TDT is very effective for agents with the computational power to run it on top of a powerful epistemology that accurately notices other TDT-like agents whose decisions will correlate with their own.
This does not describe humans on juries worried about the decisions of other unspecified humans on other juries.
TDT is very effective for agents with the computational power to run it on top of a powerful epistemology that accurately notices other TDT-like agents whose decisions will correlate with their own.
Every formal, well-justified decision theory is that way. That’s no reason to avoid tractable approximations of them.
If you have no problem basing your decision on “greater good” (utilitarian) considerations, or Golden Rule/Categorical Imperative type reasoning, all of which are computationally intractable and do not perfectly describe humans, then you shouldn’t make this objection to TDT either.
I am not saying to avoid tractable approximations. But it is not likely that other people on other juries are using or know about TDT or any tractable approximations.
TDT’s optimality does not hinge on whether others use it; it’s just that TDT can handle other TDTers as a special case. It is still optimal for interacting with non-TDTers, so this doesn’t show why he shouldn’t be trying to approximate it here.
Do you think that any reasonable approximation of TDT would say anything other than ignore those unspecified other people who are not TDT-like and may as well be pieces of cardboard with whatever decision they will make written on them, and just do the right thing for the case you are presented with?
Wait—so you think tractable approximations of TDT will still lead to the right conclusion, but that Psy-Kosh still shouldn’t use it, even for practice?
In any case, yes, at the appropriate level of granularity, there are relevant correlations between yours and others decision processes—specifically, at the level of whether you deem it optimal to use discretion to bend application of laws in ways that favor your values. See komponisto’s and my long comments on this discussion for more detail.
But FWIW, yes, I do find this reasoning from acausal correlations (that TDT relies on) to be problematic and worthy of more scrutiny.
Wait—so you think tractable approximations of TDT will still lead to the right conclusion, but that Psy-Kosh still shouldn’t use it, even for practice?
No. I advised him of the output of me running my tractable approximation of TDT: focus on the right thing for the presented case.
In any case, yes, at the appropriate level of granularity, there are relevant correlations between yours and others decision processes—specifically, at the level of whether you deem it optimal to use discretion to bend application of laws in ways that favor your values. See komponisto’s and my long comments on this discussion for more detail.
This is incorrect. Suppose you and other humans start out using variations CHDT (Crazy Human Decision Theory), which have some level of correlation with each other, but do not take this correlation into account when making decisions. Then you learn about TDT, think it is a good idea, and incorporate some of its ideas into some variation THDT (Timeless Human Decision Theory), which ends up being something like run CHDT but override some decisions if principles derived from TDT make it seem like a good idea. So you look at this jury nullification problem, and start running CHDT which returns that it seems like a good policy. But TDT says that if your decisions with others are correlated with other agents, you should take that into account. And you don’t like what you predict would happen if everyone embraced jury nullification, so you change your decision. At this point you have destroyed the correlation between your decision and the normal CHDT using humans because you overrode the component of your decision theory that works the same way.
This could work if the other humans are also using THDT, because then they could override their CHDT answers for the same reasons you do. But we are talking about juries drawn from a population of mostly non-rationalists who have never heard of TDT. The decisions you make differently because you know about TDT do not correlate with their decisions.
Most of the interesting effects of TDT do seem to relate to the case of other agents also using TDT, afaik. (Although there’s that other post about the TDT paradox, of course...)
What if TDT is optimally ethical?
You know, its major selling point to begin with?
TDT is very effective for agents with the computational power to run it on top of a powerful epistemology that accurately notices other TDT-like agents whose decisions will correlate with their own.
This does not describe humans on juries worried about the decisions of other unspecified humans on other juries.
Every formal, well-justified decision theory is that way. That’s no reason to avoid tractable approximations of them.
If you have no problem basing your decision on “greater good” (utilitarian) considerations, or Golden Rule/Categorical Imperative type reasoning, all of which are computationally intractable and do not perfectly describe humans, then you shouldn’t make this objection to TDT either.
I am not saying to avoid tractable approximations. But it is not likely that other people on other juries are using or know about TDT or any tractable approximations.
TDT’s optimality does not hinge on whether others use it; it’s just that TDT can handle other TDTers as a special case. It is still optimal for interacting with non-TDTers, so this doesn’t show why he shouldn’t be trying to approximate it here.
Do you think that any reasonable approximation of TDT would say anything other than ignore those unspecified other people who are not TDT-like and may as well be pieces of cardboard with whatever decision they will make written on them, and just do the right thing for the case you are presented with?
Wait—so you think tractable approximations of TDT will still lead to the right conclusion, but that Psy-Kosh still shouldn’t use it, even for practice?
In any case, yes, at the appropriate level of granularity, there are relevant correlations between yours and others decision processes—specifically, at the level of whether you deem it optimal to use discretion to bend application of laws in ways that favor your values. See komponisto’s and my long comments on this discussion for more detail.
But FWIW, yes, I do find this reasoning from acausal correlations (that TDT relies on) to be problematic and worthy of more scrutiny.
No. I advised him of the output of me running my tractable approximation of TDT: focus on the right thing for the presented case.
This is incorrect. Suppose you and other humans start out using variations CHDT (Crazy Human Decision Theory), which have some level of correlation with each other, but do not take this correlation into account when making decisions. Then you learn about TDT, think it is a good idea, and incorporate some of its ideas into some variation THDT (Timeless Human Decision Theory), which ends up being something like run CHDT but override some decisions if principles derived from TDT make it seem like a good idea. So you look at this jury nullification problem, and start running CHDT which returns that it seems like a good policy. But TDT says that if your decisions with others are correlated with other agents, you should take that into account. And you don’t like what you predict would happen if everyone embraced jury nullification, so you change your decision. At this point you have destroyed the correlation between your decision and the normal CHDT using humans because you overrode the component of your decision theory that works the same way.
This could work if the other humans are also using THDT, because then they could override their CHDT answers for the same reasons you do. But we are talking about juries drawn from a population of mostly non-rationalists who have never heard of TDT. The decisions you make differently because you know about TDT do not correlate with their decisions.
Most of the interesting effects of TDT do seem to relate to the case of other agents also using TDT, afaik. (Although there’s that other post about the TDT paradox, of course...)