- Location
- Book Fort
- Pronouns
- They/Them
Seconding this.
-
Going forward, links to content from Twitter (now X) will no longer be embedded. Pre-existing links should not be affected.
Marked for Death: A Rational Naruto Quest
- Thread starter eaglejarl
- Start date
-
- Tags
- naruto rationalist romance armageddon is looking better and better damn you kishimoto explosions everywhere explosives solve everyone if in doubt surrender to the mercy of zabuza it seemed like a good idea at the time looks like a really bad case of being a jōnin my god. where does the lore end? precommitting is the winning plan screaming in kagome screw the rules i have explosives something bad might still happen sometimes you gotta sociopath stabby zabby subsection c of paragraph 9 should include we can weaponize this so hard open voting
- Location
- Yes.
Not definite but what I could find.
My understanding of decision theory is not that extensive, but ... does this actually hold up? This is more-or-less a brinkmanship situation, if I understand it correctly; the payoff matrix is something like:
| Us, Oro | -Akatsuki Switch | +Akatsuki Switch |
| -Oro kills us | 0, -1 | -1, -2 |
| +Oro kills us | -100, 0 | -(100+x), -y |
Assuming that we indeed prefer Oro to beat Akatsuki even if he killed us before the assault (which I don't necessarily agree with) and that we're dealing with an Oro who would prefer to kill us before the assault (without which there's no reason to set up this switch.) Values are approximate, obviously, with the key assumption being that Oro doesn't care about killing us that much. x is how much we care about Akatsuki being more likely to resurrect Pain even if we're dead, y is how much Oro cares about Akatsuki being forewarned.
The complication is that x is probably not that large, especially not compared to y - we'd be dead anyway, while Oro would have to live with the consequences. Suppose y were 1000 and x were 20 - then would it make sense for us to set it up? It would still only have the purpose of intimidating Oro, but we'd be threatening him with something really bad and ourselves with something relatively mild; it seems reasonable to expect that Oro might yield first. Also, we have first-mover advantage, in that we've already locked ourselves into having the switch or not before we meet Oro and he gets the chance to kill us or not (though I suppose Oro could lock himself into killing us or not by precommitment).
As I've said, I'm not in favour of setting up this switch - I think the marginal benefit for us is tiny compared to just having the Cannai switch, and I think it might cause Oro to plan to kill us later even if he wasn't before - but it's not obvious to me that decision theory necessarily says Oro shouldn't yield.
Actually, insight - I think my key concern here is that Oro deciding to kill us regardless to disincentivize us from setting up a deadman's switch seems to me to only work if Oro expects us to be able to predict that he'd do that (à la Newcomb's paradox). I'm not at all convinced that we actually can predict Orochimaru's actions that well, especially in a scenario where Oro would be imposing such a massive cost on himself by killing us.
(Apologies if this has some major flaw that I just haven't seen - as I said, not all that familiar with decision theory.)
Hmmmmm @Left-Hand Mutant I am pretty convinced by this. Any counterarguement?
- Location
- The Soviet Republic of Canuckistan
- Pronouns
- He/Him
The below is well (~270) under 300 words - I'm not sure if the 399 is to leave room for expansion etc. but I think I've got the tone and content, plus a clarifying statement or two.
I will vote against plans which have us threaten Orochimaru to his face, unprompted. We are a lootbox and we are not dangerous enough to merit a sudden assassination. If he decides he wants us dead he's going to be interrogating us first.
That said, I think I would vote for plans which set up contingencies providing the remainder of the Akatsuki with Runecrafting capabilities if Orochimaru takes possession of the Rift, kills us, and does not return to Leaf. If Orochimaru moves against us, we can let him know and go from there. I think this would be (relatively) straightforward on the Cannai side: if we die, read anyone except the Akatsuki in. If they appear, demand proof that Orochimaru never returned to Leaf - the Slug Princess' word will do just fine. If they can get it, then they can have Runecrafting.
Last edited:
Noumero
Omit Needless Words
Oh yeah, that's another thing. We're almost certainly not dealing with such an Oro, so that part of the dead man's switch is unnecessary and only serves to antagonize Oro more.
Yep. Well, technically, he doesn't need to expect us to predict him specifically, he needs to expect us to model how a smart-enough agent would behave and then assume that he is a smart-enough agent. Not yielding to threats is a convergently good strategy, such that many agents would independently reinvent it; it doesn't rely on the peculiarities of Oro's mind specifically.
Or, more in the realm of how human beings work: it would also work if Oro has learned social heuristics where he becomes irrationally defiant if he perceives someone trying to pump him for resources using unreasonable threats. Having a reputation for a "bad temper" is one possible way to implement something like the above, but in a more compute-efficient manner (i. e., via deontology, not consequentialism). And the local maximum here would be at "fly off the handle if I'm being threatened with an irrational lose/lose setup, behave rationally if someone set up a rational win/lose situation".
