pjeby comments on Post Your Utility Function - Less Wrong

28 Post author: taw 04 June 2009 05:05AM

You are viewing a comment permalink. View the original post to see all comments and the full post content.

Article Navigation

Comments (273)

Sort By: Best

Show:

You are viewing a single comment's thread. Show more comments above.

Comment author: pjeby 05 June 2009 10:21:08PM -1 points [-]

Don't take it lightly, it's a well-vetted and well-understood position, extensively discussed and agreed upon.

It's extensively discussed and agreed upon, that that is how we (for certain definitions of "we") would like it to be, and it certainly has desirable properties for say, building Friendly AI, or any AI that doesn't wirehead. And it is certainly a property of the human brain that it orients its preferences towards what it believes is the outside world - again, it has good consequences for preventing wireheading.

But that doesn't make it actually true, just useful.

It's also pretty well established as a tenet of e.g., General Semantics, that the "outside world" is unknowable, since all we can ever consciously perceive is our map. The whole point of discussing biases is that our maps are systematically biased -- and this includes our preferences, which are being applied to our biased views of the world, rather than the actual world.

I am being descriptive here, not prescriptive. When we say we prefer a certain set of things to actually be true, we can only mean that we want the world to not dispute a certain map, because otherwise we are making the supernaturalist error of assuming that a thing could be true independent of the components that make it so.

To put it another way, if I say, "I prefer that the wings of this plane not fall off", I am speaking about the map, since "wings" do not exist in the territory.

IOW, our statements about reality are about the intersection of some portion of "observable" reality and our particular mapping (division and labeling) of it. And it cannot be otherwise, since to even talk about it, we have to carve up and label the "reality" we are discussing.

Comment author: loqi 07 June 2009 07:12:56AM * 0 points [-]

IOW, our statements about reality are about the intersection of some portion of "observable" reality and our particular mapping (division and labeling) of it. And it cannot be otherwise, since to even talk about it, we have to carve up and label the "reality" we are discussing.

It's funny that you talk of wordplay a few comments back, as it seems that you're the one making a technically-correct-but-not-practically-meaningful argument here.

If I may attempt to explore your position: Suppose someone claims a preference for "blue skies". The wirehead version of this that you endorse is "I prefer experiences that include the perception I label 'blue sky'". The "anti-wirehead" version you seem to be arguing against is "I prefer actual states of the world where the sky is actually blue".

You seem to be saying that since the preference is really about the experience of blue skies, it makes no sense to talk about the sky actually being blue. Chasing after external definitions involving photons and atmospheric scattering is beside the point, because the actual preference wasn't formed in terms of them.

This becomes another example of the general rule that it's impossible to form preferences directly about reality, because "reality" is just another label on our subjective map.

As far as specifics go, I think the point you make is sound: Most (all?) of our preferences can't just be about the territory, because they're phrased in terms of things that themselves don't exist in the territory, but at best simply point at the slice of experience labeled "the territory".

That said, I think this perspective grossly downplays the practical importance of that label. It has very distinct subjective features connecting in special ways to other important concepts. For the non-solipsists among us, perhaps the most important role it plays is establishing a connection between our subjective reality and someone else's. We have reason to believe that it mediates experiences we label as "physical interactions" in a manner causally unaffected by our state of mind alone.

When I say "I prefer the galaxy not to be tiled by paperclips", I understand that, technically, the only building blocks I have for that preference are labeled experiences and concepts that aren't themselves the "stuff" of their referents. In fact, I freely admit that I'm not exactly sure what constitutes "the galaxy", but the preference I just expressed actually contains a massive number of implicit references to other concepts that I consider causally connected to it via my "external reality" label. What's more, most people I communicate with can easily access a seemingly similar set of connections to their "external reality" label, assuming they don't talk themselves out of it.

