>So for the Right box commitment, there’d actually be a one in a trillion trillion chance of getting an empty Left. Although I would have committed to taking the Right box in that case, so it wouldn’t matter.
Wrong? Omega only predicts your behaviour in the bomb-visible case, you need not commit to anything in the other. Doubly surprising because "transparent Newcombs" formulations and discussions are often confused over what exactly is predicted, and the Bomb descriptions specifically fixes that.
I also wonder if this is the best example. It involves believing that a 1/epsilon event happened, *only* if you make a certain choice, which is strange at best. Why did you chose it over counterfactual mugging?
On the theory itself: Like probably many LWers, I have thought of some version of the Great Committment, and I think it would only make you updateless about things you learn after - the fully updateless version would seem worse to your current decisiontheory beforehand. In some ways this is good, it avoids truely insane stuff like "controlling whether you exist" (link), but in many ways it also doesnt go far enough, and it sure isnt theoretically elegant.
I agree that the practical impact is limited in the modest reading, but its somewhat important as a theoretical difference. Many potential counterfactual muggings are blocked by this, and other ideas that you should act mostly for Tegmark worlds youre not in.
EDT and CDT both think that the fundamental thing to evaluate is actions, while FDT thinks the fundamental thing to evaluate is algorithms for decision-making (or as Leonard Savage would have said, the one "grand" decision of what to do in every single contingency in every possibility that might come up, the full plan for one's life). CDT and FDT both think that we evaluate these things causally, by seeing what would be brought about from actually implementing the thing, while EDT says to evaluate it by what it is correlated with.
I think that CDT and FDT are both right, but address different questions. There's one question of what the rational act is, and one question of what the rational decision algorithm is. We naively think that these two should agree, in that whatever the rational decision algorithm is, and whatever act it outputs, must be whatever the rational act is. But Newcomb problems just show that this is wrong.
The rational decision algorithm to choose doesn't output the rational act. But that's just familiar from the conflict between act and rule utilitarianism.
Hi Kenny! Thanks for the comment. Weirdly enough, I actually ultimately disagree. As you know, proponents of FDT/UDT are doing something much stranger, in that they actually claim that the correct act is the locally irrational one. I think that they are correct, but that they just don't explain properly why this is the case. (My most recent post goes into much more detail on my view of the sociology of the situation, I'd be interested what you think). My personal opinion is that the study of decision theory inevitably leads to metaphysics - that it doesn't really make sense to talk about what is "rational" without getting into metaphysics.
1) It might be psychologically possible for me to make “the great commitment”, but it’s not as easy as flipping a switch and it’s not clear that it would really be the most productive use of my time and effort. Humans seem to function fine without having to make crazy pre-commitments, so the upside is unclear, and there is also the risk of making a mistake and locking in bad pre-commitments. Future AIs with better intelligence and ability to self-modify seem like they’ll be in a better position to worry about this stuff.
2) It actually might make a big difference to the future whether we make the commitments in a CDT way or an EDT way. EDT might suggest looking for ways to acausally cooperate with distant civilizations (Evidential Cooperation in Large Worlds), whereas CDT (and the commitments CDT recommends making) doesn’t. So the traditional academic philosophical debate might actually be more important than the LessWrong discourse
1) Since it’s a “moment of understanding” type thing, I think it might be as easy as flipping a switch!
The thing is, I purposely emphasized the vagueness of it so that there wouldn’t be worries of “oh, did I commit myself to something crazy” later - the correct notion of updatelessness is left completely open to further investigation! (and I agree it’s very non-obvious, cf the screenshots in the post). I claim that you shouldn’t be worried because if it turns out that there’s something crazy, you can just decide that you didn’t mean anything like that in the first place.
I do think there’s some practical upside in better understanding what we do on a day-to-day basis. cf “the hostile telepaths problem” and “newcomblike problems are the norm”. But of course I agree that it’s not a huge practical concern.
2) CDT definitely would if you use logical counterfactuals! I’ve kinda stopped thinking EDT vs CDT as a hugely important live debate ever since coming across Abram Demski’s EDT=CDT thing.