Like, think about it as an evolutionary process. Oro's mind generates candidate policies like:
- (1) "if someone tries to blackmail me, I give in".
- (2) "if someone tries to blackmail me, I kill them even if the dead man's switch is already created, and then I just eat the consequences".
- (3) "if someone tries to create situations where they win even if I kill them, I negotiate with them".
- (4) "if someone tries to create situations where they win even if I kill them, I kill them anyway and then just eat the consequences".
So, even if no conscious calculation enters into it, he'd naturally gradient-descend towards using (2, 3) over (1, 4).
Last edited:
See I don't plan to even mention the dead man's switches unless Oro starts looking malicious. Hopefully this just literally never comes up. In the event it does, however, is it better to pessimize his utility functions or not. Keep in mind that he is still human and he does respond to threats.
If you've reached the "let's all die in a fire together" stage then I think there's some value to pessimizing your opponent's utility function. This isn't meant as a threat that you're negotiating over, this is a "if you defect I will act against you to the greatest extent possible"
Disclaimer: I don't actually know any decision theory lol
The problem I have with this is that I'm not convinced this strategy actually is what any smart-enough agent would use. I feel like there's an implicit assumption here that, on average, the reduction in number of threats from having a reputation of not yielding to them outweighs the cost of suffering the effect of the threats, and I'm not sure that holds here.
In many high-profile cases, this is probably true. Considering "don't negotiate with terrorists" on the part of a state; the actual cost of indivudual terror attacks to a state is fairly small, since terror attacks are generally orchestrated by much weaker actors, and the immediate cost of negotiating is often significant (large amounts of money/resources/etc.). Contrastingly, what they do in response is extremely visible, and if they negotiate little stops the same group from coming back for more later. So the cost of not yielding is comparatively low, the immediate cost of yielding is comparatively high, and the increase in threats as a result of yielding is comparatively high.
But in this case, none of those things hold. The immediate cost to Oro of yielding is probably fairly low - he can always kill us later, it's not like we plan to spread runecrafting to the entire world right after we beat Akatsuki. The cost of not yielding is extremely high - Akatsuki being forewarned of his attack and the method of it, dramatically hurting his chances of victory and therefore increasing the risk of Pain's resurrection, which he considers an existential threat. And the reputational effects, far from being visible on the global stage, are unlikely ever to be widely known - it seems unlikely that we'd want it publicized that we threatened to sell out Oro to the people everyone else considers to be insane super-terrorists. Even considering the effect on our own future interactions with Oro, we wouldn't be arbitrarily seeking him out to extract value from him, we'd be coming to a meeting he requested with a threat that would only trigger if he tried to backstab us; if he wants to avoid that happening he could just ... not try to backstab us.
All in all, it seems entirely plausible to me that for sufficiently serious threats for sufficiently minor immediate costs and sufficiently limited reputational effect, the correct thing to do is just to yield to such threats.
(To reiterate, I'm not in favour of this idea. I think Oro is very very unlikely to try to kill us immediately, if he does the Akatsuki switch adds very little to the Cannai switch, and it might cause Oro to kill us later after the threat is expired and before we can set up another one. But if we were certain Oro wanted to kill us immediately (and for some reason we went to meet him anyway) and had no other options, I might then consider it a sensible option.)
I notice that your theory of "optimal game theory" would immediately result in total nuclear war if even one nuclear powers subscribed to it.
This should be a hint that your ideal isn't actually optimal, nor is it a natural position that people are intuitively drawn to.
The both correct and intuitive play is much closer to "I'll respect threats that are reasonable* and/or contribute to good social norms, but will stop respecting threats if you try to bully me with them."
"If you nuke me I'll nuke you back" or "If you kill me I'll take you out with me" are both intuitively reasonable* and pro-social - they make nuking/murdering your neighbours unprofitable! These retaliatory policies result in a much better equilibrium for everyone, including the existing nuclear powers!
Both culture and individual strategy tend to converge towards this.
I'm travelling and can't spend the time to fully engage in this (sorry). Offhand, this might be easy+informative related reading.
*"reasonable" can roughly be factored out as "proportional to the ammount your or your society's values were violated, and legible/guessable that you'd react to".
People are tend to care a lot about protecting their life and their child's life, and It's kinda obvious that threatening those would provoke retaliation. Retaliating if these are threatened is thus "reasonable", perticularly in a world with out police.
Demanding your co-workers new car for free isn't reasonable. That's not one of your terminal values(or if it is, you're so insane/weird you can't be safely interacted with anyways), and nobody would intuitively guess that not turning over their car is grounds for retaliation in your eyes. This behaviour is crazy, and should be ignored, even if you get punished for it.