The territory concept plays a similar role to that of an opaque reference in a programming language. Its state may not be invariant, but its identity is. I don't have to know any true facts concerning its actual structure for it to be meaningful and useful. Just as photons aren't explicitly required to subjectively perceive a blue sky, the ontological status of my territory concept doesn't really change its meaning or importance, which is acquired through its intimate connection to massive amounts of raw experience.

Claiming my preferences about the territory are really just about my map is true in the narrow technical sense that it's impossible for me to refer directly to "reality", but doing so completely glosses over the deep, implicit connections expressed by such preferences, most primarily the connection between myself and the things I label "other consciousnesses". In contrast, the perception of these connections seems to come for free by "confusing" the invariant identity of my territory concept with the invariant "existence" of a real external world. The two notions are basically isomorphic, so where's the value in the distinction?

Comment author: pjeby comment score below threshold [+] (21 children)

Comment author: pjeby 07 June 2009 06:25:29PM -4 points [-]

You seem to be saying that since the preference is really about the experience of blue skies, it makes no sense to talk about the sky actually being blue.

That depends on whether you're talking about "blue" in terms of human experience, or whether you're talking about wavelengths of light. The former is clearly "map", whereas discussing wavelengths of light at least might be considered "about" the territory in some sense.

However, if you are talking about preferences, I don't think there's any way for a preference to escape the map. We can define how the mapped preference relates to some mapped portion of the territory, but preferences (on human hardware at least) can only be "about" experience.

And that's because all of our hypothetical preferences for states of the actual world are being modeled as experiences in order to compute a preference.

The entire context of this discussion (for me, anyway) has been about preferences. That doesn't preclude the possibility of relating maps to territory, or maps to other people's maps. I'm just saying that human beings have no way to model their preferences except by modeling experiences, in human-sensory terms.

The "anti-wirehead" version you seem to be arguing against is "I prefer actual states of the world where the sky is actually blue".

But when this statement is uttered by a human, it is almost invariably a lie. Because inside, the person's model is an experience of "blue sky". The preference is about experiences, not the state of the world.

Even if you phrase this as, "I prefer the sky to be actually blue, even if I don't know it", it is still a lie, because now you are modeling an experience of the sky being blue, plus an experience of you not knowing it.

The two notions are basically isomorphic, so where's the value in the distinction?

Well, it makes clear some of the limits of certain endeavors that are often discussed here. It dissolves confusions about the best ways to make people happy, and whether a world should be considered "real" or "virtual", and whether it's somehow "bad" to be virtual.

But the most important practical benefit is that it helps in understanding why sometimes the best thing for a person may be to update their preferences to match the constraints of reality, rather than trying in vain to make reality fit their preferences.

Consider the "blue sky" preference. The experience of a blue sky is likely associated with, and significantly colored by things like the person's favorite color, the warmth of the sun or the cool breeze of that perfect day one summer when they were in love. For another person, it might be associated with the blinding heat of the desert and a sensation of thirst... and these two people can then end up arguing endlessly about whether a blue sky is obviously good or bad.

And both are utterly deluded to think this their preferences have anything to do with reality.

I am not saying that they disagree about how many angstroms the light of a blue sky is, or that the sky doesn't really exist or anything like that. I'm saying their preference is (and can only be) about their maps, because the mere fact of a blue sky has no inherent "preferability", without reference to some other purpose.

Even if we try to say in the abstract that it is good because it's correlated with things about our planet that make life possible, we can only have that preference because we're modeling an experience of "life" that we label "good". (And perhaps a depressed teenager would disagree, using a differently-labeled experience of "life"!)

This does not mean preferences are invalid or meaningless. It doesn't mean that we should only change our preferences and ignore reality. However, to the extent that our preferences produce negative experiences, it is saner to remove the negative portion of the preference.

Luckily, human beings are not limited to, or required to have, bidirectional preferences. Feeling pain at the absence of something is not required in order to experience pleasure at its presence, in other words. (Or vice versa.)

