On “describing base on algorithim vs choice” I believe the point was being made that Omega/Nature does not go “Are you running algorithm number 1163527? If so, you get screwed.” You can have two algorithms that are functionally equivalent (same input output pairs) yet operationally different.
On “Why doesn’t CDT onebox?” I think it would be very helpful to imagine the CDT agent not as a person, but as a mechanical calculator. Its quite hard to faithfully simulate a specific decision algorithm in one’s head, because all of our decision making intuition is constantly trying to sneak in, and I’ve confused myself many times by “imagining an XYZ agent” but giving them more tools than XYZ prescribes. It’s not that CDT considers it more reasonable to use a causal model than to just pick the winning choice, it’s that a causal model is all a CDT agent has (plus a little bit of extra circuitry).
My guess at a possible confusion you might be having, is that when you are thinking of causal models, you are including things that are subjunctively dependent, when a causal model isn’t framed/designed to handle subjunctively dependent things. From the FDT paper:
The basic intuition behind FDT is that there is some respect in which predictor-like things depend upon an agent’s future action, and lesion-like things do not. We’ll call this kind of dependence subjunctive dependence to distinguish it from (e.g.) straightforward causal dependence and purely statistical dependence.
Since we often automatically take subjunctive dependence into account, it feels like a causal model should take it into account. But the existing well defined causal models (here’s an intro, using the models pioneered by Judea Pearl) don’t comment on subjunctive dependence, and you end up with CDT agents 2-boxing.
My entire understanding of FDT and how it varies from CDT or EDT is based on Eliezer and Nate’s paper that I linked to, so I will only comment on what I think they mean. If someone was complaining that “Omega favors irrationality” and expanded that statement to “The choice they reward favors irrational persons”, I would make the argument that rewarding people based on the choice/action they take is not an “unfair” scenario.
Some excerpts from FDT:
The standard defense of two-boxing is that Newcomb’s problem rewards irrationality. Indeed, it is always possible to construct dilemmas that reward bad decision theories. As an example, we can imagine a decision rule that says to pick the action that comes earliest in the alphabet (under some canonical encoding). In most dilemmas, this rule does poorly; but the rule fares well in scenarios where a reliable predictor rewards exactly the agents that follow this rule, and punishes everyone else. The causal decision theorist can argue that Newcomb’s problem is similarly constructed to reward EDT and FDT agents. [...] Newcomb’s predictor is not filling boxes according to how the agent arrives at a decision; she is only basing her action on a prediction of the decision itself. While it is appropriate to call dilemmas unfair when they directly reward or punish agents for their decision procedures, we deny that there is anything unfair about rewarding or punishing agents for predictions about their actions. It is one thing to argue that agents have no say in what decision procedure they implement, and quite another thing to argue that agents have no say in what action they output.
So that’s the “cannon” take on “Omega favors irrationality”.
On the second half of your comment:
I’d agree that you one-boxing would be evidence that Omega had given you the Million $. If you went through Newcomb’s problem using “Of the actions I could take, which one would provide the most Bayesian evidence that I was going to get lots of utility?” you would one-box. However, that process is not the decision process specified by CDT. We’re now talking about EDT (which as this comment points out, is sorta wonky).
If that doesn’t feel satisfactory, and it still feels like CDT would one-box, it might be useful to explicitly go through the exact decision process that you think CDT prescribes.
On “describing base on algorithim vs choice” I believe the point was being made that Omega/Nature does not go “Are you running algorithm number 1163527? If so, you get screwed.” You can have two algorithms that are functionally equivalent (same input output pairs) yet operationally different.
On “Why doesn’t CDT onebox?” I think it would be very helpful to imagine the CDT agent not as a person, but as a mechanical calculator. Its quite hard to faithfully simulate a specific decision algorithm in one’s head, because all of our decision making intuition is constantly trying to sneak in, and I’ve confused myself many times by “imagining an XYZ agent” but giving them more tools than XYZ prescribes. It’s not that CDT considers it more reasonable to use a causal model than to just pick the winning choice, it’s that a causal model is all a CDT agent has (plus a little bit of extra circuitry).
My guess at a possible confusion you might be having, is that when you are thinking of causal models, you are including things that are subjunctively dependent, when a causal model isn’t framed/designed to handle subjunctively dependent things. From the FDT paper:
Since we often automatically take subjunctive dependence into account, it feels like a causal model should take it into account. But the existing well defined causal models (here’s an intro, using the models pioneered by Judea Pearl) don’t comment on subjunctive dependence, and you end up with CDT agents 2-boxing.
My entire understanding of FDT and how it varies from CDT or EDT is based on Eliezer and Nate’s paper that I linked to, so I will only comment on what I think they mean. If someone was complaining that “Omega favors irrationality” and expanded that statement to “The choice they reward favors irrational persons”, I would make the argument that rewarding people based on the choice/action they take is not an “unfair” scenario.
Some excerpts from FDT:
So that’s the “cannon” take on “Omega favors irrationality”.
On the second half of your comment:
I’d agree that you one-boxing would be evidence that Omega had given you the Million $. If you went through Newcomb’s problem using “Of the actions I could take, which one would provide the most Bayesian evidence that I was going to get lots of utility?” you would one-box. However, that process is not the decision process specified by CDT. We’re now talking about EDT (which as this comment points out, is sorta wonky).
If that doesn’t feel satisfactory, and it still feels like CDT would one-box, it might be useful to explicitly go through the exact decision process that you think CDT prescribes.