In an alternate universe, on an alternate earth, all smokers, and only smokers, get brain cancer. Everyone enjoys smoking, but many resist the temptation to smoke, in order to avoid getting cancer. For a long time, however, there was no known cause of the link between smoking and cancer.
Twenty years ago, autopsies revealed tiny black boxes implanted in the brains of dead persons, connected to their brains by means of intricate wiring. The source and function of the boxes and of the wiring, however, remains unknown. There is a dial on the outside of the boxes, pointing to one of two positions.
Scientists now know that these black boxes are universal: every human being has one. And in those humans who smoke and get cancer, in every case, the dial turns out to be pointing to the first position. Likewise, in those humans who do not smoke or get cancer, in every case, the dial turns out to be pointing to the second position.
It turns out that when the dial points to the first position, the black box releases dangerous chemicals into the brain which cause brain cancer.
Scientists first formed the reasonable hypothesis that smoking causes the dial to be set to the first position. Ten years ago, however, this hypothesis was definitively disproved. It is now known with certainty that the box is present, and the dial pointing to its position, well before a person ever makes a decision about smoking. Attempts to read the state of the dial during a person’s lifetime, however, result most unfortunately in an explosion of the equipment involved, and the gruesome death of the person.
Some believe that the black box must be reading information from the brain, and predicting a person’s choice. “This is Newcomb’s Problem,” they say. These persons choose not to smoke, and they do not get cancer. Their dials turn out to be set to the second position.
Others believe that such a prediction ability is unlikely. The black box is writing information into the brain, they believe, and causing a person’s choice. “This is literally the Smoking Lesion,” they say. Accepting Andy Egan’s conclusion that one should smoke in such cases, these persons choose to smoke, and they die of cancer. Their dials turn out to be set to the first position.
Still others, more perceptive, note that the argument about prediction or causality is utterly irrelevant for all practical purposes. “The ritual of cognition is irrelevant,” they say. “What matters is winning.” Like the first group, these choose not to smoke, and they do not get cancer. Their dials, naturally, turn out to be set to the second position.
Entirely Useless 
How did scientists discover that the black box is pre-set if they can’t look at it while a person is alive? Can you find people with the dial in the first position who died before they got the chance to smoke?
LikeLike
In general, we could flesh out the details of the story in many ways without changing the implication, namely that Newcomb’s problem and the Smoking Lesion are simply the same problem for all practical purposes.
Perhaps the scientists discovered that the dial is pre-set by means of some nanobot motion detector which would detect any movement of the dial, but not its current location. Trying to detect the current location would presumably cause the explosion.
Of course, stepping outside the story, we all know why you are not allowed to look at it while the person is alive: because you can choose this algorithm: “Check if the box is set to the first position. If so, do not smoke. If it is set to the second position, smoke,” and then the correlation will be necessarily broken. But there is no difference in this regard between the Lesion and Newcomb: take the case of Omega who flies away after the boxes are filled or not. You can break the correlation by implementing this algorithm: “Check the box for the million. If the million is there, take both boxes. If it is not, take only the empty one.” The natural response would be that in such a situation the box will be empty, and the person’s lust for money will lead him to take the thousand. The corresponding answer for the lesion, if we allow checking, would be that if you determine that you are set to cancer, you realize you are going to get cancer anyway, so you might as well smoke. And of course only people who have the dial set to the first position, ever check.
In other words, there are various ways to deal with the possibility of checking, but this makes no relevant difference between Newcomb and the Lesion.
“Can you find people with the dial in the first position who died before they got a chance to smoke?” This would depend on how we flesh out the details of the story; apparently people choose only once whether to be smokers or non-smokers, which isn’t terribly similar to our world. But my inclination is to say, “Sure, some people died the day before they made their choice, and some of them had it set to the first position, and some to the second.”
LikeLike
I don’t think this argument works. If you initially intend to think this way, but it turns out that the box is writing to your brain to cause your choice, the box overwrites your intention.
Of course, where the box is infallible as stated, it doesn’t matter what your initial intention was in the writing version. So you might as well assume that it is reading your intention since in that case it does matter what your intention is.
But if the box merely creates a tendency rather than a certain choice, then if that is what it is in fact doing (rather than reading your brain) then it is rational to smoke as in the smoking lesion problem.