Awareness of this fact, combined with an awareness that it is really the experience we prefer (and mainly, the somatic markers we have attached to the experience) makes it plain that the logical thing to do is to remove the negative label, and leave any positive labels in place.

However, if we think that our internal labeling of experience has something to do with "reality", then we are likely to engage in the confusion of thinking that removing a negative label of badness will somehow create or prolong badness in the territory.

And for that matter, we may be under the mistaken impression that changing the reality out there will change our experience of it... and that is often not the case. As the saying goes, if you want to make a human suffer, you can either not give them what they want, or else you can give it to them! Humans tend to create subgoals that get promoted to "what I want" status, without reference to what the original desired experience was.

For example, looking for blue skies...

When what you really want is to fall in love again.

Comment author: loqi 07 June 2009 09:35:16PM * 0 points [-]

Wow, -5! People here don't seem to appreciate this sort of challenge to their conceptual framework.

I'm just saying that human beings have no way to model their preferences except by modeling experiences, in human-sensory terms.

I agree, but I wonder if I failed to communicate the distinction I was attempting to make. The human-sensory experience of being embedded in a concrete, indifferent reality is (drugs, fantasies, and dreams aside) basically constant. It's a fundamental thread underlying our entire history of experience.

It's this indifference to our mental state that makes it special. A preference expressed in terms of "reality" has subjective properties that it would otherwise lack. Maybe I want the sky to be blue so that other people will possess a similar experience of it that we can share. "Blueness" may still be a red herring, but my preference now demands some kind of invariant between minds that seemingly cannot be mediated except through a shared external reality. You might argue that I really just prefer shared experiences, but this ignores the implied consistency between such experiences and all other experiences involving the external reality, something I claim to value above and beyond any particular experience.

Even if you phrase this as, "I prefer the sky to be actually blue, even if I don't know it", it is still a lie, because now you are modeling an experience of the sky being blue, plus an experience of you not knowing it.

This is where the massive implicit context enters the scene. "Even if I don't know it" is modeled after experience only in the degenerate sense that it's modeled after experience of indifferent causality. A translation might look like "I prefer to experience a reality with the sorts of consequences I would predict from the sky being blue, even if I don't consciously perceive blue skies". That's still an oversimplification, but it's definitely more complex than just invoking a generic memory of "not having known something" and applying it to blue skies.

The two notions are basically isomorphic, so where's the value in the distinction?

Well, it makes clear some of the limits of certain endeavors that are often discussed here. It dissolves confusions about the best ways to make people happy, and whether a world should be considered "real" or "virtual", and whether it's somehow "bad" to be virtual.

I don't see how any of that is true. I can easily think of different concrete realizations of "real" and "virtual" that would interact differently with my experience of reality, thus provoking different labellings of "good" and "bad". If your point is merely that "real" is technically underspecified, then I agree. But I don't see how you can draw inferences from this underspecification.

For another person, it might be associated with the blinding heat of the desert and a sensation of thirst... and these two people can then end up arguing endlessly about whether a blue sky is obviously good or bad.

And both are utterly deluded to think this their preferences have anything to do with reality.

I'm going to have to turn your own argument against you here. To the extent that you have a concept of reality that is remotely consistent with your everyday experience, I claim that "in reality, blue skies are bad because they provoke suffering" is a preference stated in terms of an extremely similar reality-concept, plus a suffering-concept blended together from first-hand experience and compassion (itself also formed in terms of reality-as-connected-to-other-minds). For you to say it has "nothing to do with reality" is pure semantic hogwash. What definition of "reality" can you possibly be using to make this statement, except the one formed by your lifetime's-worth of experience with indifferent causality? You seem to be denying the use of the term to relate your concept of reality to mine, despite their apparent similarity.

However, to the extent that our preferences produce negative experiences, it is saner to remove the negative portion of the preference.