Ok, so, on how I think that DBDT can generate true updatelessness.
You imagine an idealized rational agent with all the information you had at the critical point. Then you ask yourself if such an agent would adopt the disposition to pay counterfactual muggers he encounters or not. (or *the one* counterfactual mugger the one time he will encounter him). Since your CP-self didn't know how the coin would fall, he adopts the disposition to pay (from his perspective it's higher EV).
This wouldn't require you to actually have considered the scenario at CP, you can just think about what disposition your idealized past self would have adopted.
Also idk if I was arguing past you a bit because I like 99% agree with you. Especially with that DBDT and commitment theory are actually orthogonal to what is rational, I think that's a really great point. I think the only differences between this DBDT version and your commitment theory is that
a) DBDT talks of dispositions instead of commitments ("adopt the disposition that is good to act out of" instead of "commit to those actions that would have been good to have committed to. I think I like this a bit more, since I feel like commitments are something more specific, like a "self-set external rule" while dispositions are just "the way you're inclined to act" (you can solve Parfit's Hitchhiker by being disposed to always keep your promises, and I think this plausibly involves no committment in the sense above. But maybe that's also just a verbal difference idk).
b) Fisher allows for those "Critical Points", and I really like that idea. I think it's probably true that our dispositions can change throughout our life (and that we can actively work on them), and the critical points adress that imo. Also I think setting an "universal critical point" at the beginning of time would actually mirror your "great commitment". And the search for where the critical points for our actual actions are would probably mirror your proposed search of which situations are, in which way, covered by the great commitment.
Yeah, that makes sense. The issue is that disposition-based decision theory still talks about many different dispositions, right? But yeah, the way I frame the great commitment as a psychological attachment at the end is kind if like a disposition, so something like that definitely works.
Yes, but I think if you want one overarching "great disposition" you can choose one disposition that governs any possible decision, with the CP (which is the information cutoff, basically) at the beginning of time.
But obviously it's impossible to actually adopt anything close to that (or to program a machine that does that, there's just way way way way way too many possibilities to consider) but what you can actually do is adopting different dispositions with different CP's for different decisions. (even if the "great disposition" would be more optimal)
I think it is possible though, like I argue in the post. It doesn't need to cover all the possibilities, it can be vague and the details can be filled in later. And if you do many dispositions, you'll fail on situations you didn't anticipate, like we discussed on Twitter
The thing I thought you obviously can't do is figuring out how exactly the "great disposition" looks like and then check for every decision what the great disposition tells you to do. Because for that you'd need to consider every possible path the universe could have taken (or at least all the paths were "you" exist). And I think this is also impossible in commitment language, the great commitment will always stay somewhat vague.
But yeah you can say that you can figure out if some particular course of action you consider is part of the great disposition/commitment without knowing the whole thing (if that's what you mean).
I wonder though why evolution has equipped humans with such strong "brute CDT" intuitions in so many cases. Maybe it's just an easier/more economic way of thinking that usually works just fine?
"even though we have no reason to think these other versions exist".
Huh? These other versions actually exist in a Tegmarkian multiverse, for example. Not saying that's necessarily the correct view, but "no reason" seems much too strong here.
> You know that Omega has a failure rate of one in a trillion trillion in this situation.
I am a bit confused by the wording of this bomb scenario.
Does this failure rate mean a failure rate of prediction or a failure rate of putting the bomb or not bomb correctly into the box despite having made a correct prediction?
Nice piece! It occurred to me a while ago that the claims that FDT gives different recommendations from EDT in cases like XOR blackmail could be avoided if we are thinking about a special case of EDT where we assume there are many diverging copies of you (as in the MWI, or if you are a simulation and a bunch of copies are made with small random perturbations), and you have a utility function that treats all your copies equally and tries to maximize average utility for all of them. This is equivalent to thinking about what pre-commitments you'd want to make before all the copies diverge in order to maximize expected utility. It also occurred to me later that the FDT recommendation to cooperate with other FDT agents (not divergent copies of you) in a one-shot prisoner's dilemma could be justified if we imagine that all pre-commitments are made behind a veil of ignorance about which agent we'll be, so that if there's a combination of choices both can make that allow them each to do better than the Nash equilibrium (i.e. the Pareto optimal choices differ from the Nash equilibrium, as in the prisoner's dilemma where both are better off if they cooperate than if they both defect, but the Nash equilibrium is that both defect), all such agents should pre-commit to choosing that.
