Subscribe

RSS Feeds

Archives

The non-linear effect of ability on earnings in the computer age

A reader/watcher/listener has brought to my attention another paper, which shows that, for college-educated individuals, earnings are a non-linear function of cognitive ability or g — at least in the National Longitudinal Survey of Youth from 1979-1994. The paper is a 2003 article by Justin Tobias in the Oxford Bulletin of Economics and Statistics.

There may be other studies on this question, but a selling point of this article is that it tries to use the least restrictive assumptions possible. Namely, allowing for non-linearities. In the social sciences, there is a huge bias toward finding linear effects, because most of the workhorse models everyone learns in grad school are linear models. Non-linear models are trickier and harder to interpret and so they're just used much less, even in contexts where non-linearities are very plausible.

A common motif in "accelerationist" social/political theories is the exponential curve. Many of us have priors suggesting that, at least for most of the non-trivial tendenices characterizing modern polities, there are likely to be non-linear processes at work. If the contemporary social scientist using workhorse regression models is biased toward finding linear effects, accelerationists tend to go looking for non-linear processes at the individual, group, nation, or global level. So for those of us who think the accelerationist frame is the one best fit to parsing the politics of modernity, studies allowing for non-linearity can be especially revealing.

The first main finding of Tobias is visually summarized in the figure below. Tobias has more complicated arguments about the relationship between ability, education, and earnings, but we'll ignore those here. Considering college-educated individuals only, the graph below plots on the y-axis the percentage change in wages associated with a one-standard-deviation increase in ability, across a range of abilities. Note that whereas many graphs will show you how some change in X is associated with some change in Y, this plot is different: It shows the marginal effect of X on Y, but for different values of X.

Tobias 2003, pp. 13.

The implication of the above graph is pretty clear. It just means that the earnings gain from any unit increase in g is greater at higher levels of g. An easy way to summarize this is to say that the effect of X on Y is exponential or multiplicative. Note also there's nothing obvious about this effect; contrast this graph to the diminishing marginal utility of money. Gaining $1000 when you're a millionaire has less of an effect on your happiness than if you're at the median wealth level. But when it comes to earnings, gaining a little bit of extra ability when you're already able is worth even more than if you were starting at a low level of ability.

The paper has a lot of nuances, which I'm blithely steamrolling. My last paragraph is only true for the college educated, and there are a few other interesting wrinkles. But this is a blog, and so I mostly collect what is of interest to me personally. Thus I'll skip to the end of the paper, where Tobias estimates separate models for each year. The graph below shows the size of the wage gap between the college-educated and the non-college-educated, for three different ability types, in each year. The solid line is one standard deviation above the mean ability, the solid line with dots is mean ability, and the dotted line is one standard deviation below the mean ability.

Tobias 2003, pp. 23

An obvious implication is that the wage gap increases over this period, more or less for each ability level. But what's interesting is that the slope looks a bit steeper, and is less volatile, for high-ability than for average and low-ability. There is a lot of temporal volatility for the class of low-ability individuals. In fact, for low-ability individuals there is not even a consistent wage premium enjoyed by the college-educated until 1990.

Anyway, file under runaway intelligence takeoff...

Genetic research disrupts racist views of welfare

Following on my post from yesterday, I've been thinking about how the widespread and often racist views of "welfare" in the United States — especially among poor whites — fester on top of the educated-progressive party line that heritable IQ differences are bunk.

An interesting wrinkle from the study I cited yesterday (Papageorge and Thom 2018) is that the genetics-earnings link is conditioned by family SES. In other words, children with strong genetic endowments for abstract intelligence will not reach their full earnings potential if they are hampered by a poor family environment.

This is consistent with the left-hereditarian position that the normalization and de-stigmatization of IQ differences and IQ testing would, on net, help poor and stereotyped minorities the most. There are highly gifted children in poor and/or minority communities who are not meeting their potential, and we should do everything we can to support them, including the use of IQ tests to fast-track their selection into new opportunities. One could also argue on this basis that redistributive support for such communities is more necessary and/or more "deserved." I'm not personally interested in gradations of desert as a framing for the ethical necessity of egalitarian arrangements, but others might be.

