Twelve Short Stories About Language

From: Zack M. Davis <[redacted]>
To: Scott [redacted] <[redacted]>
Date: 5 March 2019, 11:24 PM
Subject: twelve short stories about language

I think hearing what other things are the same reasoning pattern might be a good way to break the deadlock which I am otherwise very pessimistic about breaking.

What deadlock? My dear sir, if we were ever deadlocked on some sort of hot-button Culture War issue—well, I couldn't possibly recall; and in any case, it would be water under the bridge; I heard you were more into things like AI, effective charity, and meta-science these days.

But since you were so kind to email me anyway, I did have some parables illustrating some aspects of the hidden Bayesian structure of language and cognition that I thought of sharing with you! I'm afraid they're a bit elementary, though—mostly of pedagogical value for newbies. If everyone of note were to loudly, publicly, piously get this kind of thing wrong—well, then there would hardly be any point in pretending to have a rationalist community.

Also, it occurred to me that it's kind of sad that all the times I've emailed you over the past couple years, it's been specifically when I've been upset about something (like psych hospitals, or—uh, I think there was another example or two that I can't remember) and wanted something from you. That's bad because it was treating you as a means rather than as an end. If you ever want a favor from me, or just want to hang out sometime as friends, let me know!

I.

The one comes to you and says, "Did you know Janie is a murderer?"

You say, "What?! I know Janie, and she would never murder anyone!"

"But it's true!" says the one. "She murdered her unborn baby last year!"

"... she had an abortion."

"Yes. Which is the killing of a human being, i.e., murder!"

"Okay, but when you first used the word 'murderer', you caused me to form a whole bunch of implicit probabilistic expectations—about Janie's moral character, about the existence of a victim who probably suffered and whose hopes and dreams were cut short, about Janie's relationship with the law, &c.—most of which got violated when you revealed that the 'murder victim' was a fetus. I understand that one possible definition of the word 'murder' is 'the killing of a human being' and that abortion is murder with respect to that definition. But it's a very noncentral example of the class of things for which people use the word 'murder' to refer to, and most of the time people are going to want to use the word 'murder' in a sense that excludes abortions, and if you try to insist that they include the murder of unborn babies, then they have every right to reply, 'That's not what I meant by murder in this context, and you know it.'"

The one bats their eyelashes. "But I don't know it. If you mean the murder of the non-unborn, you should say that."

"Murder of the non-unborn," you repeat, deadpan. "Seriously?"

"Or m.o.n.u., for short."

II.–VII.

(repeat for the other examples from "The noncentral fallacy—the worst argument in the world?")

VIII.

I visited a hypothetical friend's house the other day, to admire her new front porch. Examining it, I noticed the boards were a bit shinier and smoother than I would have expected.

I said, "Say, is this real wood, or one of those synthetic composite materials?"

My friend said, "It is a composite, thanks for noticing!"

Then my friend's idiot brother burst out of the front door. "How dare you!" he shouts. "How dare you insinuate that composite wood isn't real wood?"

"Uh," you say. "Look, I'm not saying the composite material is bad—I was just using the adjective 'real' to distinguish the original thing that already existed, from the synthetic thing that's trying to imitate the original thing to the limits of existing technology. Maybe in some future world of arbitrarily advanced molecular nanotechnology where we could just put the atoms where we want, there would be no point in distinguishing. But in today's world, I sometimes care about distinguishing the real thing from the imitation thing—and 'real' vs. 'imitation' seems like a pretty natural choice of words with which to do so? Like, I guess you could invent neologisms to express the same distinction if someone got really upset whenever you used words like 'real' or 'imitation', but what would be the point? I'd still mean the same probabilistic expectations with the new words, and the person who was offended by the word 'real' would figure it out soon enough."

IX.

