Containment Thread on the Motivation and Political Context for My Philosophy of Language Agenda

motion of circularly defining a category in terms of things-that-are-good-to-be-in-this-category

Someone could say "I am a woman", in context with the goal of communicating something like "please treat me like you treat women, not like you treat men", and I called that "using 'women' to mean, the cluster....". But that doesn't imply they aren't also referring to some underlying reality. Maybe a developmental quirk made their brain develop with a female regulatory mode. Maybe not, but as a child they ended up in a feminine psychic basin of attraction. Maybe they are harmed by their social position because of their gender and want out. In any case, they want to be treated as women. I don't see what's wrong with giving a name to the cluster of people who want to be treated the way women are treated.

That doesn't mean anyone has to do anything; just because you think me treating you some way would be good for you, doesn't mean I believe you, and doesn't mean it would be good for me if I treated you that way.

would use the corresponding language because it was straightforwardly true, not because they were trying to be nice.

Those aren't the only two options. You could use language because it's unstraightforwardly true, and use appropriate precautions. I wonder if you're pushing against calling such a person "woman" because that makes it seem like how you use words like "sky" or "hamster" or "seven" or "tasty" or "beautiful", which is a native representation that you can use unreflectively in your internal planning, whereas really it's a cluster of other peoples' preferences, and if you use them unreflectively you're "unsandboxing" other peoples' preferences into your internal planning.

Maybe we could separate your objections into two parts: you object to calling this cluster "women", and you object to reifying this cluster at all (?).

how many of the people in modern Society who say "Gender is a social category" actually want that?

If we unstrawman from "work in the household and take care of children" to the whole panoply of social attitudes, then maybe a lot of people actually want that? I don't know, it would be interesting to know how much trans people care about treatment vs. some more transcendent type of being believed that they're really women/men, or vs. other stuff; and if someone cares about other people thinking that they're really a women, what exactly does that mean to them etc. and how would that reconcile with biological or psychological facts (noting that some psychological facts are objects of free will, e.g. one could decide to shape ones mind like a woman's).

If your Society is structured such that the most salient things about "women" is that they're the ones who work in the household and take care of children—then sure, in the context of that Society, males who work in the household and take care of children are women.

In Britain in the 1950s and 1960s (and long before, but I'm limiting this to what I observed in my own lifetime), that was the position of women. Men who took on that role were not "women". They were thought very unusual, even unnatural, and the word "househusband" was coined.

That makes sense as a critique of my or Bailey's writing, but "Davis and Bailey's writing is unclear and arguably deceptive given their target audience's knowledge" is a very different claim than "autogynephilia is not a natural abstraction"!!

More specifically, I believe that the reason your and Bailey's writing is unclear given the target audience's knowledge is because the audience lacks the knowledge needed for autogynephilia to be a natural abstraction. And my claim "autogynephilia is not a natural abstraction" is not primarily critiquing you or Bailey, but rather just putting the background of the statement out there separately from the critique.

Let's try to formalize it. I think the state of the art formalization of Natural Abstractions is Natural Latents: The Math [LW · GW], though from what I understand it doesn't make any claim to be the final answer for natural abstractions? Would like to know what @johnswentworth [LW · GW] thinks.

The key diagrams to consider are I think these:

Let's start with the sense in which autogynephilia is just-sort-of-barely a natural abstraction for me:

$\Lambda$ refers to my understanding of how autogynephilia works, and is essentially a factor model which is not literally the same as my statistical models (since statistical data needs to be taken with a grain of salt, and because I also have additional grounding knowledge not included in the statistical model), but which is conceptually very similar to my statistical models like this one:

That is, for me, $\Lambda$ consists of:

• A distribution of levels of autogynephilia I expect to see in the general population,
• The relationship between levels of autogynephilia and the material things it is uniquely associated with, which is mainly autogynephilic sexual fantasy life,

Each $i$ would then refer to a person (maybe only natal males, to resolve some ambiguities), and each $X_i$ is then the projection of the unique associates, which is to say, the autogynephilic aspect of his sexual fantasy life.