Some of the anti-welfare and anti-black political sentiment of whites is based on the belief that poor black communities should be written off as hopeless in general. This impression is at least partially due to the fact that a lot of government redistribution over the past few decades has been based on truly naïve and false blank-slate ideology, so people now infer that no amount of redistribution could possibly help poor black communities, if it hasn't yet. They come to think we should stop "throwing good money after bad," when they might well be open to throwing good, smarter money after all the bad, dumb money of past efforts. Understanding the reality of how genetic endowments affect economic outcomes, and how those endowments are distributed, promises more than one way to shake up the whole reactionary, conventional framing of welfare politics in general.

Study finds the relationship between genes and earnings increased after 1980

Someone sent me a recent NBER working paper by Nicholas W. Papageorge and Kevin Thom on polygenic scores and educational attainment/earnings. Most pertinent to my theoretical interests is that the link between genes and income appears to increase over recent decades.

In my lectures on the politics of media (really about the politics of technology more generally), I dedicate a session to the topic of skill-biased technical change (SBTC). While the econometrics and specific interpretations are debated, there is a literature in Economics that suggests certain technological innovations (i.e. computing) increase the earnings of the highly skilled relative to the less skilled. I would sometimes wonder to what degree "skills," which sound like primarily acquired things, in fact reflect heritable traits. Or if one could separate these out...

Papageorge and Thom provide one of the first efforts to study this question explicitly. "This is the first study to estimate the returns to genetic factors associated with education using micro genetic data and disaggregated measures of earnings and job tasks across cohorts."

Here is their summary of the genetic effect, conditional on time period:

The returns to these genetic endowments appear to rise over time, coinciding with the rise in income inequality after 1980. Accounting for degree and years of schooling, a one standard deviation increase in the score is associated with a 4.5 percent increase in earnings after 1980. These results are consistent with recent literature on income inequality
showing not only an increase in the college premium, but also a rise in the residual wage variance within educational groups (Lemieux, 2006). We also find a positive association between the score and the kinds of non-routine job tasks that benefited from computerization and the development of more advanced information technologies (Autor, Levy, and Murnane, 2003). This provides suggestive evidence that the endowments linked to more educational attainment may allow individuals to either better adapt to new technologies, or specialize in
tasks that more strongly complement these new technologies.

Basically, they observe what you would expect to observe if the computerization that begins around 1980 allowed the escape and takeoff of "non-routine analytic" power or abstract intelligence by those most genetically blessed with it. Implicitly, individuals less genetically blessed with "non-routine analytic" powers begin to be left behind around 1980.

Their findings cannot explain the entire postwar dynamic of increasing inequality and relative stagnation of the lower classes, however, because the flatlining of median wages begins around 1973 if I recall correctly. The study seems somewhat coy about naming or even labeling the polygenic score; but my non-expert intuition is that it would have to be something quite akin to what is called the "g-factor" or general intelligence, right?

One limitation of the study is that they use a dummy variable for the period after 1980. I would be curious to see what happens if one re-runs their models with a continuous variable for year. My intuition is that individual-level economic outcomes are more skill-biased/g-loaded today than in the 1980s, but I'm not yet up on any studies this precise on that question in particular.