LikeLike
First, the scientists came up with the ideas of reading and writing, but not that these actually are not exhaustive. There could also be intermediate possibilities including both reading and writing. The box could be some sort of brain computer interface.
Second, there is no need for any overwriting. All adults have the boxes, so whatever your initial thoughts about the matter, that is what was written. So there is no need for overwriting. Also, even if something did get overwritten, that would just feel like changing your mind, which is a thing that happens often to human beings. So it would feel quite normal.
Third, we need to start with what we agree on, and then use that to consider the things we disagree about. You say that you “might as well assume that it is reading,” so we agree that in the story as written, it is rational to choose not to smoke, even though we disagree on the reason.
You start to discuss the disagreement when you say, “But if the box merely creates a tendency…” Let us think about this. We will modify the story:
“The correlation was very strong, but not perfect. Out of each thousand persons who chose to smoke, on average 999 had their dial set to the first position, and died of cancer. On average one person had his dial set to the second position, and he did not get cancer. Likewise, out of each thousand persons who chose not to smoke, on average 999 had their dial set to the second position, and they did not get cancer. And on average one had his dial set to the first position, and that unfortunate fellow got cancer even though he did not smoke.”
“The scientists still did not know for sure whether the box was reading or writing.”
I assume you agree with me that it is still rational not to smoke in this situation, especially since it could just be that you have a predictor which is just nearly perfect.
On with the story:
“Two years later, the scientists conclusively ruled out the possibility of prediction. They now know that the box is writing, but only in a way that creates a very strong tendency, so that 99.9% of people follow the tendency.”
“The great philosopher Simon said, ‘Since the box merely creates a tendency rather than a certain choice, and this is what it is in fact doing, rather than reading, then it is rational to smoke as in the smoking lesion problem.’ Thousands of people were persuaded by his eloquence. Unsurprisingly, 999 out of 1000 of those convinced had their dials set to the first position, and choosing to smoke, they died of cancer. Likewise, 999 out of 1000 of the evidential decision theorists, who refused to be convinced by Simon’s arguments, chose not to smoke, and had their dials set to the second position. They survived, unlike Simon’s followers.”
Are you still in favor of following the philosopher’s opinion and joining the dead group, rather than following the evidential decision theorists and surviving?
LikeLike
What we agree on:
I agree that it is correct to 1-box in Newcomb’s problem (i.e. not to smoke in this problem if it is reading and not writing).
What we disagree on:
You say whether it is reading or writing doesn’t matter, and you seem to be arguing from that that it is correct to not smoke in the smoking lesion problem.
I disagree, thinking it is correct to smoke in the smoking lesion problem.
Your first modified story:
““The scientists still did not know for sure whether the box was reading or writing.”
I assume you agree with me that it is still rational not to smoke in this situation, especially since it could just be that you have a predictor which is just nearly perfect.”
Yes, as I would agree that it is correct not to smoke in this case.
Your second modified story:
“Two years later, the scientists conclusively ruled out the possibility of prediction. They now know that the box is writing, but only in a way that creates a very strong tendency, so that 99.9% of people follow the tendency.”
“The great philosopher Simon said, ‘Since the box merely creates a tendency rather than a certain choice, and this is what it is in fact doing, rather than reading, then it is rational to smoke as in the smoking lesion problem.’ Thousands of people were persuaded by his eloquence. Unsurprisingly, 999 out of 1000 of those convinced had their dials set to the first position, and choosing to smoke, they died of cancer. Likewise, 999 out of 1000 of the evidential decision theorists, who refused to be convinced by Simon’s arguments, chose not to smoke, and had their dials set to the second position. They survived, unlike Simon’s followers.”
You say:
“Are you still in favor of following the philosopher’s opinion and joining the dead group, rather than following the evidential decision theorists and surviving?”
That’s a very presumptuous way of stating it. The scenario we are considering here is this:
1) The dials are set before considering the arguments and making the decision, and the setting does not change as a result of the decision, and is not originally set by reading anything about your brain.
2A) If the dial is set to the first position, you get brain cancer with certainty.
2B) If the dial is set to the second position, you certainly don’t get brain cancer.