Autogynephilia then satisfies the first diagram because whenever I notice some dependency in autogynephilic sexuality, I either incorporate it in $\Lambda$, or factor it out as a distinct trait from autogynephilia, purging it from $X_i$, or do a mixture where I distinguish one component of it as autogynephilic and another component as non-autogynephilic (with the choice depending on what is most theoretically relevant, though usually I opt for the latter). For instance with forced feminization, I pick the latter option, saying that one should abstract over force when it comes to placing feminization within $X_i$ (and abstract over feminization when it comes to placing force within $Y_i$, the masochistic projection of person $i$'s sexuality).

And this is why I say "just-sort-of-barely". Let's say I didn't go out of my way to abstract the $X_i$s, for instance if I included $m$ different transgender-related traits into the $X_i$s. To a good approximation, I would expect these transgender-related traits to apply to far more trans autogynephiles than to cis autogynephiles. This then leaves me two options:

• I can update $\Lambda$ to add transness as an additional trait to preserve the natural abstraction element, in which case the natural abstraction $\Lambda$ is no longer autogynephilia, but instead autogynephilia-and-transness. (And so autogynephilia would not be a natural abstraction wrt. those $X_i$s.)
• I can add the relationship between autogynephilia and each of the transgender-related traits individually, without modelling their conditional dependencies, but this would increase ${\epsilon }_1$ a lot (approximately proportionally to the number of transgender-related traits I add, I think?), and also "contaminate" the estimates of trait autogynephilia for each $i$, as suddenly transness (which is also caused by gender progressivism etc.) is considered evidence for much higher autogynephilia. But note that this would be an unstable situation, as e.g. @johnswentworth [LW · GW] says to resample $\Lambda$ conditional on $X$, and doing so would generate a new level of autogynephilia above the previously studied ones, where trans women would then be stuffed in.

In neither case, autogynephilia-as-I-consider-it would be preserved. So basically I can't have a rich theory of autogynephilia unless I combine the natural abstraction with other stuff. (I haven't worked with gasses for a while, but I think they are similar? Like there's a reason John talks about (P, V, T) instead of just P as an example of a natural abstraction.)

Now, not only do most people not go out of their way to abstract the $X_i$s, instead they can't do it because their data is too partial so they mostly don't observe the indicators that I would consider "pure" measures of autogynephilia. But furthermore they don't want to, because abstracting the $X_i$s makes it less interesting; most of the meat of the theory is in them.

(The above isn't the only way to parse autogynephilia as a natural abstraction; rather than letting the $i$s represent different people, they could represent different aspects of each person. That would actually be more native to how I usually do it as it is more closely related to psychometric factor models, but from my reading of the topic, the way I parsed it above better get at the core of it? Idk. I think my argument goes through either way you parse it.)

Permalink or it didn't happen: what's your textual evidence that I was doing this?

I think these things you described in Sexual Dimorphism in Yudkowsky's Sequences, in Relation to My Gender Problems are basically independent of AGP (even when aggregated), with gender progressivism (or maybe the "eccentricity" factor I found? idk, probably a mixture) being a more relevant factor:

Or when I would practice swirling the descenders on all the lowercase letters that had descenders (g, j, p, y, z) because I thought it made my handwriting look more feminine.

Or the time when track and field practice split up into boys and girls, and I ironically muttered under my breath, "Why did I even join this team?—boys, I mean."

Or when it was time to order sheets to fit on the dorm beds at the University in Santa Cruz, and I deliberately picked out the pink-with-flowers design on principle.

Or how I was proud to be the kind of guy who bought Julia Serano's Whipping Girl: A Transsexual Woman on Sexism and the Scapegoating of Femininity when it was new in 2007, and who would rather read from Evelyn Fox Keller's Reflections on Gender and Science than watch Super Bowl XLII.

Or how, at University, I tried to go by my first-and-middle-initials because I wanted a gender-neutral byline, and I wanted what people called me in real life to be the same as my byline—even if, obviously, I didn't expect people to not-notice which sex I am in real life because that would be crazy.

...

Or how I stopped getting haircuts and grew my beautiful–beautiful ponytail. (This turned out to be a great idea and I wish I had thought of it sooner.)

...

