of course huge and contains lots of irrational minds, but here I am arguing about universality of rationality.
My point, as I stated above, is that every argument I’ve heard against universality of morality applies just as well to rationality.
I agree with your statement:
I would also agree with the following:
The possible mind design space is of course huge and contains lots of immoral minds, but here I am arguing about universality of morality.
But rationality is defined by external criteria—it’s about how to win (=achieve intended goals). Morality doesn’t have any such criteria. Thus, “rational minds” is a natural category. “Moral minds” is not.
Yeah: CEV appears to just move the hard bit. Adding another layer of indirection.
To take Eliezer’s statement one meta-level down:
What did he mean by “I tried that...”?
I’m not at all sure, but I think he means CFAI.
Possibly he means this.
He may have soleved it, but if only he or someone else could say what the solution was.
Can you give examples of beliefs that aren’t about anticipation?
Beliefs about things that are outside our future light cone possibly qualify, to the extent that the beliefs don’t relate to things that leave historical footprints. If you’ll pardon an extreme and trite case, I would have a belief that the guy who flew the relativistic rocket out of my light cone did not cease to exist as he passed out of that cone and also did not get eaten by a giant space monster ten minutes after. My anticipations are not constrained by beliefs about either of those possibilities.
In both cases my inability to constrain my anticipated experiences speaks to my limited ability to experience and not a limitation of the universe. The same principles of ‘belief’ apply even though it has incidentally fallen out of the scope which I am able to influence or verify even in principle.
Beliefs that aren’t easily testable also tend to be the kind of beliefs that have a lot of political associations, and thus tend not to act like beliefs as such so much as policies. Also, even falsified beliefs tend to be summarily replaced with new untested/not-intended-to-be-tested beliefs, e.g. “communism is good” with “correctly implemented communism is good”, or “whites and blacks have equal average IQ” with “whites and blacks would have equal average IQ if they’d had the same cultural privileges/disadvantages”. (Apologies for the necessary political examples. Please don’t use this as an opportunity to talk about communism or race.)
Many “beliefs” that aren’t politically relevant—which excludes most scientific “knowledge” and much knowledge of your self, the people you know, what you want to do with your life, et cetera—are better characterized as knowledge, and not beliefs as such. The answers to questions like “do I have one hand, two hands, or three hands?” or “how do I get back to my house from my workplace?” aren’t generally beliefs so much as knowledge, and in my opinion “knowledge” is not only epistemologically but cognitively-neurologically a more accurate description, though I don’t really know enough about memory encoding to really back up that claim (though the difference is introspectively apparent). Either way, I still think that given our knowledge of the non-fundamental-ness of Bayes, we shouldn’t try too hard to stretch Bayes-ness to fit decision problems or cognitive algorithms that Bayes wasn’t meant to describe or solve, even if it’s technically possible to do so.
I believe the common to term for that mistake is “no true Scotsman”.
What do we lose by saying that doesn’t count as a belief? Some consistency when we describe how our minds manipulate anticipations (because we don’t separate out ones we can measure and ones we can’t, but reality does separate those, and our terminology fits reality)? Something else?
So if someone you cared about is leaving your future light cone, you wouldn’t care if he gets horribly tortured as soon as he’s outside of it?
I’m not clear on the relevance of caring to beliefs. I would prefer that those I care about not be tortured, but once they’re out of my future light cone whatever happens to them is a sunk cost- I don’t see what I (or they) get from my preferring or believing things about them.
Yes, but you can affect what happens to them before they leave.
Before they leave, their torture would be in my future light cone, right?
Oops, I just realized that in my hypothetical scenario by someone being tortured outside your light cone, I meant someone being tortured somewhere your two future light cones don’t intersect.
Indeed; being outside of my future light cone just means whatever I do has no impact on them. But now not only can I not impact them, but they’re also dead to me (as they, or any information they emit, won’t exist in my future). I still don’t see what impact caring about them has.
Ok, my scenario involves your actions having an effect on them before your two light cones become disjoint.
Right, but for my actions to have an effect on them, they have to be in my future light cone at the time of action. It sounds like you’re interested in events in my future light cone but will not be in any of the past light cones centered at my future intervals- like, for example, things that I can set in motion now which will not come to fruition until after I’m dead, or the person I care about pondering whether or not to jump into a black hole. Those things are worth caring about so long as they’re in my future light cone, and it’s meaningful to have beliefs about them to the degree that they could be in my past light cone in the future.
The best illustration I’ve seen thus far is this one.
(Side note: I desire few things more than a community where people automatically and regularly engage in analyses like the one linked to. Such a community would actually be significantly less wrong than any community thus far seen on Earth. When LessWrong tries to engage in causal analyses of why others believe what they believe it’s usually really bad: proffered explanations are variations on “memetic selection pressures”, “confirmation bias”, or other fully general “explanations”/rationalizations. I think this in itself is a damning critique of LessWrong, and I think some of the attitude that promotes such ignorance of the causes of others’ beliefs is apparent in posts like “Our Phyg Is Not Exclusive Enough”.)
I agree that that post is the sort of thing that I want more of on LW.