3A) If the dial is set to the first position, the box writes to your brain to give you a very strong tendency to smoke.
3B) If the dial is set to the second position, the box writes to your brain to give you a very strong tendency not to smoke.
In this scenario, absolutely nothing about your decision, or what decision theory you use, etc. affects whether you get brain cancer or not – it is predestined by your box, which has nothing to do with your thinking.
One way the box might give you a strong tendency not to smoke is by writing to your brain giving you a strong tendency to be an evidential decision theorist. So, the fact that you are an evidential decision theorist is strong evidence that your dial is set to the second position. But, the evidential decision theory did nothing to help you in this respect. It is merely an unfortunate side effect of the fortunate position of your dial. The evidential decision theory is merely taking credit for the dial setting that it did nothing to affect. The only actual effect of evidential decision theory here is giving you a small chance to fail to smoke (while still getting brain cancer) if you managed to talk yourself into evidential decision theory despite having the dial set to the first position.
On the other hand if you decide to smoke, this is evidence your dial is set to the first position, but the decision theory you use is doing nothing to cause the dial setting. It is just giving you some chance to smoke, without getting brain cancer, if you manage to accept a sensible decision theory despite having the dial set to the second position.
Note this all contrasts with Newcomb’s problem where it is rational to, before the predictor examines your brain, reprogram yourself to precommit to 1-box. In this case, the 1-boxing decision theory is doing real work for you (when the predictor examines you).
LikeLike
If I understand your response correctly, it says this:
“In that situation, I would say: if I smoke, there is a small chance I won’t get cancer. On the other hand, if I don’t smoke, I will probably turn out to be one of the lucky ones that don’t get cancer. But my decision theory will have had nothing to do with this. So I’m going to smoke, even though I will have a 999/1000 chance of getting cancer, since my tendency to go forward with reasoning like this is very likely because my dial is set to the first position.”
Do you agree that you are saying this?
LikeLike
Yes, I agree
LikeLiked by 1 person
I would say that’s a pretty bad ritual of cognition, and would prefer to be “lucky”, rather than dead. I’m not sure if there’s more to say on this question, except perhaps this:
We never have 100% probabilities in real life. So even if the correlation were perfect, as in the original story, we would not know for sure that it is perfect. So if it were discovered we were in the writing case, with the perfect correlation, you would still choose to smoke, even though it would essentially certain that doing that would kill you?
LikeLike
If the correlation is perfect, and the box is writing, not reading, then there is no point pretending that you have any control over whether you smoke or not.
About the “we never have 100% probabilities”: this also applies to whether the box is just writing or also doing some reading. So, the decision would have to depend on the relative probabilities of escaping the correlation v. the box reading, and the relative severity of the cancer v. the benefit of smoking.
LikeLike
“there is no point pretending that you have any control over whether you smoke or not”
This is the same as saying, “If determinism is true, there is no point in pretending that you can make choices.”
You have no choice about whether to make choices, even if determinism is true; and in the same way you would have to decide to smoke or not smoke, even in the writing case.
LikeLike
Following on from this example of yours, Lets decrease the correlation further to something which we can find in the real world. Instead of 999 out of a thousand, Out of each thousand persons who chose to smoke, on average 800 had their dial set to the first position, and died of cancer. On average 200 had their dials set to the second position, did not get cancer. Likewise, out of each thousand persons who chose not to smoke, on average 800 had their dial set to the second position, and they did not get cancer. And on average 200 had their dial set to the first position, and those unfortunate fellows got cancer even though they did not smoke.
800 of the 1000 evidential decision theorists, refusing Simons arguments, chose not to smoke, and had their dials set to the second position. 800 of the 1000 still survived.
In other words, the above example proves that In the real world, evidential decision theorists can in fact prevent cancer by basing their decision purely on correlation data. Agreed?
LikeLike
In the example you should expect that you are less likely to get cancer if you choose not to smoke based on the correlation. In this sense it shows that you can “prevent” cancer. However, it does not prove that it prevents it causally, since it may be that the condition of the dial causes you to use EDT.
