Newcomb's Paradox

If the predictor is able to predict that you will two-box, because making such a choice matches your rational profile, then so should the predictor be able to predict that someone would one-box in the case where, according to you, they choose to do so irrationally (and are rewarded for their irrationality). This brings into question your assumption that the subject is unable, through their Phase 2 deliberation, to rationally determine the content of the opaque box, and, by extension, also brings into question your judgement that one-boxing is irrational in such circumstances.

The strategy that you advocate for an agent who begins deliberating in Phase 1, while effective, reminds me of this delightful passage in M.R. Ayers’ book The Refutation of Determinism:

“But it is absurd to represent deliberation as a procedure of trying to influence oneself in a particular direction, since if someone knows in which direction he wants to influence himself, his deliberation is already done. Nor is deliberation a process of trying to bring oneself to perform some act successfully, no matter what. The end of deliberation is not to screw oneself up to the point of acting, but to determine rationally which course of action to follow.” p.152

It is worth noting that there is a tradition in decision theory (so-called causal decision theory, or CDT) that is right to reject what might be called naive evidentialism: the view that the mere fact that an act would be evidence of a good outcome gives one a reason to perform it. Consider a variant of the Newcomb scenario (sometimes called the “smoking lesion” case) where a gene both causes a craving for smoking and independently causes cancer. There, the bare fact that abstaining would be evidence of lacking the gene gives one no genuine reason to abstain, since one’s choice is causally downstream of the gene rather than upstream of it.

CDT correctly insists that what matters is causal structure, not mere evidential relevance. The trouble is that CDT then applies an impoverished conception of causation to Newcomb’s problem that recognizes only efficient causation pushing forward from the present moment and that has no place for rational causation as a distinct causal form. The gene in the medical case operates entirely independently of the agent’s deliberative process, bypassing agency altogether. The Newcombian predictor, by contrast, achieves their reliability precisely because rational causation is efficacious: what they track is the very reasoning process that will issue in the choice.

The disanalogy between Newcomb and the medical cases is therefore itself a causal point, which is why resolving it doesn’t require retreating to evidentialism but rather enriching our ontology of causation.

What Ayers’ comment quoted above also highlights is the peculiar structure of rational causation, where an agent’s reasons for acting are upstream from their decision, but the agent’s deliberative process and their psychological dispositions share a common rational source rather than standing in a simple temporal sequence. When viewed in that way, we can say that the predictor’s decision to fill up (or leave empty) the opaque box has the same rational causal source as the agent’s decision: namely their reasons for acting. So, when an agent decides to one-box during Phase 2, while it is true that at that time the predictor had already determined the content of the opaque box, it is not true that this determination was independent from the agent’s rational deliberative process. The predictor’s action and the agent’s rational decision have the same rational source.

I think you are misremembering what that whole thing was about. It wasn’t that important, but basically, I am saying that our hypothetical participants in V1 of the Newcomb game go through the entirety of Phase 1 not knowing they will be faced with the dilemma, or even knowing about the dilemma in theory.

Me and you know about the dilemma in theory. Therefore, we are not one of those participants. And yet, I am using pronouns like “you” and “we” in this post, describing us as the participants. But we couldn’t be, not in our current state. We’d need our memories regarding the whole dilemma to be deleted.

I am not sure what you are trying to argue with everything after “I concede this”. Box 2 is not like a basket of apples you look into. You don’t look into it.

But, if you agree that P exists, then you must agree with the mathematical expected utility formulae in my OP

When? When do they miss out on it? During Phase 1, when they live their lives as TBs, completely oblivious that their otherwise rational attitude will cost them 1 million in the future (this is V1, so they cannot know). The “choice” of missing out on the $1M is located in Phase 1, and it is not a choice, rather their nature as people who do not conflate correlation with causation.

When they enter Phase 2, by virtue of being TBs, they have already “missed out” on the $1M. They’ve already lost the main prize of the Newcomb game, through no fault of their own.

When they then, at this point, take two boxes, instead of one, they are choosing to go home with $1000 instead of $0.

Their choice to go home with $1000 instead of $1M however… never happened. It wasn’t a choice.

The expected return calculation is:

OB = XY + 0(1-X)
TB = 1,000X + (1,000 + Y)(1-X)

Where X is the success rate of the prediction and Y is the value of the second box.

If X = 0.99 then:

OB = 0.99Y
TB = 1,000 + 0.01Y

These are roughly equal if Y = 1,020.41.

So if the value of the second box is less than $1,020.41 then it’s rational to choose both boxes, else it’s rational to choose the second box.

And yet [almost] none of these supposed rational people leave with that figure.