This doesn't make sense to me. Whether or not an experience is "negative" is a function of our preferences. If a preference "produces" negative experiences, then either they're still better than the alternative (in which case it's a reasonable preference, and it's probably worthwhile to change your perception of the experience) or they're not (in which case it's not a true preference, just delusion).

Luckily, human beings are not limited to, or required to have, bidirectional preferences. Feeling pain at the absence of something is not required in order to experience pleasure at its presence, in other words. (Or vice versa.)

That's a property of pain and pleasure, not preference. I may well decide not to feel pain due to preference X being thwarted, but I still prefer X, and I still prefer pleasure to the absence of pleasure.

Awareness of this fact, combined with an awareness that it is really the experience we prefer (and mainly, the somatic markers we have attached to the experience) makes it plain that the logical thing to do is to remove the negative label, and leave any positive labels in place.

This is where I think your oversimplification of "experience vs reality" produces invalid conclusions. Those labels don't just apply to one experience or another, they apply to a massively complicated network of experience that I can't even begin to hold in my mind at once. Given that, your logic doesn't follow at all, because I really don't know what I'm relabeling.

This relates to a general reservation I have with cavalier attitudes toward mind-hacks: I know full well that my preferences are complex, difficult to understand, and grossly underspecified in any conscious realization, so it's not at all obvious to me that optimizing a simple preference concerning one particular scenario doesn't carry loads of unintended consequences for the rest of them. I've had direct experience with my subconsciously directed behavior "making decisions for me" that I had conscious reasons to optimize against, only later to find out that my conscious understanding of the situation was flawed and incomplete. I think that ignoring the intuitive implications of an external reality leads to similar contradictions.

Comment author: pjeby 08 June 2009 12:47:54AM 0 points [-]

You seem to mostly be arguing against a strawman; as I said, I'm not saying reality doesn't exist or that it's not relevant to our experiences. What I'm saying is that the preferences are composed of map, and while there are connections between that map and external reality, we are essentially deluded to think our preferences refer to actual reality, and that this delusion leads us to believing that changing external reality will change our internal experience, when more often the reverse is more likely true. (That is, changing our internal experience will more likely result in our taking actions that will actually change external reality.)

Note, however that:

Whether or not an experience is "negative" is a function of our preferences.

Here you seem to be arguing my point. The experience is a function of preferences, the preferences are a product of, and point to, other experiences, in a self-sustaining loop that sometimes might as well not be connected to outside reality at all, for all that it has anything to do with what's actually going on.

Lucky people live in a perpetual perception of good things happening, unlucky people the opposite, even when the same events are happening to both.

How can we say, then, that either person's perceptions are "about" reality, if they are essentially unconditional? Clearly, something else is going on.

If we disagree at this point, I'd have to say it can only be because we disagree on what "about" means. When I say preferences are not "about" reality, it is in the same sense that Robin Hanson is always saying that politics is not "about" policy, etc.

Clearly, preferences are "about" reality in the same sense that politics are "about" policy. That is, reality is the subject of a preference, in the same way that a policy might be the subject of a political dispute.

However, in both cases, the point of the ostensible activity is not where it appears to be. In order for politics to function, people must sincerely believe that it is "about" policy, in precisely the same way as we must sincerely believe our preferences are "about" reality, in order to make them function -- and for similar reasons.

But in neither case does either the sincerity or the necessity of the delusion change the fact that it's nonetheless a delusion.

Comment author: loqi 08 June 2009 02:35:40AM 1 point [-]

If we disagree at this point, I'd have to say it can only be because we disagree on what "about" means.

I don't think I disagree with any of the above, except to dispute its full universality (which I'm not sure you're even arguing). To attempt to rephrase your point: Our interactions with reality create experiences filtered through our particular way of characterizing such interactions. It's these necessarily subjective characterizations (among other things), rather than the substance of the interaction itself, which generate our preferences. When reflecting on our preferences, we're likely to look right past the interpretive layer we've introduced and attribute them to the external stimulus that produced the response, rather than the response itself.