LikeLike
Thanks. Yup, I meant “prevent” cancer in a non-causal way. Just a question, how did you arrive at the conclusion that 999 out of 1000 EDT’s who chose not to smoke BECAUSE of their knowledge of the study will also have their dials set in the second/anti-cancer position, just like the non-smokers IN the study? Why are the non-smokers in the study identical to the non-smokers who choose not to smoke (or think they are “choosing” not to smoke) BECAUSE of the study. i.e. both have 999 out of a 1000 with dials set to the second position? Does this also apply when the proportion is 900 out of a 1000, 800 out of a 1000, and so forth?
LikeLike
In reality it could turn out various ways when you control for decision theory. I simply posited (in the story) that the statistics did not turn out differently, to make not smoking the correct decision. Depending on how it actually turns out, smoking could be the correct decision. The point is that what is important is correlation and not causation.
LikeLike
Fair enough. Lets see whether I’ve got this. In our world today, the current big issues are the correlations between carbon emissions and global warming, and smoking and cancer. Does this mean that, despite the absence of knowledge of causation, based on the alien implant logic, the correct or “rational” decision for people in the real world is to cut carbon emissions and stop smoking based purely on this correlational data?
LikeLike
Yes, assuming you care about avoiding global warming or cancer.
LikeLike
There’s plenty of point pretending for regular determinism, as the mental attitude of taking your decisions seriously will in fact lead to better of outcomes. Here, it won’t, only the box setting does.
(I considered clarifying the “no point pretending” statement after posting it, as it includes an implied assertion that you don’t have control which is very unclear in this case, then decided it was too hard to figure out what “you” and “control” actually means and left it there as the statement itself is literally true, even if the implied assertion is unclear.)
LikeLike
In the infallible case, choosing not to smoke will lead to a better outcome than choosing to smoke. The box setting will correspond.
You say it is pointless to choose to smoke or not in that situation. So what would you do instead of choosing? Lie down and die?
LikeLike
Well no.
see:
http://lesswrong.com/lw/r0/thou_art_physics/
The difference is, in this case the causal diagram really does look like the first one in that linked post, not the later ones, with the box taking the spot of “physics” in that diagram. There is also, of course, an arrow from the box to you, setting your smoking decision, but that is irrelevant – the point is that whether you get cancer is wholly decided by the box setting and no causation of that flows from your decision to smoke, your prior tendencies which would affect your decision, your decision theory on how to decide to smoke, any arguments from anyone considering decision theory, etc.
There is correlation, but the causation flows entirely directly from the box and your mental state is not even an intermediate step in that causation.
Whereas, if you decide to lie down and die, your death follows directly from that decision, which is a completely different situation.
LikeLike
You didn’t tell me yet what you would choose to do in the writing situation. Saying, “but I don’t have a choice in that situation,” does not say what you would do. And since it would be written that you would choose to do something, you would choose to do something. What would you choose to do?
“in this case the causal diagram really does look like the first one in that linked post”… Wrong. It looks like the second, with “physics” being “physics and the box.”
“There is also, of course, an arrow from the box to you, setting your smoking decision, but that is irrelevant – the point is that whether you get cancer is wholly decided by the box setting and no causation of that flows from your decision to smoke”… Even in Eliezer’s post, we could make another flow chart : Physics ten thousand years ago > Physics and me today… and all the causality would flow from ten thousand years ago to today, and none from today to ten thousand years ago.
My argument is that causality doesn’t matter; getting the good result does, and you do that by not smoking.
LikeLike
In the writing situation, I would prefer to choose to smoke, but I would end up doing so if and only if the dial were set to the first setting. For all other choices, I would choose normally.
” It looks like the second, with “physics” being “physics and the box.””
No, it does look like the first diagram, and does not look like the second. Recall that:
a) the box creates the cancer by directly releasing chemicals depending on the dial setting
b) the dial setting is not selected by reading the mind
So, there is no causal arrow pointing from “me” to “cancer” as shown in the second diagram. There is only a causal arrow from the box to “cancer” (and another causal arrow from the box to “me”, not shown on that diagram, but unnecessary to the point).
“Even in Eliezer’s post, we could make another flow chart : Physics ten thousand years ago > Physics and me today… and all the causality would flow from ten thousand years ago to today, and none from today to ten thousand years ago.”