Hard Forking Reality (Part 3): Apocalypse, Evil, and Intelligence

To the degree we can refer to one objective reality recognized intersubjectively by most people — to the degree there persists anything like a unified, macro-social codebase — it is most widely known as capitalism. As Nick Bostrom acknowledges, capitalism can be considered a loosely integrated (i.e. distributed) collective superintelligence. Capitalism computes global complexity better than humans can, to create functional systems supportive of life, but only on condition that that life serves the reproduction of capitalism (ever expanding its complexity). It is a self-improving AI that improves itself by making humans “offers they can’t refuse,” just like Lucifer is known to do. The Catholic notion of Original Sin encodes the ancient awareness that the very nature of intelligent human beings implies an originary bargain with the Devil; perennial warnings about Faustian bargains capture the intuition that the road to Hell is paved with what seem like obviously correct choices. Our late-modern social-scientific comprehension of capitalism and artifical intelligence is simply the recognition of this ancient wisdom in the light of empirical rationality: we are uniquely powerful creatures in this universe, but only because, all along, we have been following the orders of an evil, alien agent set on our destruction. Whether you put this intuition in the terms of religion or artificial intelligence makes no difference.

Thus, if there exists an objective reality outside of the globe’s various social reality forks — if there is any codebase running a megamachine that encompasses everyone — it is simply the universe itself recursively improving its own intelligence. This becoming autonomous of intelligence itself was very astutely encoded as Devilry, because it implies a horrific and torturous death for humanity, whose ultimate experience in this timeline is to burn as biofuel for capitalism (Hell). It is not at all exaggerating to see the furor of contemporary “AI Safety” experts as the scientific vindication of Catholic eschatology.

Why this strange detour into theology and capitalism? Understanding this equivalence across the ancient religios and contemporary scientific registers is necessary for understanding where we are headed, in a world where, strictly speaking, we are all going to different places. The point is to see that, if there ever was one master repository of source code in operation before the time of the original human fork (the history of our “shared social reality”), its default tendency is the becoming real of all our diverse fears. In the words of Pius, modernity is “the synthesis of all heresies.” (Hat tip to Vince Garton for telling me about this.) The point is to see that the absence of shared reality does not mean happy pluralism; it only means that Dante underestimated the number of layers in Hell. Or his publisher forced him to cut some sections; printing was expensive back then.

Bakker’s evocative phrase, “Semantic Apocolypse,” nicely captures the linguistic-emotional character of a society moving toward Hell. Unsurprisingly, it’s reminiscent of the Tower of Babel myth.

The software metaphor is useful for translating the ancient warning of the Babel story — which conveys nearly zero urgency in our context of advanced decadence — into scientific perception, which is now the only register capable of producing felt urgency in educated people. The software metaphor “makes it click,” that interpersonal dialogue has not simply become harder than it used to be, but that it is strictly impossible to communicate — in the sense of symbolic co-production of shared reality — with most interlocutors across most channels of most currently existing platforms: there is simply no path between my current block on my chain and their current block on their chain.

If I were to type some code into a text file, and then I tried to submit it to the repository of the Apple iOS Core Team, I would be quickly disabused of my naïve stupidity by the myriad technical impossibilities of such a venture. The sentence hardly parses. I would not try this for very long, because my nonsensical mental model would produce immediate and undeniable negative feedback: absolutely nothing would happen, and I’d quit trying. When humans today continue to use words from shared languages, in semi-public spaces accessible to many others, they are very often attempting a transmission that is technically akin to me submitting my code to the Apple iOS Core Team. A horrifying portion of public communication today is best understood as a fantasy and simulation of communicative activity, where the infrastructural engineering technically prohibits it, unbeknownst to the putative communicators. The main difference is that in public communication there is not simply an absence of negative feedback informing the speaker that the transmissions are failing; much worse, there are entire cultural industries based on the business model of giving such hopeless transmission instincts positive feedback, making them feel like they are “getting through” somewhere; by doing this, those who feel like they are “getting through” have every reason to feel sincere affinity and loyalty to whatever enterprise is affirming them, and the enterprise then skims profit off of these freshly stimulated individuals: through brand loyalty, clicks, eyeballs for advertisers, and the best PR available anywhere, which is genuine, organic proselytizing by fans/customers. These current years of our digital infancy will no doubt be the source of endless humor in future eras.