It seems to me like Steve_Rayhawk’s comment is all about anticipation- I hold position X because I anticipate it will have Y impact on the future. But I think I see the disconnect you’re talking about- the position one takes on global warming is based on anticipations one has about politics, not the climate, but it’s necessary (and/or reduces cognitive dissonance) to state the political position in terms of anticipations one has about the climate.
I don’t think public stated beliefs have to be about anticipation- but I do think that private beliefs have to be (should be?) about anticipation. I also think I’m much more sympathetic to the view that rationalizations can use the “beliefs are anticipation” argument as a weapon without finding the true anticipations in question (like Steve_Rayhawk did), but I don’t think that implies that “beliefs are anticipation” is naive or incorrect. Separating out positions, identities, and beliefs seems more helpful than overloading the world beliefs.
You seem to be modeling the AGW disputant’s decision policy as if he is internally representing, in a way that would be introspectively clear to him, his belief about AGW and his public stance about AGW as explicitly distinguished nodes;—as opposed to having “actual belief about AGW” as a latent node that isn’t introspectively accessible. That’s surely the case sometimes, but I don’t think that’s usually the case. Given the non-distinguishability of beliefs and preferences (and the theoretical non-unique-decomposability (is there a standard economic term for that?) of decision policies) I’m not sure it’s wise to use “belief” to refer to only the (in many cases unidentifiable) “actual anticipation” part of decision policies, either for others or ourselves, especially when we don’t have enough time to be abnormally reflective about the causes and purposes of others’/our “beliefs”.
(Areas where such caution isn’t as necessary are e.g. decision science modeling of simple rational agents, or largescale economic models. But if you want to model actual people’s policies in complex situations then the naive Bayesian approach (e.g. with influence diagrams) doesn’t work or is way too cumbersome. Does your experience differ from mine? You have a lot more modeling experience than I do. Also I get the impression that Steve disagrees with me at least a little bit, and his opinion is worth a lot more than mine.)
Another more theoretical reason I encourage caution about the “belief as anticipation” idea is that I don’t think it correctly characterizes the nature of belief in light of recent ideas in decision theory. To me, beliefs seem to be about coordination, where your choice of belief (e.g. expecting a squared rather than a cubed modulus Born rule) is determined by the innate preference (drilled into you by ecological contingencies and natural selection) to coordinate your actions with the actions and decision policies of the agents around you, and where your utility function is about self-coordination (e.g. for purposes of dynamic consistency). The ‘pure’ “anticipation” aspect of beliefs only seems relevant in certain cases, e.g. when you don’t have “anthropic” uncertainty (e.g. uncertainty about the extent to which your contexts are ambiently determined by your decision policy). Unfortunately people like me always have a substantial amount of “anthropic” uncertainty, and it’s mostly only in counterfactual/toy problems where I can use the naive Bayesian approach to epistemology.
(Note that taking the general decision theoretic perspective doesn’t lead to wacky quantum-suicide-like implications, otherwise I would be a lot more skeptical about the prudence of partially ditching the Bayesian boat.)
I’m describing it that way but I don’t think the introspection is necessary- it’s just easier to talk about as if he had full access to his mind. (Private beliefs don’t have to be beliefs that the mind’s narrator has access to, and oftentimes are kept out of its reach for security purposes!)
I don’t think I’ve seen any Bayesian modeling of that sort of thing, but I haven’t gone looking for it.
Bayes nets in general are difficult for people, rather than computers, to manipulate, and so it’s hard to decide what makes them too cumbersome. (Bayes nets in industrial use, like for fault diagnostics, tend to have hundreds if not thousands of nodes, but you wouldn’t have a person traverse them unaided.)
If you wanted to code a narrow AI that determined someone’s mood by, say, webcam footage of them, I think putting your perception data into a Bayes net would be a common approach.
Political positions / psychology seem tough. I could see someone do belief-mapping and correlation in a useful way, but I don’t see analysis on the level of Steve_Rayhawk’s post coming out of a computer-run Bayes net anytime soon, and I don’t think drawing out a Bayes net would help significantly with that sort of analysis. Possible but unlikely- we’ve got pretty sophisticated dedicated hardware for very similar things.
Hmm. I’m going to need to sleep on this, but this sort of coordination still smells to me like anticipation.
(A general comment: this conversation has moved me towards thinking that it’s useful for the LW norm to be tabooing “belief” and using “anticipation” instead when appropriate, rather than trying to equate the two terms. I don’t know if you’re advocating for tabooing “belief”, though.)
(Complement to my other reply: You might not have seen this comment, where I suggest “knowledge” as a better descriptor than “belief” in most mundane settings. (Also I suspect that people’s uses of the words “think” versus “believe” are correlated with introspectively distinct kinds of uncertainty.))
Beliefs about primordial cows, etc. Most people’s beliefs. He’s talking descriptively, not normatively.
Don’t my beliefs about primordial cows constrain my anticipation of the fossil record and development of contemporary species?
I think “most people’s beliefs” fit the anticipation framework- so long as you express them in a compartmentalized fashion, and my understanding of the point of the ‘belief=anticipation’ approach is that it helps resist compartmentalization, which is generally positive.