"Note that Omega’s prediction is made after you choose Room 2, if you do choose that."
Perhaps it's worth noting that that's not true in Spencer's original problem statement, which (in note 6) says that the Frustrater made the prediction yesterday.
Didn’t see the footnote—thanks for this! I might be wrong, but this seems like a mistake on Spencer’s part? If Omega made the prediction before the choice of which room to go in, then I’d think they’d prefer Room 2 (and then picking A or B) to Room 1 for the same reason they’d pick A or B over C.
I can see Spencer’s later reasoning; he seems to be thinking that, from the perspective of the first choice, which dependency hypothesis holds gives evidence as to what the agent will later choose (which it does), so that the agent takes into account their probability of picking A given that Omega predicted A. In other words, they reason as follows: “Assume Omega predicted A; then if I pick Room 1, I get 35. If I pick Room 2, then I’ll probably pick Box A since Omega is accurate, and I’ll probably get nothing. Thus, Room 1 is better.”
This way of taking into account one’s future choices seems to be against the spirit of CDT. It would imply, for example, that after you’re scanned in Newcomb’s problem, the CDT agent would pay $1 to have someone force them to one-box, which is not a verdict I think CDTers would like. Did I screw up somewhere?
(My way of understanding CDT would have the CDTer choose from among strategies at earlier nodes, under the condition that choices at later nodes also maximize causal EU, analogous to how subgame perfect equilibrium is defined)
At any rate, even if Spencer’s argument is a-okay, I hope at least that having the prediction made after the first choice still works so that my own setup is internally okay.
I think CDT enters Room 1 regardless of when the prediction happens. Even if the prediction already happened, Carl the CDT'er knows that after entering Room 2, he will take one of the boxes - just as he does in the original Frustrater problem. And just like in Death in Damascus, Carl never actually believes he can "cheat" the predictor. He knows he won't make any money.
I wonder if you could cash out some of the FDT counterpossible reasoning stuff via some sort of relevance logic. You wouldn't need to actually adopt the relevance logic (or whatever) as "true," much less provide a semantics for counterpossibles, just base your calculations partially on what adopting this logic would imply. "If X is a theorem in some formal system I reject, then do A, otherwise do B," etc.
You make it sound as if the main problem with counterpossibles is that we haven't figured them out yet. But I would argue that the exact opposite is the case: We know *very much* about them (the literature on them is enormous) and have good grounds to think that vacuism is true, e.g. because non-vacuism would break principles that seem obviously true, such as substitution of necessarily co-referring terms salva veritate or basically any basic logical inference rule ("If conjunction elimination were to mislead us and Trump were president, then Trump would be president" would not come out true).
Almost all of the major theories of counterfactuals entail vacuism and all the non-vacuist proposals either have totally insane implications or are hopelessly ad hoc, despite non-vacuists having spent decades to make it work. But, of course, if vacuism is true, then FDT is dead in the water.
Object-level, I just don’t think vacuism is very likely to be true. (For those reading this, Dominik’s language may seem to suggest that there’s some kind of consensus in favor of vacuism, which is not so.) IMO it seems like it just has momentum from the bygone pre-hyperintensional period in metaphysics/language. Case in point: that normal counterfactuals are merely sensitive to intension is not even remotely obvious, once we see what kinds of things are hyperintensional. A priori we should expect the opposite: there’s a close connection between counterfactuals and relations of dependence, the latter of which can be hyperintensional (grounding definitely, causation probably). Plus there’s independent reason to think ordinary counterfactuals are hyperintensional (eg the seesaw case).
It’s strictly true that people have been trying to make nonvacuism work for decades, but misleading. If I’m being super generous I could say the hyperintensionality boom (a necessary thing to have as far as nonvacuism goes) started in the 90s with Essence and Modality, but the real progress was made largely in the 2010s. That there isn’t a satisfactory substantive account of counterpossibles at this time is basically no evidence as to the truth of vacuism. As if we should expect views that are true to be the ones that are easy to develop.
Meta-level, I don’t agree that FDT would be dead in the water of vacuism were true. Not any more than a skeptical view of counterfactuals would be fatal to CDT. You’d just reformulate CDT without counterfactuals. Counterfactuals are useful, and therefore they’re either straightforwardly significant or a good heuristic for something else; if the latter, then CDT can be formulated in terms of what they’re a heuristic for. Same with counterpossibles. (One may reasonably worry about goalpost-moving here, but I do think some amount of reformulation is tolerable. I’d expect something at least as similar to Yudkowsky-FDT as Pearl-style CDT is to counterfactual-style CDT.)
> (For those reading this, Dominik’s language may seem to suggest that there’s some kind of consensus in favor of vacuism, which is not so.