Do you agree with seeing "the great commitment" as potentially a special case of EDT, or something different? Either way, do you think "EDT + imagining the most desirable pre-commitments from behind the veil of ignorance and sticking to those" would ever differ in its recommendations from FDT on any specific problems?
Yes, the thing you're thinking of is roughly EDT with alternate realities / updateless EDT! It is also my actually preferred framing. Seems like we converged on something similar :)
Though updateless vs updateful is generally taken to be an orthogonal axis to EDT vs CDT (this is Christiano's 2x2 if you want to ask an LLM about it), so commitment theory (which is basically just a different framing for updatelessness) wouldn't be a special case of EDT.
In terms of whether it ever differs from FDT, FDT is more of an umbrella term (the MIRI/OP exchange on decision theory would be a good thing to read here), so it's hard to say.
>But wait - Omega was predicting me, as usual, so this is definitely covered by the great commitment.
why is it covered? Also
>LessWrong-style decision theory
is an extreme name; I doubt almost anyone from LessWrong takes the left box. You are talking about Yudkowsky's FDT (Functional Decision Theory), and while many people on LW love Yudkowsky's Sequences including myself, they don't include FDT.
- If you let CDT make a grand commitment it turns into Son-of-CDT which is not the same as FDT and loses all precommitment races. https://www.lesswrong.com/w/son-of-cdt.
- @Easwaran: FDT does not evaluate 'algorithms' as things that can be varied by intervention. It evaluates 'outputs of rigid algorithms' as the variable whose value it itself is determining.
I'm afraid that from my perspective these are common misunderstandings of logical decision theory, not isomorphs that might be easier to explain.
Hi Eliezer, I address this in a footnote, will reproduce it here:
"A key point is that humans (and likely almost all agents in the universe when they realize that they should make the great commitment) face basically no Son-of-CDT issue - there is no one that has been predicting their pre-great-commitment self with enough fidelity for it to matter. (It would have to be high enough fidelity to distinguish between making the great commitment and adopting FDT - hard to imagine, since they are so similar)."
"...even though that would mean changing the past."
I must object to this framing. Imagine there's a calculator that calculated 56+8 yesterday, but I don't know what it answeres. So I put 56+8 into my calculator and it answers 64. Does my calculator "change the past"? No, I just learn the other calculator must also have answered 64.
It's really no different for FDT. The agent reaches a decision, and knows the predictor must have predicted it.
Let’s take Bomb*, where you need to make a decision that makes your current reality impossible (or extremely unlikely). I think this is fairly glossed as “trying to change the past”.
If it makes your current reality impossible, you were never there to begin with! You (your function) are instantiated at different points in spacetime, and each "you" feels the exact same from the inside. If one of these instatiations of you is impossible given your decision, then that version of you was never there to begin with.
Either way, there is no “changing the past”. I’m quite surprised there are FDT proponents - if I can count you among them - who say this. If a theory implies you can change the past, it's almost definitely wrong given our understanding of physics!
If we don't have an additional metaphysical claim (like Tegmark IV or platonism about computations) then there is nowhere that my computation is instantiated other than the situation I'm in. So it can't be impossible, since it's actually instantiated.
(To be clear, my actual opinion is that you have to dissolve the notion of existence. But that's also a metaphysical claim).
I guess I think the way people loosely talk about it sometimes is kind of like that. Joe Carlsmith also frames it in that way at least directionally. But yeah I'm not too attached to it.
