Thanks for writing this up and sharing. I find myself pretty sympathetic to the idea that people generally do better when they focus on the first order consequences of their actions and I appreciate this as a formalization of that intuition.
bracketing could provide action-guidance in the face of cluelessness and in particular support neartermism
I interpret âsupport neartermismâ to mean something like âdonating to AMF is better than donating to AI safety.â Is that what you mean? If so, do you have an example where this is true?
I buy that itâs an argument against very crazy town actions like acausal trade, but most âlongtermistâ worldviews place a relatively high probability on e.g. AI killing a bunch of bed net recipients soon, so I would expect that you canât actually bracket out just the far future ones and find that AMF is c-better. (See e.g. EA and Longtermism: not a crux for saving the world.)
Bracketing doesnât in general recommend focusing on the âfirst order consequencesâ, in the sense people usually use that term (e.g. the first step in some coarse-grained causal pathway). There can be locations of value Iâ where weâd think A >_{Iâ} B if we only considered first order consequences, yet A [incomparable]_{Iâ} B all things considered. Conversely, there can be locations of value Iâ that are only affected by higher-order consequences, yet A >_{Iâ} B.
Not sure exactly what you mean by âgenerally do betterâ, but just to be clear: Bracketing is its own theory of what it means to âdo betterâ as an impartial altruist, not a formalization of a heuristic for getting higher EV. (Jesse says as much in the summary.)
On your second point: Even if AI would plausibly kill the bed net recipients soon, we also need to say whether (1) any concrete intervention weâre aware of would decrease AI risk in expectation, and (2) that intervention would be more cost-effective for bracketed-in welfare than the alternatives, if so.
Iâm skeptical of (1), briefly, because whether an intervention prevents vs. causes an AI x-risk seems sensitive to various dynamics that we have little evidence about + are too unfamiliar for us to trust our intuitions about. More on this here.
Re: (2), if weâre bracketing out the far-future consequences, I expect itâs hard to argue that AI risk work is more cost-effective than the best animal welfare opportunities. (Less confident in this point than the previous one, conditional on (1).)
Thanks for writing this up and sharing. I find myself pretty sympathetic to the idea that people generally do better when they focus on the first order consequences of their actions and I appreciate this as a formalization of that intuition.
I interpret âsupport neartermismâ to mean something like âdonating to AMF is better than donating to AI safety.â Is that what you mean? If so, do you have an example where this is true?
I buy that itâs an argument against very crazy town actions like acausal trade, but most âlongtermistâ worldviews place a relatively high probability on e.g. AI killing a bunch of bed net recipients soon, so I would expect that you canât actually bracket out just the far future ones and find that AMF is c-better. (See e.g. EA and Longtermism: not a crux for saving the world.)
Thanks Ben â a few clarifications:
Bracketing doesnât in general recommend focusing on the âfirst order consequencesâ, in the sense people usually use that term (e.g. the first step in some coarse-grained causal pathway). There can be locations of value Iâ where weâd think A >_{Iâ} B if we only considered first order consequences, yet A [incomparable]_{Iâ} B all things considered. Conversely, there can be locations of value Iâ that are only affected by higher-order consequences, yet A >_{Iâ} B.
Not sure exactly what you mean by âgenerally do betterâ, but just to be clear: Bracketing is its own theory of what it means to âdo betterâ as an impartial altruist, not a formalization of a heuristic for getting higher EV. (Jesse says as much in the summary.)
On your second point: Even if AI would plausibly kill the bed net recipients soon, we also need to say whether (1) any concrete intervention weâre aware of would decrease AI risk in expectation, and (2) that intervention would be more cost-effective for bracketed-in welfare than the alternatives, if so.
Iâm skeptical of (1), briefly, because whether an intervention prevents vs. causes an AI x-risk seems sensitive to various dynamics that we have little evidence about + are too unfamiliar for us to trust our intuitions about. More on this here.
Re: (2), if weâre bracketing out the far-future consequences, I expect itâs hard to argue that AI risk work is more cost-effective than the best animal welfare opportunities. (Less confident in this point than the previous one, conditional on (1).)