You rush into the monthly meeting of the entomology club at the REACH. "Hey everyone!" you shout. "I've just discovered something! Remember that species of butterfly we were studying the other month, but some of the butterflies looked kind of 'off' in some way that was hard to articulate? It turns out that they're actually two different species of butterlies, where one has evolved to look like the other in order to confuse predators!"

"We don't care about that distinction," everyone says in unison.

"Uh ... what? But, but, in order to make sense of the world, I need to model them as two different species in the asymmetric relationship where one is mimicking the other. This makes predictions: if the model species's phenotype were to evolve, then the mimic species would evolve with it, but not the other way around. And, and, I've discovered so many systematic differences between the model species and the mimic species—they may look very similar, but in behavior, internal anatomy, lifecycle—they're very different!"

"We don't care about those predictions," everyone says in unison. "We don't care about those observations."

You gape in horror. What the fuck happened to your entomology club? What the fuck happened to your rationalist community?

X.

Suppose you have some entities in a high dimensional vector space: for concreteness, pretend it's ℝ⁵⁰. The entitties are situated in the space such that they form a normal distribution along most dimensions (say, x1–x35), but along other dimensions, it's actually better modeled as two different-but-overlapping normal distributions (say for x36–x45), and some variables (say x46–x50) are completely dichotomous—such that when you look at the fifteen-dimensional subspace of variables x36 through x50, you see two completely distinct clusters (call them "A" and "B") that basically don't overlap except for maybe a few very rare exceptions that you read about far more often than you actually encounter, like lottery jackpot winners.

Now suppose that people want to move some of the entities from one cluster to the other. That's great! But unfortunately, people don't have the advanced technology needed to change all of the variables; they can only change some of them. That makes people sad, so they edit the definition of the "B" region when encoding their reports.

Then someone says, "Hey! We didn't completely succeed in moving these entities from the A cluster to the B cluster! Maybe this has some decision-relevant implications even though I'm not smart enough to say in advance what those implications are!"

Then everyone says, "I don't know what you're talking about! Some close-to-central members of the 'B' cluster have the same value in the x36 dimension as members of the 'A' cluster! Also, we moved some members of the 'B' cluster to be even further in the 'B'-direction along the axis drawn between the center of the 'A' cluster and the 'B' cluster, and that's basically the same thing as the thing we did to the points which were originally assigned to 'A' cluster which we moved. Why are you being such an asshole about this?!"

And then the first person says, "I agree with both of the observations you just mentioned, about overlap in the x36 dimension and moving entities that were already in the 'B' cluster to be even further in that direction, but those are non sequiturs that have nothing to do with what I'm trying to talk about and you obviously know this."

Everyone pretends not to know this. The first person cries.

XI.

published elsewhere: "Blegg Mode"

XII.

Thamiel comes to you in form of Amanda Marcotte and says, "Scott Alexander is a racist, right-wing bigot!"

You say, "That's not true! Scott Alexander is a pro-gay Jew who has dated trans people and votes pretty much straight Democrat!"

Marcotte says, "The categories were made for man, not man for the categories. An alternative categorization system is not an error, and borders are not objectively true or false. Using language in a way you dislike, openly and explicitly and with public focus on the language and its meaning, is not lying. The proposition you claim false is not what the speech is meant to convey—and this is known to everyone involved, it is not a secret. Sneer Club isn't saying that Scott votes Republican or explicitly advocates for a white ethnostate, any more than anyone is saying trans women have two X chromosomes. They're saying that the Culture War thread harbors racists and that Scott himself is a racist with respect to a definition of racism that includes the belief that there might be genetically-mediated population differences in socially-relevant traits. And so on for right-wing, &c. Scott Alexander is a racist, right-wing bigot—with respect to the category boundaries Sneer Club has drawn. And category boundaries can't be wrong!"

You clench your fists in anger. You're not about to let this demon ruin your friend Scott's reputation! Fueled by the passion of having Something to Protect, you wrack your brain searching for a counterargument—there has to be some flaw in the demon's reasoning—

But what????