Yes, and that’s totally different from this situation! Physics from 10 thousand years ago controls the future via the present (including one’s own state), but the box does not control the cancer via the smoking decision. It controls the cancer purely through its own release of chemicals – the effect on the smoking decision is just a side effect.
“My argument is that causality doesn’t matter; getting the good result does, and you do that by not smoking.”
I’m all for getting the best results, but, you aren’t getting good results from your decision theory, but from the dial setting which isn’t affected by your decision theory.
LikeLike
“In the writing situation, I would prefer to choose to smoke.” That seems pretty foolish.
Again, my point is that being caused to be in a good situation is just as good as causing it. There is no need to stamp your foot and say if you don’t get to cause it, you don’t want it.
LikeLike
You’ve left out a fourth possible kind of response: those who note that the argument is utterly irrelevant, and yet still smoke, get cancer, and die, with dials found to be in the first position.
Underlying the assumption that this fourth group doesn’t exist is the assumption of free will. Ignoring this fourth group suggests begging the question. And when we try and connect your analogy to our real world, we see that this fourth group exists, at least to the extent that we can trust people’s communications about their own beliefs.
The reality is that no one who disbelieves in free will believes that it matters in the least. (I avoid the word “determinism” because it invites arguments about randomness, which I think are irrelevant to the free will question.) If we haven’t free will, then we haven’t the free will to disbelieve free will. Where it matters is where we treat questions differentially– where we say that the subject has free will in one domain, but not in another. The argument from the free will skeptic is that no such distinction is appropriate. So let’s change your argument a little bit and see what comes out.
There’s a switch that will make eventually people believe themselves to be Napoleon. For a long time, this switch is assumed to be set at birth– but it really doesn’t matter for the free will argument when it is set, or what switches it. (Note that all three of your groups exist in regards to this switch, and that my fourth group also exists.) One day the CIA discover a magic spell and find that every target at which they cast it eventually believes him or herself to be Napoleon, and post-mortem examination shows the Napoleon switch set appropriately. Statistical analysis shows this to happen much more often than mere chance would suggest, so we assume causality. This is used by the CIA to Napoleonize and thus nullify political dissidents. But many CIA operatives find this ethically distasteful and refuse to participate in the program. A second switch is discovered: one which correlates perfectly with which CIA operatives participate. As more and more of these operatives die and are examined, this correlation is established beyond any reasonable doubt. Indeed, the CIA eventually discover a second spell which appears to cause CIA operatives to tolerate Napoleonization. However, not all CIA operatives are willing to use this second spell.
Now: does it make any sense to have one opinion about the first switch, yet a different opinion about the second switch? Does it make any sense to hold that behaviors of people suffering from schizophrenia are any more or less free than behaviors of people not?
LikeLike
It’s kind of amusing that I had to remove your comment from the spam filter in order to publish it. Apparently even an algorithm was able to detect that it was not relevant to the post.
The post is about decision theory, not about free will.
LikeLike
I get where you are coming from OP, and I can certainly see why a 100% correlation may rule out the effect of causal decision theory. In your example, If anyone in the past could have used CDT to convince himself to smoke despite not having the lesion, he would have appeared as a person who smoked but did not get cancer, decreasing the odds of a smoker getting cancer from 100/100 to 99/100 (because he is now included in the group of smokers, but being a smoker without the lesion he does not die of cancer, ‘ruining’ the perfect correlation of the smoker group). Thus, if there was even ONE such person who used CDT, the correlation would not have been perfect, as it is. Hence, the perfect correlation is proof that in the infinite scheme of things which would include CDT users, no one could have resisted smoking if they had the lesion.
But suppose the correlation is not 100%, but say 80%. Lets modify it to this:
1. 100 people have the smoking lesion and all smoke, but only 80 of them get cancer
2. 100 people do not have the smoking lesion and don’t smoke, but 20 of them get cancer
Would your conclusion be any different? For example, does the 20% of people who get cancer despite not smoking reflect the numbers of smokers who use EDT to give up smoking and but still end up dying because they have the lesion?
And if so, is it correct to say, based on these stats, that there is a 20% chance that anyone who gives up cancer has the lesion, while 80% of those who give it up do not? Thus, that 20% who give up smoking will still die, but 80% won’t?