People are tend to care a lot about protecting their life and their child's life, and It's kinda obvious that threatening those would provoke retaliation. Retaliating if these are threatened is thus "reasonable", perticularly in a world with out police.
Demanding your co-workers new car for free isn't reasonable. That's not one of your terminal values(or if it is, you're so insane/weird you can't be safely interacted with anyways), and nobody would intuitively guess that not turning over their car is grounds for retaliation in your eyes. This behaviour is crazy, and should be ignored, even if you get punished for it.
It's actually less extreme but pretty close to a traditional deadman switch of "If you murder me, we both explode".
Which is famously a good way to deter someone for murdering you, as long as you otherwise behaved well.
Just ask Mari/Kei about it, don't do it unless they're all in. and if we do set it we won't even tell Oro unless things get dangerous/threatening.
Plus, Kei can just lie and tell Hazou she used it in the deadman. Best of both worlds.
Last edited:
Noumero
Omit Needless Words
Here's a bunch of scenarios:
- Suppose you live across the street from a coworker, who is about to kill you because he wants your job. If you had the opportunity, would you choose to have an asteroid fall on your neighbourhood later that day, killing your coworker, but also your entire family and any bystanders?
- Suppose you are a leader of a nation A with nuclear weapons, waging a conventional (so far) war with another nuclear nation B. You value human life and don't want to kill B's civilians. B's ideology, however, is hellish by your values; you'd prefer to destroy the world to B winning. You are losing the war. If you fire nukes, the other nation will fire as well. Here's a twist, however: the amount of nuclear weapons at your disposals is insufficient to totally destroy each others' armies, and you're confident that your nation is more well-positioned to win in a world after the nuclear exchange. The exchange will, however, cause countless civilian casualties. Do you fire the nukes?
- Suppose the situation is as above, but you're B instead. You don't want to kill A's civilians; A's victory is worse than extinction; you're winning; you see them launch the nukes. You expect to lose in the world after the nuclear exchange, with, say, probability 80%. But if you don't launch the nukes, you will lose with 99.9% probability. Do you launch the nukes?
If you're at the "let's all die in a fire together" stage, that doesn't actually mean deliberately optimizing for everyone to lose, like killing everyone you know just to spite your coworker. Even at that stage, you're still optimizing for your values. It's just that sometimes there are situations in which your enemy commits to a course of actions such that the only way to preserve as much value as possible is to play your part in tanking the world's value by everyone's standards. Both A and B would be poorer in a post-exchange world, but A would lose totally unless they nuke, and B would lose totally if they don't nuke back. So both nuke, and create a world-state that's massively worse by both standards, yet without having ever swerved away from expected-utility maximization.
So in the counterfactual where Oro kills us, we should only give the notes to the Akatsuki if we expect that to result in a better-by-our-values world. Not because we want to make a worse-by-Oro's-values world regardless of how bad it becomes by our values.
... No? Also it's not mine.
Here's a different framework for thinking about it: commitment races.
(Note that these are not functional-decision-theory consequentialists.)
So in this case, the question is whether Oro is the kind of person who'd already precommitted to ignore all game-theoretic threats, regardless of their scale.
Note that this framework is much, much messier and more arbitrary than FDT, likely leads to countless misunderstandings regarding what sorts of things any given consequentialist did or didn't commit to (since the structure of these commitments now does depend on the details of each individual agent's mind and history, rather than FDT's much cleaner and detail-agnostic structure), and so everyone sets off each other's dead man's switches all the time. So sufficiently smart agents probably do converge to FDT instead.
Last edited:
That's the kind of position notorious for failing horribly unless almost everyone else in your environment is a very perticular kind of nerd, and you make sure to make extremely public declarations/examples of your policy.
Orochimaru has not grown in an environment that would reward this position, or make it look like a good idea.
If Mari hasn't heard of Oro holding this policy, then he hasn't tried to make an example of it, which is the first thing you'd do if you had such a policy.
I don't necessarily feel like this is reflective of the situation since it's an outcome maximally pessimistic of my values, but no worse by my opponent's. Since he's dead either way.
See, IMO the fact of Orochimaru's betrayal changes the calculus on desirable world states. If we knew, for a fact, that he would betray us, I might tip Akatsuki off so they could kill him, or die trying, and we could try our luck on the victor afterwards.
Tipping Akatsuki off is not to give them victory, but to hurt Orochimaru. Which becomes a priority of mine if he betrays us. I'm not saying that we should actually do this or what my exact calculus is. But I think it's worth considering that betrayal can shift desired world states.
It's the official policy of the United States government. "We do not negotiate with terrorists".