If we discovered that, in life, one-boxers are generally more successful, it would be ridiculous to say that this shows that two-boxing is the wrong thing to do. Yet, it seems this is what one-boxers argue.

As I said many times, two-boxers always win more money than one-boxers in the situation described in the problem.

I don’t think so. We know the situation, we know the rules, and we know that people who one box win the most money. I don’t take it for granted that it’s accurate to say this “punishes rationality”. That’s just assuming your conclusion. I don’t assume your conclusion.

I don’t assume that you’re being rational by losing a game whose rules you know, when you see people who you look down on, winning with more money. It just sounds like sad whining to say they’re being rewarded for irrationality. They know the rules just like you, and they’re winning. Don’t whine about them winning, just… go win. You can one box too.

The Newcombian predictor, by contrast, achieves their reliability precisely because rational causation is efficacious: what they track is the very reasoning process that will issue in the choice.

This still doesn’t matter and CDT is right. Even if the predictor predicts the reasoning process that will happen in the room, you picking one box or two still doesn’t change the content of the box in front of you.

It’s assuming that conflating correlation with causation is irrational. You have yet to rigorously defend your stance against my earlier arguments that you were and are conflating the two.

If you try to sketch a causal picture, you will be incapable of seeing how one-boxing can be rational. But you’ll still see that OBs win more than TBs in general, in the Newcomb Game. You’ll also see that OBs might not actually win more in life, because conflating correlation with causation will usually harm you in real life.

Why did they win? Because in Phase 1, they were an OB. The entity saw this, and predicted they’d be an OB in Phase 2. The entity thus put $1M in the box. Then came Phase 2, and the OB was then guaranteed at least $1M.

This is why the OBs are winning over TBs. Do you agree?

You’ve undermined your own previous argument. An agent cannot be right less than 100% of the time when guessing what other agents will do who all act the same. If the predictor is wrong some of the time, but is attempting to be right all of the time, then the only possible situation is that not all agents it predicts act the same way.

As such, there has to be the possibility of being the 1 out of 100 who acts differently

It doesn’t change the content from what it was when you entered the room. But it does determine what this content is (and was). More precisely, the rational ground for your choice and the rational ground for the predictor’s action are one and the same since the latter tracks (and anticipates, but doesn’t cause) the first.

In an article that I wrote in 2009 (and had posted in the first of the three incarnations of The Philosophy Forum) I had highlighted a formal analogy between the CDT argument for two-boxing and a practical implication of van Inwagen’s Consequence argument for incompatibilism:

Suppose you are sitting at a restaurant table and are poised to order some dessert. Chocolate ice cream is listed on the menu and you want some. Maybe you will order some. If you do, you are likely to get some. But then, a Laplacean demon would already have known with certainty that you will get some. The demon knows this because the total state of the universe in the past was such that you will get some ice cream, and you can’t do anything about it since you have no control over the past states of the universe or the laws of nature. (This was one step in van Inwagen’s Consequence Argument.) But then you shouldn’t bother ordering any ice cream. If you don’t order any then you are unlikely to get any, for sure; but in that case that would simply mean that the state of the universe was such that you necessarily wouldn’t have ordered any. The demon would have known that already. So you won’t be able to blame yourself if you don’t order any ice cream. You simply can’t hold yourself responsible for what you will do. You might just as well do nothing, wait, and see what it is that the waiter will bring.

A more reasonable option is to order ice cream. Doing so places you in a special epistemic relation with a future event: knowing that it will happen through deciding to make it happen, and thereby ensuring that whatever the past state of the universe was, it was such that you would get ice cream. The one-boxer is in the same epistemic relationship with respect to the content of the opaque box and the action of the predictor. They can know that the opaque box will contain the big prize through deciding to one-box and thereby ensuring that the predictor will have filled it up.

Which is only possible if some agents are not perfectly rational agents, because all perfectly rational agents will make the same choice; if it is more rational to choose one box then all perfectly rational agents will choose one box and if it is more rational to choose two boxes then all perfectly rational agents will choose two boxes.

So really there are two different questions being asked:

  1. Which choice will all perfectly rational agents make?
  2. Which choice will you make?

It stands to reason that if (2) differs from (1) then you are making the less rational choice.

Which isn’t always what we want. Remember that there are four possible outcomes:

  1. I pick two and was predicted to pick one = I win $1,001,000
  2. I pick one and was predicted to pick one = I win $1,000,000
  3. I pick two and was predicted to pick two = I win $1,000
  4. I pick one and was predicted to pick two = I win $0

You are in the 1% if you are in scenario 4, but you don’t want to be in scenario 4. So it’s not just enough to trick the predictor; you specifically want to trick the predictor into scenario 1.