LikeLike
The real question is whether the correlation remains after controlling for a person’s decision theory. In the 100% case, the correlation remains no matter what you control for. In the 80% case, it is possible that the correlation will go away or change in some way when we control for the decision theory which is being used. So in your example, suppose we discover that the 20 people who smoke and do not get cancer are mainly using CDT. The 80 people who smoke and get cancer or mainly just choosing by their preference, without any specific theory. The 100 people who do not smoke are mainly using EDT, or at least think they are (I will explain the qualification below.)
In a situation somewhat along these lines, we would have no reason to believe that not smoking would improve our chances of not getting cancer, because the correlation between a person’s choice and cancer vanishes once we control for the decision theory being used. But when I say “we have no reason,” I mean precisely that we have no evidence. And consequently, someone actually using EDT in this situation will choose to smoke. This is why the people in the example think they are using EDT, but are not; or rather, they are applying it mistakenly.
In other words: EDT recommends not smoking when there is a correlation between smoking and cancer even after controlling for a person’s decision theory. It recommends smoking if the correlation vanishes after such controlling.
What about a more realistic situation, where we have not actually tested this, and we do not know whether the correlation remains or not? In this situation, your priors might lead you to believe that the correlation would very likely vanish if you could control for the theory. And to the extent that you believe this, you should choose to smoke even if you follow EDT. This is the real reason for the intuition that EDT recommended a mistaken course of action in the Smoking Lesion. But this was a misinterpretation of EDT from the very beginning.
LikeLike
Actually the specific breakdown I offered was a mistaken way of making my point, but I think you will get the idea anyway, so I’m not going to bother to correct it.
LikeLike
Thanks for your reply OP. I would like to focus on paragraph 3 and part of paragraph 4 of your reply first:
“In other words: EDT recommends not smoking when there is a correlation between smoking and cancer even after controlling for a person’s decision theory. It recommends smoking if the correlation vanishes after such controlling.”
“What about a more realistic situation, where we have not actually tested this, and we do not know whether the correlation remains or not? In this situation, your priors might lead you to believe that the correlation would very likely vanish if you could control for the theory…”
How do you estimate whether the correlation would likely vanish if you could control for the theory? Would it be as simple as, say, estimating the correlation between smoking and cancer from 1990 to 2000, publishing the results for the world, and then after publishing, doing a second correlation test between smoking and cancer from 2000 to 2010 to ascertain the effect on the correlation of any decision theories that would have inevitably resulted from the publishing of the first study?
So for example, if the second study showed a slight drop of the correlation from 80% to 65% for smoking and cancer, then we can estimate that the 15% drop probably came from new smokers using CDT (non smokers take up smoking because they believe in CDT, but don’t get cancer because they don’t have the lesion anyway).
and if there is a slight drop of the correlation from 80% to 70% for non-smoking and non-cancer, then we can estimate that the 10% drop probably came from smokers using EDT to give up smoking (these smokers give up smoking because they believe in EDT, but still get cancer because they still have the lesion).
Would you agree with this method of estimation?
LikeLike
Theoretically, a process like that might work. In the real world, probably not. Although Egan did not realize this, his intuition that one should smoke came from a concrete idea of how such a lesion would probably work. A lesion is physical damage, so Egan has a strong prior that it should cause fairly simple physical effects, not complicated mental ones. That is, the lesion might give you a desire to smoke, but it is unlikely to make you think in any particular way. This means that he expects the correlation to go away after controlling for a person’s decision theory.
In practice it would be hard to test this, though, because in real life the vast majority of people (probably upwards of 99%) would not use a decision theory at all. So you would be unlikely to get the kind of changes that you mentioned.
LikeLike
Ok, so what you are saying is if most people in the study do not use a decision theory, then it is highly likely that the correlation goes away after controlling for decision theory i.e. non-smokers still get cancer after using EDT to stop smoking after reading the study?
On the other hand, do you believe the effecient market hypothesis can have an impact here? that when looking back to past information involving human beings i.e. correlation between non-smokers and cancer, that it is reasonable to conclude that the majority of people were in fact using rational decisions based on all publicly available information to stop smoking including decision theories?
LikeLike