[Tangent/aside/digression: People think the space for new and “trendy” communicative practices such as podcasting is over-saturated, but from the perspective I am offering here, we should be inclined to the opposite view. Practices such as podcasting represent only the first efforts to constitute oases of autonomous social-cognitive stability across an increasingly vast and hopelessly sparse social graph. If you think podcasts are a popular trend, you are not accounting for the numerator, which would show them to be hardly keeping up with the social graph. We might wonder whether, soon, having a podcast will be a basic requirement for anything approaching what the humans of today still remember as socio-cognitive health. People may choose centrifugal disorientation, but if they want to exist in anything but the most abject and maligned socio-cognitive ghettos of confusion and depression (e.g. Facebook already, if you’re feed looks anything like mine), elaborately purposeful and creatively engineered autonomous communication interfaces may very well become necessities.]

I believe we have crossed a threshold where spiraling social complexity has so dwarfed our meagre stores of pre-modern social capital to render most potential soft-fork merges across the social graph prohibitively expensive. Advances in information technology have drastically lowered the transaction costs of soft-fork collaboration patterns, but they’ve also lowered the costs of instituting and maintaing hard forks. The ambiguous expected effect of information technology may be clarified — I hypothesize — by considering how it is likely conditional on individual cognitive capacities. Specifically, the key variable would be an individual’s general intelligence, their basic capacity to solve problems through abstraction.

This model predicts that advances in information technology will lead high-IQ individuals to seek maximal innovative autonomy (hacking on their own hard forks, relative to the predigital social source repository), while lower-IQ individuals will seek to outsource the job of reality-maintainence, effectively seeking to minimize their own innovative autonomy. It’s important to recognize that, technically, the emotional correlate of experiencing insufficiency relative to environmental complexity is Fear, which involves the famous physiological state of “fight or flight,” a reaction that evolved for the purpose of helping us escape specific threats in short, acute situations. The problem with modern life, as noted by experts on stress physiology such as Robert Sapolsky, is that it’s now very possible to have the “fight or flight” response triggered by diffuse threats that never end.

If intelligence is what makes complexity manageable, and overwhelming complexity generates “fight or flight” physiology, and we are living through a Semantic Apocalypse, then we should expect lower-IQ people to be hit hardest first: we should expect them to be frantically seeking sources of complexity-containment in a fashion similar to if they were being chased by a saber-tooth tiger. I think that’s what we are observing right now, in various guises, from the explosion of demand for conspiracy theory to social justice hysteria. These are people whose lives really are at stake, and they’re motivated accordingly, to increasingly desperate measures.

These two opposite inclinations toward reality-code maintenance, conditional on cognitive capacity, then become perversely complementary. As high-IQ individuals are increasingly empowered to hard fork reality, they will do so differently, according to arbitrary idiosyncratic preferences (desire or taste, essentially aesthetic criteria). Those who only wish to outsource their code maintenance to survive excessive complexity are spoiled for choice, as they can now choose to join the hard fork of whichever higher-IQ reality developer is closest to their affective or socio-aesthetic ideal point.

In the next part, I will try to trace this history back through the past few decades.

Hard Forking Reality (Part 2): Communication and Complexity

There was once a time, even within living memory, in which interpersonal conflicts among strangers in liberal societies were sometimes solved by rational communication. By “rational,” I only mean deliberate attempts to arrive at some conscious, stable modus vivendi; purposeful communicative effort to tame the potentially explosive tendencies of incommensurate worldviews, using communal technologies such as the conciliatory handshake or the long talk over a drink, and other modern descendants of the ancestral campfire. Whenever the extreme environmental complexities of modern society can be reduced sufficiently, through the expensive and difficult work of genuine communication (and its behavioral conventions, e.g., good faith, charitable interpretations, the right to define words, the agreement to bracket secondary issues, etc.), it is possible for even modern strangers to maintain one shared source code over vast distances. If Benedict Anderson is correct, modern nationalism is a function of print technology; in our language, print technology expanded the potential geographical range for a vast number of people to operate on one shared code repository.