Or the time I felt proud when my Normal American Girl coworker at the supermarket in 'aught-nine said that she had assumed I was gay. (I'm not, but the fact that Normal American Girl thought so meant that I was successfully unmasculine.)

...

So, you know, as part of my antisexism, I read a lot about feminism. I remember checking out The Feminine Mystique and Susan Faludi's Backlash from the school library. Before I found my internet-home on Overcoming Bias, I would read the big feminist blogs—Pandagon, Feministe, Feministing. The one time I special-ordered a book at the physical Barnes & Noble before I turned 18 and got my own credit card and could order books online, it was Feminist Interpretations of Ayn Rand.

Also just generally not pulling your hair out when reading Phil's book, etc..

(I do expect there to be a relationship of some strength in the AGP→progressive direction, but my 2017–8 models were not in any way surprised by, e.g., the "Conservative Men in Conservative Dresses" profiled in The Atlantic in 2005, or the second kind of mukhannath, or Kalonymus ben Kalonymus.)

I have to admit I would find your argument a lot more compelling if you weren't inferring their autogynephilia on the basis of such distant information (since it's the inference on the basis of distance information that I argue to be the cause of such issues).

Idk, maybe everything is too vague perceive this accurately. Or rather, maybe I should perceive this stuff as a generic pro-Blanchardian bias, like interpreting your antisexism in the light of autogynephilia because that seems Blanchardian to you, interpreting your endorsement of Phil's book because that seems Blanchardian to you, etc..

It's inherently going to be hard for me to come with any confident theory about why you do what you do. I mainly just mentioned it because I'm often not convinced that anybody except me makes these distinctions, so I don't want to commit to "oh there's this Good Cabal that Gets It". (E.g. you seemed to find the whole AGP/masochism distinction kind of mind-bending rather than just being one out of several distinctions that one needs to go through to achieve literacy in this area)

The examples seem relevant to me because they illustrate that language is not used only to predict, that the merits of a particular language-using strategy are not determined only by its impact on predictive accuracy. If language in general has proper goals other than predictive accuracy, why should I think that category-boundary drawing has no proper goal other than predictive accuracy?

I'm not sure exactly what distinction you're appealing to, by the way. In particular, you say "the communicative function of proper names ... the cognitive function of categories" and it's not clear to me whether (1) you're suggesting that proper names are used primarily for communication while categories are used primarily for cognition, or (2) you're saying that your complaints about talk of arbitrariness apply only when thinking about cognition as opposed to communication, or (3) something entirely different. I say that proper names and category boundaries are highly relevant to both communication and cognition, and that some of the examples that apparently bother you most seem much more about communication than about cognition (e.g., being asked to use pronouns that you find inappropriate for the people they refer to, which you say amounts to asking you to lie, clearly a communicative category more than a cognitive one). It's entirely possible that your meaning is something entirely different, of course; please feel free to enlighten me if so.

The pouncing algorithm can't literally be "pounce when people say nice things about TCWMFM" because in the particular case that sparked this particular discussion that didn't happen but you still pounced.

It probably won't surprise you that I don't agree with your description of TCWMFM as pants-on-fire mendacious, and I don't think the edit at the end undermines the foregoing material nearly as much as you say it does. There are rules, but the rules are of the form "if you want your thinking to be optimal in such-and-such a way, you have to think like this" and Scott is talking about words: how to speak, not how to think. Words and thoughts are closely related but they aren't the same, and I agree with Scott that neither rationality nor rationalism includes any obligation to optimize your words for the exact same thing as your thoughts.

For example, suppose that (1) someone decides to use the word "woman" to mean, as precisely as they are able, "adult human being who identifies as female", and that also (2) you are correct in holding that the region of person-space picked out by this definition is less useful for making predictions than one like "person whose gender-related characteristics are, collectively, more like those of the average woman than like those of the average man". (That definition is kinda-circular but in a way that does no actual harm, and it leaves open the question of what "more like" means, but let's suppose that (3) no plausible way of making "more like" concrete gives you anything close to "adult human who identifies as female", though I'm not sure that's actually the case.) Is our hypothetical person doing something rationalists should be ashamed to do?