Nate Soares in "Notes on "Can you control the past?"": "Parfit's hitchhiker, at the point where you're standing at the ATM, is much like this. The alleged problem statement is "you may either lose $0 or $10,000, and you're going to choose to lose $10,000". At which point we're like "Hold on a sec, the problem statement makes an assertion about my choice, which I can refute. What happens if I refute the problem statement?" At which point the question-poser is like "haha oops, yeah, if you refute the problem statement then you die alone in the desert". At which point, yeah, when the logical decision theorist closes their eyes and imagines stiffing the driver, then (under the assumption that the driver is accurate) they're like "oh dang, this would refute my observations; what happens in that case again? right, I'd die alone in the desert, which is worse than losing $10,000", and then they pay."
This is what I mean, it's definitely shading into "you can change the past" territory.
Well no, it doesn’t leave the question open (of whether it’s “rational” or not), it very consciously sidesteps it. You should just be so attached to this notion of “acting correctly” that you cannot help but follow it, no matter what seems rational at that point.
Simple, your past self would have precommitted to paying (since it would have been ignorant about which way the coin will come up), so you pay.
Right, I commented to quickly. It's like: "Do whatever you would have precommitted to do." You're right that it sidesteps my question; and indeed solves CM neatly. Nice.
Two notes on Bomb:
>So for the Right box commitment, there’d actually be a one in a trillion trillion chance of getting an empty Left. Although I would have committed to taking the Right box in that case, so it wouldn’t matter.
Wrong? Omega only predicts your behaviour in the bomb-visible case, you need not commit to anything in the other. Doubly surprising because "transparent Newcombs" formulations and discussions are often confused over what exactly is predicted, and the Bomb descriptions specifically fixes that.
I also wonder if this is the best example. It involves believing that a 1/epsilon event happened, *only* if you make a certain choice, which is strange at best. Why did you chose it over counterfactual mugging?
On the theory itself: Like probably many LWers, I have thought of some version of the Great Committment, and I think it would only make you updateless about things you learn after - the fully updateless version would seem worse to your current decisiontheory beforehand. In some ways this is good, it avoids truely insane stuff like "controlling whether you exist" (link), but in many ways it also doesnt go far enough, and it sure isnt theoretically elegant.
https://www.lesswrong.com/posts/4MYYr8YmN2fonASCi/you-re-in-newcomb-s-box
oh, you’re right of course. fixing that.
https://x.com/reconfigurthing/status/2075168426746687852
I agree that the practical impact is limited in the modest reading, but its somewhat important as a theoretical difference. Many potential counterfactual muggings are blocked by this, and other ideas that you should act mostly for Tegmark worlds youre not in.
I feel like I've been gesturing at this point in my academic decision theory papers about FDT.
Here's the recent (open-access) one, with Ben Levinstein and Ted Shear: https://link.springer.com/article/10.1007/s11238-025-10080-w
Here's the older one that probably isn't open access (but the pdf should be available on my website, at kennyeaswaran.org/research): https://www.jstor.org/stable/48692220
EDT and CDT both think that the fundamental thing to evaluate is actions, while FDT thinks the fundamental thing to evaluate is algorithms for decision-making (or as Leonard Savage would have said, the one "grand" decision of what to do in every single contingency in every possibility that might come up, the full plan for one's life). CDT and FDT both think that we evaluate these things causally, by seeing what would be brought about from actually implementing the thing, while EDT says to evaluate it by what it is correlated with.
I think that CDT and FDT are both right, but address different questions. There's one question of what the rational act is, and one question of what the rational decision algorithm is. We naively think that these two should agree, in that whatever the rational decision algorithm is, and whatever act it outputs, must be whatever the rational act is. But Newcomb problems just show that this is wrong.
The rational decision algorithm to choose doesn't output the rational act. But that's just familiar from the conflict between act and rule utilitarianism.
