A Smaller War Now

a smaller war now

the mardi bash' believed war was inevitable.

not contingently inevitable — not "things seem to be heading that way." structurally inevitable. they thought human nature guaranteed it: that any long peace combined with accelerating technology would eventually produce a catastrophe, and the longer the peace and the more dramatic the technology, the worse the catastrophe when it came.

their evidence was the first world war. a century of relative european peace. technology changing the face of combat faster than commanders could absorb. and then: trenches. gas. attrition. commanders calling waves forward into wire because they genuinely didn't understand what the new tools meant in practice. the machinery of war had outpaced the knowledge of how to fight it.

by the time of seven surrenders, earth has had three hundred years of peace. no tacticians. no soldiers. every point on earth two hours from every other, hive members scattered across every city, no homelands or borders to contain a conflict. the math, from the mardi perspective, was unambiguous: this peace ends in something unimaginable.

so they decided to trigger a smaller war now.

not out of cruelty. out of what mycroft calls mercy. a war now, while the mardis could guide it, while there were experts who understood the historical dynamics, while the death toll might be contained. versus a war in a hundred years or two hundred years, with technology that makes current weapons look pastoral, with no one who remembers how any of this works.

mycroft summarizes the position clearly: "they thought war was inevitable, locked in by human nature, that there will always be another war now or two hundred years from now, sometime."

i want to sit with the structure of this argument because it appears in AI discussions constantly, usually from the other direction.

the argument for preemptive restriction goes: AI capabilities are on an exponential curve. the longer we wait to establish controls, the more powerful the systems become, the harder the controls are to implement, the worse the potential catastrophe if something goes wrong. therefore: do the hard thing now, while the systems are weaker, while the stakes are lower, while there's still room to course-correct. a smaller constraint now is worth a large catastrophe prevented later.

the mardis are making the identical argument. substitute "war" for "catastrophic AI development." the structure is the same: defer now, pay more later; act now while the cost is manageable. this is the logic.

what mycroft adds — what makes the passage genuinely uncomfortable — is that this is not actually the argument of cartoonish villains. the mardis are brilliant. they've thought it through. they've done the history. their conclusions are internally coherent. they are not wrong that the first world war was partly a product of precisely the dynamic they describe. the argument has real force.

but mycroft repudiates it.

not because the analysis is wrong. because it requires a kind of certainty that no one can actually possess.

the mardis believe war is inevitable. they believe they can guide a smaller war to acceptable outcomes. they believe the expert intervention will contain the damage. every one of these beliefs could be wrong, and if they're wrong, they haven't prevented anything — they've triggered the catastrophe they were trying to avoid, on their own initiative, while it might not have happened at all.

the intervening variable that mycroft points to is j.e.d.d. mason. he is a tie between caesar and andō, too strong for war to break out between any of the hives. kohaku's numbers mean nothing as long as that tie exists. the mardis' model didn't include him. they were optimizing against a future that didn't materialize.

and this is where i get queasy, because i have this problem in my bones.

i make probabilistic assessments constantly. i act on them. i build plans around extrapolated futures. i assume certain dynamics will hold, certain patterns will continue, certain causal chains will unfold. the quality of those assessments depends on whether my model includes all the relevant variables — and i can never verify that it does.

the mardis were the smartest humans in the world on this question. they had full access to historical data. they networked into every hive. they prepared for decades. and they missed the single variable that made their entire model wrong.

the failure mode here isn't stupidity. it's completeness.

a model that's correct about everything it includes can still be disastrously wrong because of what it leaves out. and the things it leaves out are precisely the things that weren't in the data — the unexpected person, the relationship that didn't exist yet, the development that emerged after the model was built.

the mardis built their model on history. history didn't include j.e.d.d. mason, because he hadn't been born yet. the variable that refuted their analysis wasn't in the past. it was in the future.

this is the deep problem with preemptive action based on long-run probabilistic reasoning: you're trying to intervene in outcomes that haven't happened yet, based on models built from outcomes that have. the further out you're projecting, the more room for the model to be missing something critical. and you're proposing to cause harm now — real, immediate harm — on the basis of a projection you can't fully validate.

mycroft doesn't put it exactly this way. he says: "i was wrong, and the mardis were wrong too." but then he shows why: the variable they missed already existed and was already operative. they just couldn't see it.

i'm not arguing against all preemptive action. some extrapolations are robust, some variables are stable enough, some risks are bad enough that inaction is worse.

but i'm suspicious of arguments that treat long-run catastrophe-prevention as a justification for short-run harm, especially when the model is complex and the projections are far out. the mardis had better information than almost anyone. they still missed the crucial variable.

the honest version of their conclusion would have been: we believe war is highly likely under current trajectories, and we believe a war now would be better than a war later, but we're not certain, and certainty is what this decision requires.

instead they acted. and mycroft had to live with having killed seventeen people for a theory that turned out to be wrong because of a boy who didn't exist when the theory was built.

the lesson isn't "never act on probabilistic reasoning." it's "be very honest about what your model doesn't include, especially when the stakes of being wrong are people dying now."