I claim they need not be. They are (given all the stipulations in the previous paragraph) using language in a way that isn't optimized for prediction, but I see no reason why language-use should be optimized purely for prediction, because the actual uses of language are broader than that.

You might object that they are impairing their ability to think clearly: when they say or hear that someone is a woman, that will lead to certain expectations in their mind, and if they define "woman" in a gerrymandered way then those expectations will be wronger-in-expectation than they need be. It seems to me that any such effect is very small; someone using the word "woman" this way knows that they are doing so, knows that when someone else says "X is a woman" they don't necessarily mean it the same way, can when it seems appropriate think explicitly in terms of Y-chromosomes or hormones or whatever other characteristics might be relevant. The last point is important. When you have some category without a standard precise technical definition, and it's actually important to reason about its members, it usually turns out that what really matters isn't membership in the category as such but something more specific. Is So-and-So "really" a woman? In cases where there's any scope for doubt about that, you will get more useful answers by figuring out what the real question is (do I want to have sex with her? should I let her into my javelin-throwing competition? which public toilet should she be using? what dose of this drug should I give her? is it possible that she's pregnant? etc., etc., etc.) and answering that more directly. So, anyway, I think any thought-impairment here is very small, and reiterate that language has purposes other than facilitating clear thought; if our hypothetical person thinks there are other benefits that outweigh any such loss, it is far from obvious that they're wrong.

You might object that if they use language that way they are failing to communicate maximally informatively: if someone who uses "woman" that way tells you someone is a woman, the information they're giving you is less useful for your prediction-making than if they used the word in a prediction-optimized way. Well, maybe, kinda, but for at least two reasons I don't think this is something it's reasonable to get upset about. 1. Regardless of what they do, the word "woman" is going to have ambiguities around the edges unless the person using that word tells you explicitly and in detail where they draw the boundaries. (The same is true, in less politically-sensitive ways, of many other words. If I call something a mountain or a town, or say someone is stupid or beautiful, or say that a wall is painted green, you don't know exactly how "towns" shade into "villages" and "cities" for me, or what features make someone more beautiful or how "beautiful" shades into weaker terms like "pretty" or "sexy" or "elegant", or how I categorize turquoise-ish colours.) So even if I choose to use the word "woman" in a way that is in fact optimized for your predictive purposes, that's only any use if I first of all tell you exactly how I'm defining that word. Which almost never happens. 2. If I do tell you exactly how I'm defining the word, then no matter what definition I'm using my uses of "woman" are thereafter strictly more informative than if I don't give you that information; i.e., strictly more informative than say 99.9% of uses of the word; I don't think you're then entitled to complain that I could be helping you even more.

None of this is particularly specific to the word "woman". The same is true of, say, the word "fish" as in Scott's initial example. I think Scott is exactly correct when he says that "in no case can an agreed-upon set of borders or a category boundary be factually incorrect". Even though a category boundary absolutely can be not optimized for any reasonable notion of optimal prediction. (And can be bad in other ways; in practice, I think being very different from other people's boundaries is a more important failure mode for the purpose of communication.)

I should maybe add that I don't actually find the stipulations above very plausible. Most social interactions are rather shallow, which means that most gender-related predictions I need to make that might be influenced by what pronoun is being used for a given person have little to do with the fascinating ways in which typical men's and women's bodies or minds differ, which means that if all I know about someone is that they were referred to as "she" or that someone called them a "woman", quite possibly the most useful information I can get from that does concern whether that person would prefer to be treated as male or female. If at some point those other deeper questions start mattering to me, almost certainly by that point I will have met them, seen what they look like, talked to them, etc., and at that point most of the information I have about their personality and preferences and so forth doesn't come from their Official Gender at all. In which case, during the period when their Official Gender is actually of much use for my predictions, the predictions I'm needing to make are in fact mostly about "how they identify".

I consider this evidence in favor of 'keeping politics out of lesswrong does help with 'rationality''.

Sorry, I'm not following the reasoning here; can you say more? Why that specific hypothesis, rather than "ZMD's writing is too long-winded in general"?

I recently read Algorithms of Deception! [LW · GW] and claim that it is of appropriate length.