Most philosophers writing on counterfactuals are vacuists. This is obviously true. Even among those who specifically *entertain* the (in my opinion: crazy) thesis that counterpossibles can be non-trivial, there are - in my estimation - probably more vacuists than non-vacuists. So, despite the obvious selection-effect, non-vacuism still loses. So any appeal to authority obviously goes in favour of vacuism.
But I didn't just appeal to popular consensus - I provided arguments to think that non-vacuism is insane. You provided no good response to the CRAZY result that "If A and B, then B" might turn out not true if your theory is right.
> Not any more than a skeptical view of counterfactuals would be fatal to CDT.
????? Obviously a sceptical view of counterfactuals would be fatal to CDT, that is completely uncontroversial. But a sceptical view of counterfactuals is utter nonsense that is endorsed by almost no one, for good reason.
If you think CDT could be reformulated without counterfactuals, then you haven't understood CDT.
It seems that CDT would recommend two-boxing given a causal structure where Omega’s prediction cannot be affected backwards in time by my choice of two-boxing. If I am not fully sure of that, then it could be rational under CDT to one box. That backwards-in-time causality is not possible is an additional assumption, that intuitive one-boxers implicitly question. One-boxing is rational just the way superstition is: “just in case” given we don’t really know the causal structure of reality. Does this make sense?
> If, alternatively, we do assume the lesion correlates with how you think about decision theory—so, the lesion causes you to be more likely to refrain on the basis that doing so is evidence of not getting cancer—then this just becomes a version of Newcomb’s problem where Omega is replaced by a gene, and it no longer becomes so counterintuitive to refrain from smoking.
I'm unclear as to how we get here. Up until that point, it seems like you'd established that EDTs and CDTs should both choose to smoke. In that case, what's the effect of this lesion (the new version) supposed to be? I don't have a clear interpretation of what you mean here (happy to respond to a clarified version), but typically I've found that when forced to specify details, something breaks.
Here's my attempt at a general explanation:
Suppose EDT supports not-smoking. For smoking to be evidence of the lesion, we require that the lesion causes some EDTs to smoke (in spite of their decision theory). Thus the EDT's thought process is "my decision theory tells me not to smoke, so conditional on smoking, I've probably been coerced by the lesion, which would be bad news" – but in this case, what's the EDT meant to do? "Try really hard not to be coerced"? Note that even if this is an available action, then even though it will increase your success rate at not-smoking, it still doesn't convince you that you don't have the lesion ("I tried hard enough that the lesion wouldn't have been able to coerce me into smoking. So I still don't have any evidence for/against the lesion")
"Note that Omega’s prediction is made after you choose Room 2, if you do choose that."
Perhaps it's worth noting that that's not true in Spencer's original problem statement, which (in note 6) says that the Frustrater made the prediction yesterday.
Didn’t see the footnote—thanks for this! I might be wrong, but this seems like a mistake on Spencer’s part? If Omega made the prediction before the choice of which room to go in, then I’d think they’d prefer Room 2 (and then picking A or B) to Room 1 for the same reason they’d pick A or B over C.
I can see Spencer’s later reasoning; he seems to be thinking that, from the perspective of the first choice, which dependency hypothesis holds gives evidence as to what the agent will later choose (which it does), so that the agent takes into account their probability of picking A given that Omega predicted A. In other words, they reason as follows: “Assume Omega predicted A; then if I pick Room 1, I get 35. If I pick Room 2, then I’ll probably pick Box A since Omega is accurate, and I’ll probably get nothing. Thus, Room 1 is better.”
This way of taking into account one’s future choices seems to be against the spirit of CDT. It would imply, for example, that after you’re scanned in Newcomb’s problem, the CDT agent would pay $1 to have someone force them to one-box, which is not a verdict I think CDTers would like. Did I screw up somewhere?
(My way of understanding CDT would have the CDTer choose from among strategies at earlier nodes, under the condition that choices at later nodes also maximize causal EU, analogous to how subgame perfect equilibrium is defined)
At any rate, even if Spencer’s argument is a-okay, I hope at least that having the prediction made after the first choice still works so that my own setup is internally okay.
At any rate, thanks so much for pointing this out!
I think CDT enters Room 1 regardless of when the prediction happens. Even if the prediction already happened, Carl the CDT'er knows that after entering Room 2, he will take one of the boxes - just as he does in the original Frustrater problem. And just like in Death in Damascus, Carl never actually believes he can "cheat" the predictor. He knows he won't make any money.
So it seems to me your setup is fine!
I wonder if you could cash out some of the FDT counterpossible reasoning stuff via some sort of relevance logic. You wouldn't need to actually adopt the relevance logic (or whatever) as "true," much less provide a semantics for counterpossibles, just base your calculations partially on what adopting this logic would imply. "If X is a theorem in some formal system I reject, then do A, otherwise do B," etc.