Hi Kenny! Thanks for the comment. Weirdly enough, I actually ultimately disagree. As you know, proponents of FDT/UDT are doing something much stranger, in that they actually claim that the correct act is the locally irrational one. I think that they are correct, but that they just don't explain properly why this is the case. (My most recent post goes into much more detail on my view of the sociology of the situation, I'd be interested what you think). My personal opinion is that the study of decision theory inevitably leads to metaphysics - that it doesn't really make sense to talk about what is "rational" without getting into metaphysics.
(this is the other post https://substack.com/home/post/p-206278520)
I think this is a good contribution to the debate
A couple points:
1) It might be psychologically possible for me to make “the great commitment”, but it’s not as easy as flipping a switch and it’s not clear that it would really be the most productive use of my time and effort. Humans seem to function fine without having to make crazy pre-commitments, so the upside is unclear, and there is also the risk of making a mistake and locking in bad pre-commitments. Future AIs with better intelligence and ability to self-modify seem like they’ll be in a better position to worry about this stuff.
2) It actually might make a big difference to the future whether we make the commitments in a CDT way or an EDT way. EDT might suggest looking for ways to acausally cooperate with distant civilizations (Evidential Cooperation in Large Worlds), whereas CDT (and the commitments CDT recommends making) doesn’t. So the traditional academic philosophical debate might actually be more important than the LessWrong discourse
Thanks Bz!
1) Since it’s a “moment of understanding” type thing, I think it might be as easy as flipping a switch!
The thing is, I purposely emphasized the vagueness of it so that there wouldn’t be worries of “oh, did I commit myself to something crazy” later - the correct notion of updatelessness is left completely open to further investigation! (and I agree it’s very non-obvious, cf the screenshots in the post). I claim that you shouldn’t be worried because if it turns out that there’s something crazy, you can just decide that you didn’t mean anything like that in the first place.
I do think there’s some practical upside in better understanding what we do on a day-to-day basis. cf “the hostile telepaths problem” and “newcomblike problems are the norm”. But of course I agree that it’s not a huge practical concern.
2) CDT definitely would if you use logical counterfactuals! I’ve kinda stopped thinking EDT vs CDT as a hugely important live debate ever since coming across Abram Demski’s EDT=CDT thing.
Ok, so, on how I think that DBDT can generate true updatelessness.
You imagine an idealized rational agent with all the information you had at the critical point. Then you ask yourself if such an agent would adopt the disposition to pay counterfactual muggers he encounters or not. (or *the one* counterfactual mugger the one time he will encounter him). Since your CP-self didn't know how the coin would fall, he adopts the disposition to pay (from his perspective it's higher EV).
This wouldn't require you to actually have considered the scenario at CP, you can just think about what disposition your idealized past self would have adopted.
Also idk if I was arguing past you a bit because I like 99% agree with you. Especially with that DBDT and commitment theory are actually orthogonal to what is rational, I think that's a really great point. I think the only differences between this DBDT version and your commitment theory is that
a) DBDT talks of dispositions instead of commitments ("adopt the disposition that is good to act out of" instead of "commit to those actions that would have been good to have committed to. I think I like this a bit more, since I feel like commitments are something more specific, like a "self-set external rule" while dispositions are just "the way you're inclined to act" (you can solve Parfit's Hitchhiker by being disposed to always keep your promises, and I think this plausibly involves no committment in the sense above. But maybe that's also just a verbal difference idk).
b) Fisher allows for those "Critical Points", and I really like that idea. I think it's probably true that our dispositions can change throughout our life (and that we can actively work on them), and the critical points adress that imo. Also I think setting an "universal critical point" at the beginning of time would actually mirror your "great commitment". And the search for where the critical points for our actual actions are would probably mirror your proposed search of which situations are, in which way, covered by the great commitment.
Yeah, that makes sense. The issue is that disposition-based decision theory still talks about many different dispositions, right? But yeah, the way I frame the great commitment as a psychological attachment at the end is kind if like a disposition, so something like that definitely works.
Yes, but I think if you want one overarching "great disposition" you can choose one disposition that governs any possible decision, with the CP (which is the information cutoff, basically) at the beginning of time.