Robin's "X is not about Y" has the flavor of general, but not universal, rules. Would you extend your analogy to include this property?

Comment author: pjeby 08 June 2009 03:08:48AM -1 points [-]

Robin's "X is not about Y" has the flavor of general, but not universal, rules. Would you extend your analogy to include this property?

Here's an interesting question for you: why is it important that you consider this non-universal? What value does it provide you for me to concede an exception, or what difference will it make in your thinking if I say "yes" or "no"? I am most curious.

(Meanwhile, I agree with your summation as an accurate, if incomplete restatement of the bulk of my point.)

Comment author: loqi 08 June 2009 05:22:54AM 0 points [-]

Because I'm trying to make sense of your position, but I don't think I can with such a strict conclusion. I don't see any fundamental reason why someone couldn't form preferences more or less directly mediated by reality, it just seems that in practice, we don't.

If you're asking why I'm bringing up universality, it seemed clear that your claims about preferences were universal in scope until you brought up "X is not about Y". "Must logically be" and "tends to be in practice" are pretty different types of statement.

Comment author: pjeby 08 June 2009 06:05:11PM -1 points [-]

You didn't answer my questions.

I mean, you said some things that sound like answers, but they're not answers to the questions I asked. Here they are again:

Why is it important that you consider this non-universal?

and

What value does it provide you for me to concede an exception, or what difference will it make in your thinking if I say "yes" or "no"?

Your almost-answer was that you don't think you can "make sense" of my position with a strict conclusion. Why is that? What would it mean for there to be a strict conclusion? How, specifically, would that be a problem?

Comment author: loqi 09 June 2009 09:41:01AM 1 point [-]

Why is it important that you consider this non-universal?

I didn't answer this because it's predicated on an assumption that has no origin in the conversation. I never claimed that it was "important" for me to consider this non-universal. As per being "liberal in what I accept" in the realm of communication, I tried to answer the nearest meaningful question I thought you might actually be asking. I thought the phrase "If you're asking why I'm bringing up universality" made my confusion sufficiently clear.

If you really do mean to ask me why I think it's important that I believe in some property of preference formation, then either I've said something fairly obvious to that end that I'm not remembering (or finding), or you're asserting your own inferences as the basis of a question, instead of its substance. I try to give people the benefit of the doubt that I've misunderstood them in such cases, rather than just assume they're speaking manipulatively.

What value does it provide you for me to concede an exception

No particular value in mind. I suppose the greatest value would be in you solidly refuting such exceptions in a way that made sense to me, as that would be a more surprising (therefore more informative) outcome. If you the concede the exception, I don't gain any additional insight, so that's of fairly neutral value.

what difference will it make in your thinking if I say "yes" or "no"?

Not really sure yet, especially in the "no" case (since in that case you may have reasons I haven't yet thought of or understood). I suppose in the "yes" case I'd have greater confidence that I knew what you were talking about if I encountered similar concepts in your comments elsewhere. This discussion has had some difference on my thinking: I don't think I understood the thrust of your point when I originally complained that your distinction lacked relevance.

Your almost-answer was that you don't think you can "make sense" of my position with a strict conclusion. Why is that? What would it mean for there to be a strict conclusion?

By strict conclusion, I mean "preferences are modeled strictly in terms of the map: it is logically impossible to a hold preference expressed in terms of something other than that which is expressed in the map". This seems very nearly true, but vulnerable to counterexamples when taken as a general principle or logical result of some other general principle. I'll elaborate if you'd like, but I thought I'd clarify that you meant it that way. If you didn't, theoretical or speculative counter-examples aren't particularly relevant.

continue this thread »

Comment author: Vladimir_Nesov 07 June 2009 10:12:52AM * 1 point [-]

Think of what difference is there between "referring directly" to the outside reality and "referring directly" to the brain. Not much, methinks. There is no homunculus whose hands are only so long to reach the brain, but not long enough to touch your nose.