You make it sound as if the main problem with counterpossibles is that we haven't figured them out yet. But I would argue that the exact opposite is the case: We know *very much* about them (the literature on them is enormous) and have good grounds to think that vacuism is true, e.g. because non-vacuism would break principles that seem obviously true, such as substitution of necessarily co-referring terms salva veritate or basically any basic logical inference rule ("If conjunction elimination were to mislead us and Trump were president, then Trump would be president" would not come out true).
Almost all of the major theories of counterfactuals entail vacuism and all the non-vacuist proposals either have totally insane implications or are hopelessly ad hoc, despite non-vacuists having spent decades to make it work. But, of course, if vacuism is true, then FDT is dead in the water.
Object-level, I just don’t think vacuism is very likely to be true. (For those reading this, Dominik’s language may seem to suggest that there’s some kind of consensus in favor of vacuism, which is not so.) IMO it seems like it just has momentum from the bygone pre-hyperintensional period in metaphysics/language. Case in point: that normal counterfactuals are merely sensitive to intension is not even remotely obvious, once we see what kinds of things are hyperintensional. A priori we should expect the opposite: there’s a close connection between counterfactuals and relations of dependence, the latter of which can be hyperintensional (grounding definitely, causation probably). Plus there’s independent reason to think ordinary counterfactuals are hyperintensional (eg the seesaw case).
It’s strictly true that people have been trying to make nonvacuism work for decades, but misleading. If I’m being super generous I could say the hyperintensionality boom (a necessary thing to have as far as nonvacuism goes) started in the 90s with Essence and Modality, but the real progress was made largely in the 2010s. That there isn’t a satisfactory substantive account of counterpossibles at this time is basically no evidence as to the truth of vacuism. As if we should expect views that are true to be the ones that are easy to develop.
Meta-level, I don’t agree that FDT would be dead in the water of vacuism were true. Not any more than a skeptical view of counterfactuals would be fatal to CDT. You’d just reformulate CDT without counterfactuals. Counterfactuals are useful, and therefore they’re either straightforwardly significant or a good heuristic for something else; if the latter, then CDT can be formulated in terms of what they’re a heuristic for. Same with counterpossibles. (One may reasonably worry about goalpost-moving here, but I do think some amount of reformulation is tolerable. I’d expect something at least as similar to Yudkowsky-FDT as Pearl-style CDT is to counterfactual-style CDT.)
> (For those reading this, Dominik’s language may seem to suggest that there’s some kind of consensus in favor of vacuism, which is not so.
Most philosophers writing on counterfactuals are vacuists. This is obviously true. Even among those who specifically *entertain* the (in my opinion: crazy) thesis that counterpossibles can be non-trivial, there are - in my estimation - probably more vacuists than non-vacuists. So, despite the obvious selection-effect, non-vacuism still loses. So any appeal to authority obviously goes in favour of vacuism.
But I didn't just appeal to popular consensus - I provided arguments to think that non-vacuism is insane. You provided no good response to the CRAZY result that "If A and B, then B" might turn out not true if your theory is right.
> Not any more than a skeptical view of counterfactuals would be fatal to CDT.
????? Obviously a sceptical view of counterfactuals would be fatal to CDT, that is completely uncontroversial. But a sceptical view of counterfactuals is utter nonsense that is endorsed by almost no one, for good reason.
If you think CDT could be reformulated without counterfactuals, then you haven't understood CDT.
It seems that CDT would recommend two-boxing given a causal structure where Omega’s prediction cannot be affected backwards in time by my choice of two-boxing. If I am not fully sure of that, then it could be rational under CDT to one box. That backwards-in-time causality is not possible is an additional assumption, that intuitive one-boxers implicitly question. One-boxing is rational just the way superstition is: “just in case” given we don’t really know the causal structure of reality. Does this make sense?
> If, alternatively, we do assume the lesion correlates with how you think about decision theory—so, the lesion causes you to be more likely to refrain on the basis that doing so is evidence of not getting cancer—then this just becomes a version of Newcomb’s problem where Omega is replaced by a gene, and it no longer becomes so counterintuitive to refrain from smoking.
I'm unclear as to how we get here. Up until that point, it seems like you'd established that EDTs and CDTs should both choose to smoke. In that case, what's the effect of this lesion (the new version) supposed to be? I don't have a clear interpretation of what you mean here (happy to respond to a clarified version), but typically I've found that when forced to specify details, something breaks.
Here's my attempt at a general explanation:
Suppose EDT supports not-smoking. For smoking to be evidence of the lesion, we require that the lesion causes some EDTs to smoke (in spite of their decision theory). Thus the EDT's thought process is "my decision theory tells me not to smoke, so conditional on smoking, I've probably been coerced by the lesion, which would be bad news" – but in this case, what's the EDT meant to do? "Try really hard not to be coerced"? Note that even if this is an available action, then even though it will increase your success rate at not-smoking, it still doesn't convince you that you don't have the lesion ("I tried hard enough that the lesion wouldn't have been able to coerce me into smoking. So I still don't have any evidence for/against the lesion")