But obviously it's impossible to actually adopt anything close to that (or to program a machine that does that, there's just way way way way way too many possibilities to consider) but what you can actually do is adopting different dispositions with different CP's for different decisions. (even if the "great disposition" would be more optimal)
I think it is possible though, like I argue in the post. It doesn't need to cover all the possibilities, it can be vague and the details can be filled in later. And if you do many dispositions, you'll fail on situations you didn't anticipate, like we discussed on Twitter
Ok wait maybe it is
The thing I thought you obviously can't do is figuring out how exactly the "great disposition" looks like and then check for every decision what the great disposition tells you to do. Because for that you'd need to consider every possible path the universe could have taken (or at least all the paths were "you" exist). And I think this is also impossible in commitment language, the great commitment will always stay somewhat vague.
But yeah you can say that you can figure out if some particular course of action you consider is part of the great disposition/commitment without knowing the whole thing (if that's what you mean).
I wonder though why evolution has equipped humans with such strong "brute CDT" intuitions in so many cases. Maybe it's just an easier/more economic way of thinking that usually works just fine?
No, that's not what I'm saying, I'm saying you can figure out whether the great commitment applies only when you actually get into a given situation.
I'm not sure if it has, many people have EDT intuitions on Newcomb's, and concepts like integrity and honor have an FDT flavor to them.
(Will comment on the post as a whole later!)
"even though we have no reason to think these other versions exist".
Huh? These other versions actually exist in a Tegmarkian multiverse, for example. Not saying that's necessarily the correct view, but "no reason" seems much too strong here.
Yes, I’m speaking to a “common sense” perspective
> You know that Omega has a failure rate of one in a trillion trillion in this situation.
I am a bit confused by the wording of this bomb scenario.
Does this failure rate mean a failure rate of prediction or a failure rate of putting the bomb or not bomb correctly into the box despite having made a correct prediction?
Of prediction!
Nice piece! It occurred to me a while ago that the claims that FDT gives different recommendations from EDT in cases like XOR blackmail could be avoided if we are thinking about a special case of EDT where we assume there are many diverging copies of you (as in the MWI, or if you are a simulation and a bunch of copies are made with small random perturbations), and you have a utility function that treats all your copies equally and tries to maximize average utility for all of them. This is equivalent to thinking about what pre-commitments you'd want to make before all the copies diverge in order to maximize expected utility. It also occurred to me later that the FDT recommendation to cooperate with other FDT agents (not divergent copies of you) in a one-shot prisoner's dilemma could be justified if we imagine that all pre-commitments are made behind a veil of ignorance about which agent we'll be, so that if there's a combination of choices both can make that allow them each to do better than the Nash equilibrium (i.e. the Pareto optimal choices differ from the Nash equilibrium, as in the prisoner's dilemma where both are better off if they cooperate than if they both defect, but the Nash equilibrium is that both defect), all such agents should pre-commit to choosing that.
Do you agree with seeing "the great commitment" as potentially a special case of EDT, or something different? Either way, do you think "EDT + imagining the most desirable pre-commitments from behind the veil of ignorance and sticking to those" would ever differ in its recommendations from FDT on any specific problems?
Yes, the thing you're thinking of is roughly EDT with alternate realities / updateless EDT! It is also my actually preferred framing. Seems like we converged on something similar :)
Though updateless vs updateful is generally taken to be an orthogonal axis to EDT vs CDT (this is Christiano's 2x2 if you want to ask an LLM about it), so commitment theory (which is basically just a different framing for updatelessness) wouldn't be a special case of EDT.
In terms of whether it ever differs from FDT, FDT is more of an umbrella term (the MIRI/OP exchange on decision theory would be a good thing to read here), so it's hard to say.
You lost me at
>But wait - Omega was predicting me, as usual, so this is definitely covered by the great commitment.
why is it covered? Also
>LessWrong-style decision theory
is an extreme name; I doubt almost anyone from LessWrong takes the left box. You are talking about Yudkowsky's FDT (Functional Decision Theory), and while many people on LW love Yudkowsky's Sequences including myself, they don't include FDT.