Comment author: loqi 07 June 2009 08:21:37PM 0 points [-]

Agreed, as the brain is a physical object. Referring "directly" to subjective experiences is a different story though.

Comment author: timtyler 05 June 2009 11:00:35PM 0 points [-]

Whether your preferences refer to your state, or to the rest of the world is indeed a wirehead-related issue. The problem with the idea that they refer to your state is that that idea tends to cause wirehead behaviour - surgery on your own brain to produce the desired state. So - it seems desirable to construct agents that believe that there is a real world, and that their preferences relate to it.

Comment author: pjeby 05 June 2009 11:07:23PM -1 points [-]

Whether your preferences refer to your state, or to the rest of the world is indeed a wirehead-related issue. The problem with the idea that they refer to your state is that that idea tends to cause wirehead behaviour - surgery on your own brain to produce the desired state. So - it seems desirable to construct agents that believe that there is a real world, and that their preferences relate to it.

I agree - that's probably why humans appear to be constructed that way. The problem comes in when you expect the system to also be able to accurately reflect its preferences, as opposed to just executing them.

This does not preclude the possibility of creating systems that can; it's just that they're purely hypothetical.

To the greatest extent practical, I try to write here only about what I know about the practical effects of the hardware we actually run on today, if for no other reason than if I got into entirely-theoretical discussions I'd post WAY more than I already do. ;-)

Comment author: timtyler 05 June 2009 11:14:09PM 0 points [-]

Presumably, if you asked such an agent to reflect on its own purposes, it would claim that they related to the external world (unless it's aim was to deceive you about its purposes for signalling reasons, of course).

For example, it might claim that its aim was to save the whales - rather than to feel good about saving the whales. It could do the latter by taking drugs or via hypnotherapy - and that is not how it actually acts.

Comment author: pjeby 05 June 2009 11:28:31PM 0 points [-]

Presumably, if you asked such an agent to reflect on its own purposes, it would claim that they related to the external world (unless it's aim was to deceive you about its purposes for signalling reasons, of course).

Actually, if signaling was its true purpose, it would claim the same thing. And if it were hacked together by evolution to be convincing, it might even do so by genuinely believing that its reflections were accurate. ;-)

For example, it might claim that its aim was to save the whales - rather than to feel good about saving the whales. It could do the latter by taking drugs or via hypnotherapy - and that is not how it actually acts.

Indeed. But in the case of humans, note first that many people do in fact take drugs to feel good, and second, that we tend to dislike being deceived. When we try to imagine getting hypnotized into believing the whales are safe, we react as we would to being deceived, not as we would if we truly believed the whales were safe. It is this error in the map that gives us a degree of feed-forward consistency, in that it prevents us from certain classes of wireheading.

However, it's also a source of other errors, because in the case of self-fulfilling beliefs, it leads to erroneous conclusions about our need for the belief. For example, if you think your fear of being fired is the only thing getting you to work at all, then you will be reluctant to give up that fear, even if it's really the existence of the fear that is suppressing, say, the creativity or ambition that would replace the fear.

In each case, the error is the same: System 2 projection of the future implicitly relies on the current contents of System 1's map, and does not take into account how that map would be different in the projected future.

(This is why, by the way, The Work's fourth question is "who would you be without that thought?" The question is a trick to force System 1 to do a projection using the presupposition that the belief is already gone.)

pjeby comments on Post Your Utility Function - Less Wrong

Article Navigation

Comments (273)

Nearest Meetups

Virtual Study Room

Recent Comments

Recent Posts

Recent Rationality Quotes

Recent Wiki Edits

Top Contributors, 30 Days

Recent Karma Awards

Comments (273)

Top Contributors, 30 Days

You'll need to login or register to do that

(Don't worry, it only takes a few seconds)

Create

Login