I'm not saying that if you're a hostage or one of your loved ones is it's what you want to hear, but that's official state policy.
- Location
- Book Fort
- Pronouns
- They/Them
But what about Project Omnicule?
If everyone is dead, then there's nobody who's not part of the omnicule. A resounding success.But what about Project Omnicule?
Nope!
The US, like everybody else, takes reasonable* threats into consideration, but ignores bullying/terrorism.
The US has been extremely carefully not to directly attack Russia, for example.
The US is another great example of what I described as "the commonsense policy everyone tends to converge on"
If the US did totally ignore all threats, we'd have a NATO-Warsaw nuclear war and most people would be dead. Their actual policy is to have clear kinds of "reasonable" threats they allow and benefit from, while ignoring bullying/exploitation.
Last edited:
- Location
- Book Fort
- Pronouns
- They/Them
Now we're cooking >:3
Ehhh probably the US actually veers too far on the side of avoiding "escalation".
I was considering the situation where the opponent can't meaningfully hurt them, but that's not really applicable here.
When nuclear war is on the table they ignore (some) things that are clearly bluster, and they avoid things that would probably escalate to nuclear war. No-Fly Zone? Absolutely 100% would cause nuclear war.
US pilots on "vacation" in Ukraine? Almost certainly wouldn't cause nuclear war, but the US doesn't do it.
- Location
- France
I fall for it every time. Alright, adding more CR levels to the wishlist, right after pushing ES to 60. I'm all in on the eaglejarl suggestion train !
We're nowhere near getting another ACE pip, right ?
- Location
- Book Fort
- Pronouns
- They/Them
Either 3 days or 3 months, pending QM Discussion.
Situation unclear, most optimistic estimate 3 days away, less optimistic 15 days, realistic/pessimistic multiple months away.
Yup, as you point out, the actual situation is a bit more nuanced, mostly due to parties often having an incentive to be a slightly vague on where exactly they draw the line. Nobody knows that "vacation" pilots would trigger war, but they also don't know it won't.
I do think it's still a central example of the kind of decision theory I've been describing.
That's the main reason why I keep saying that optimal decision theory "basically" or "very nearly" converges to [the simple dynamic I've outlined]. The ambiguity means there's a bit more to explore.
Ambiguity adds a lot of very interesting nuance to the edges of this dynamic, but is enormously to discuss, and still leaves the main dynamic intact.
It's very not obvious to me that the US is overcautious by not sending in pilots and aircraft on "vacation"(~2% chance this triggers war), but I could see reasonable people disagreeing.
We probably shouldn't get to far into this example in thread.
Last edited:
[X] Snek Moot.
Word Count <399.
Word Count <399.
- Sanity check plans for the meeting.
- Make stops on the way to Rice for Mari to gather information.
- Focus on information we would want more than Orochimaru's word to believe.
- Cancel the meeting and contact Leaf if Mari feels something fishy.
- Make stops on the way to Rice for Mari to gather information.
- Go to the meeting with Mari.
- Tone. Not transactional or one upping. We are working together for the best outcomes for everyone, not positioning ourselves to be the ones who benefit the most in the aftermath.
- Orochimaru's plan and progress.
- Tell us everything we need to know. Specific interests.
- What does he know about the fortress, Akatsuki's plans and member locations.
- What allies do we have for the fight?
- Rice?
- Leaf?
- What runes and other assets relevant to the fight should we know about?
- Share our combat relevant runes.
- Air Dome. Force Dome. Iron Earth. RERs. Ninja Radar. Air Leadener. Icarus. Storm.
- We need his help with substrate.
- Does he think we can get runes close to the rift or does he want to try to lead them to prepared ground?
- Do we attack before they open the rift or while some of them are inside?
- Offer to swap abandoned rune prep results after the fight to save mutual research time.
- Share what we learned about the Great Seal.
- Share our combat relevant runes.
- Objectives. With the tools we have what are we aiming for?
- Killing Akatsuki?
- Killing their seal masters?
- Stealing the rift?
- Do we need more information?
- Pain did not die normally. He sacrificed himself. He could be truly gone or difficult to evac if the thing Orochimaru called a "market place" brought him somewhere defended.
- Waiting until they open the portal and Hidan checks with Jashin could tell us more.
- Killing Akatsuki?
- Tell us everything we need to know. Specific interests.
- Getting ahead of conflicts. Promise we will not deny Orochimaru the rift.
- We say whatever we need to to Leaf to get their help to stop Pain.
- Whoever gets the rift, if we have access we promise to research runes to open a second rift to the Pure Lands.
- If Leaf gets it, we will hide making a second one for him.
- If he gets it, we will make the second one for Leaf to make them think they have the only one.