But how do you propose to do this? You can certainly rule out being in scenario 4 by picking two, but how do you ensure that you’re in scenario 1 and not scenario 3?

I say that there is no good way to trick yourself into scenario 1. It is rational to accept that the predictor is most likely to be correct, and that it is rational to prefer scenario 2 over scenario 3. Therefore, it is rational to choose one box.

Agree. I never denied this. I said that you haven’t demonstrated an impossibility. The predictor looks at us (just like the analogy of looking in the basket of apples) and, even before the problem is presented to us, pegs me as OB, and you and TB, and puts the money only in my box. There’s no impossibility of that.

Never looked at it actually, but it’s wrong.

2nd line should read (1-P), not P, for me, being OB. The first line should have a similar substitution for computing your probability instead of mine.

When they walk away with only 1000 in their pocket of course.

I actually don’t see free will coming into play in this scenario, but that’s just me (and at least @Michael). Some say that this free will is necessary for decision making, but then what do all decision making devices leverage deterministic processes to make the best decisions? If we cannot choose, then we’d not have evolved this expensive brain to make better ones.

:+1:

I can think of scenarios that do. My core brain is not rational, but it believes all sorts of intuitive things like time flowing and whatnot. The rational part of me knows that’s bunk, but it’s not in charge. A true rational creature would discard all the lies that evolution put there to make one fit, and the resulting rational creature would not be fit at all and would quickly be naturally selected out (punished for rationality so to speak). Anyway, yet, I agree with you here that this scenario is hardly one of them.

1 Like

Before responding to other things, are you assuming a perfect predictor? If you are assuming a perfect predictor, then there is no choice really (it’s a nonsensical question in my opinion) but if you really want to talk about it as if there was one, then yes you should one box.

I was talking about the non-perfect predictor case.

Here’s an idea for a predictor:

Before making your choice you are cloned (including your brain) and your clone is asked to make the choice. It isn’t aware that it is a clone. Whichever choice it makes is the prediction for your game.

You don’t know whether you are the real you or the temporary clone. Both the real you and the temporary clone have the same knowledge and are almost certain to reason the same way.

Do you choose one box or two boxes?

Do I care about the other version of myself? If I do, then this is a form of the Iterated Version. If I don’t, this is equivalent to Version 1 still.

Yes, but when is the event that causes Box 2 to contain nothing? Place that event on the time line, and then place your decision in Phase 2 on the time line… and then just look at it.

The predictor need not be perfect. What is at issue is the ground for their predictive success. Is it sensitive to the rational profile of the agent: whatever mechanisms there are that enable their rational deliberation process and rational behavior? If it is so sensitive, then what explains the predictor’s prediction (and thereby the content of the opaque box) is the rational considerations that ground the subject’s actions at the time when they deliberate and act. It’s true that the predictor is able to foresee the agent’s decision before the agent makes it. But this queer temporal ordering doesn’t reverse the causal relationship. It’s still the agent’s reason for acting that explains both their decision and, consequently, the content of the opaque box, with the predictor merely acting as a causal facilitator.

I assume the prediction for the clone is the ‘standard’ way?

Well, I’ll agree with @AlveK. If I care about both results, it’s an instance of the iterated version in which case I’ll one-box. If I only care about my result, then yeah I’ll two-box because it will give me $1000 more than one-boxing.

There is no prediction for the clone. It’s just asked to make a choice and then destroyed. Its answer is the prediction for your game.

Then your clone will almost certainly reason the same way, and so you will almost certainly walk away with $1,000.

Whereas I will choose to one-box, my clone will almost certainly reason the same way, and so I will almost certainly walk away with $1,000,000.

But let’s consider a variation: the clone makes its choice after you, but nonetheless determines the contents of your second box. Does this change your answer?

You say

The predictor need not be perfect

and then

It’s true that the predictor is able to foresee the agent’s decision before the agent makes it.

I understand that it is possible the predictor can make accurate predictions because he has access to our “rational profile” but it still doesn’t matter if we aren’t in a situation where he is perfect.

You were examined and the predictor predicted a deliberation and whatever. The important thing is he either placed a million or nothing in box 2 and you cannot change how the prediction was made once in the room. This means you should take all the money in front of you, i.e. take two boxes.

If you want to say “you’ll get $1000”, the fact that you take two boxes here isn’t the cause or the reason why you’ll have $1000, the cause of that was your rational profile that you already have no control over once you are in the room. So, as we are already in the room, if you took one box instead, you would have nothing.

But if you take one box, you were already the kind of person to have the rationality profile to take one box.