- If you let CDT make a grand commitment it turns into Son-of-CDT which is not the same as FDT and loses all precommitment races. https://www.lesswrong.com/w/son-of-cdt.
- @Easwaran: FDT does not evaluate 'algorithms' as things that can be varied by intervention. It evaluates 'outputs of rigid algorithms' as the variable whose value it itself is determining.
I'm afraid that from my perspective these are common misunderstandings of logical decision theory, not isomorphs that might be easier to explain.
Hi Eliezer, I address this in a footnote, will reproduce it here:
"A key point is that humans (and likely almost all agents in the universe when they realize that they should make the great commitment) face basically no Son-of-CDT issue - there is no one that has been predicting their pre-great-commitment self with enough fidelity for it to matter. (It would have to be high enough fidelity to distinguish between making the great commitment and adopting FDT - hard to imagine, since they are so similar)."
"...even though that would mean changing the past."
I must object to this framing. Imagine there's a calculator that calculated 56+8 yesterday, but I don't know what it answeres. So I put 56+8 into my calculator and it answers 64. Does my calculator "change the past"? No, I just learn the other calculator must also have answered 64.
It's really no different for FDT. The agent reaches a decision, and knows the predictor must have predicted it.
Let’s take Bomb*, where you need to make a decision that makes your current reality impossible (or extremely unlikely). I think this is fairly glossed as “trying to change the past”.
If it makes your current reality impossible, you were never there to begin with! You (your function) are instantiated at different points in spacetime, and each "you" feels the exact same from the inside. If one of these instatiations of you is impossible given your decision, then that version of you was never there to begin with.
I understand that. But to make this work in situations where you are not being simulated, you need an additional metaphysical claim, e.g. Tegmark IV
Do we need that?
Either way, there is no “changing the past”. I’m quite surprised there are FDT proponents - if I can count you among them - who say this. If a theory implies you can change the past, it's almost definitely wrong given our understanding of physics!
If we don't have an additional metaphysical claim (like Tegmark IV or platonism about computations) then there is nowhere that my computation is instantiated other than the situation I'm in. So it can't be impossible, since it's actually instantiated.
(To be clear, my actual opinion is that you have to dissolve the notion of existence. But that's also a metaphysical claim).
I guess I think the way people loosely talk about it sometimes is kind of like that. Joe Carlsmith also frames it in that way at least directionally. But yeah I'm not too attached to it.
Nate Soares in "Notes on "Can you control the past?"": "Parfit's hitchhiker, at the point where you're standing at the ATM, is much like this. The alleged problem statement is "you may either lose $0 or $10,000, and you're going to choose to lose $10,000". At which point we're like "Hold on a sec, the problem statement makes an assertion about my choice, which I can refute. What happens if I refute the problem statement?" At which point the question-poser is like "haha oops, yeah, if you refute the problem statement then you die alone in the desert". At which point, yeah, when the logical decision theorist closes their eyes and imagines stiffing the driver, then (under the assumption that the driver is accurate) they're like "oh dang, this would refute my observations; what happens in that case again? right, I'd die alone in the desert, which is worse than losing $10,000", and then they pay."
This is what I mean, it's definitely shading into "you can change the past" territory.
"... other than the situation I'm in."
But then you're just only instantiated in the predictor's mind!
Interesting idea! Though it seems to leave open te question of *why* to act on your great commitment later.
Also, how does Commitment Theory answer Counterfactual Mugging?
Thanks!
Well no, it doesn’t leave the question open (of whether it’s “rational” or not), it very consciously sidesteps it. You should just be so attached to this notion of “acting correctly” that you cannot help but follow it, no matter what seems rational at that point.
Simple, your past self would have precommitted to paying (since it would have been ignorant about which way the coin will come up), so you pay.
Right, I commented to quickly. It's like: "Do whatever you would have precommitted to do." You're right that it sidesteps my question; and indeed solves CM neatly. Nice.
Thank you🙏
Ah, this is what I talked about in a LessWrong post about unrolling a loop and choosing which bot to be.