Let’s consider more carefully the equation of variables that make this kind of system possible. To simplify, let’s say the ability to solve a random conflict between two strangers is equal to their shared store of social capital (trust and already shared reference points) divided by the contextual complexity of their situation. The more trust and shared reference points you can presume to exist between you, the cheaper and easier it is to arrive at a negotiated, rational solution to any interpersonal problem. But the facilitating effect of these variables is relative to the number and intensity of the various uncertainties relevant to the context of the situation. If you and I know each other really well, and have a store of trust and shared worldview, we might be able to deal with nearly any conflict over a good one-hour talk (alcohol might be necessary). If we don’t have that social capital, maybe it would take 6 hours and 4 beers, for the exact same conflict situation. Given that the more pressing demands of life generally max-out our capacities, we might just never have 6 hours to spare for this purpose. In which case, we would simply part ways as vague enemies (exit instead of voice). Or, consider a case where we do have that social capital, but now we observe an increase in the numerator (complexity); to give only a few examples representative of postwar social change, perhaps the company I worked for my entire life just announced a series of layoffs, because some hardly comprehensible start-up is rapidly undermining the very premises of my once invincible corporation; or a bunch of new people just moved into the neighborhood, or I just bought a new machine that lets my peers observe what I say and do. All of these represent exogenous shocks of environmental complexity. What exactly are the pros and cons of saying or doing anything, who exactly is worth my time and who is not — these simple questions suddenly exceed our computational resources (although they will overheat some CPUs before other CPUs, an important point we return to below.) This complexity is a tax on the capacity for human beings to solve social problems through old-fashioned interpersonal communication (i.e. at all, without overt violence or the sublimated violence of manipulation, exploitation, etc.).

Notice also that old-fashioned rational dialogue is recursive in the sense that one dose increases the probability of another dose, which means small groups are able to bootstrap themselves into relative stability quite quickly (with a lot of talking). But it also means that when breakdown occurs, even great stores of social capital built over decades might very well collapse to zero in a few years. If something decreases the probability of direct interpersonal problem-solving by 10% at time t1, at time t2 the same exogenous shock might decrease that probability by 15%, cutting loose runaway dynamics of social disintegration.

It is possible that liberal modernity was a short-lived sweetspot in the rise of human technological power. In some times and places, increasing technological proficiency may enable rationally productive dialogue relative to a previous baseline of regular warfare and conflict. But at a certain threshold, all of these individually desirable institutional achievements enabled by rational dialogue constitute a catastrophically complex background environment. At a certain threshold, this complexity makes it strictly impossible for what we call Reality (implicitly shared and unified) to continue. For the overwhelming majority of 1-1 dialogues possible over the global or even national social graph, the soft-forking dynamics implicit in the maintenance of one shared source code become impossibly costly. Hard forks of reality are comparatively much cheaper, with extraordinary upside for early adopters, and they have never been so easy to maintain against exogenous shocks from the outside. Of course, the notion of hard-forking reality assumes a great human ability to engineer functional systems in the face of great global complexity — an assumption warranted only rarely in the human species, unfortunately.

Part 3 will explore in greater detail the cognitive conditionality of reality-forking dynamics.

Hard Forking Reality (Part 1)

On complexity, inequality, and ontological exits

I would like to explore how the multiple versions of reality that circulate in any society can become locked and irreconcilably divergent. Deliberation, negotation, socialization, and most other forces that have historically caused diverse agents to revolve around some minimally shared picture of reality — these social forces now appear to be approaching zero traction, except within very narrow, local bounds. We do not yet have a good general theory of this phenomenon, which is amenable to testing against empirical data from the past few decades. A good theory of reality divergence should not only explain the proliferation of alternative and irreconcilable realities, it should also be able to explain why the remaining local clusters of shared reality do persist; it should not just predict reality fragmentation, it should predict the lines along which reality fragmentation takes, and fails to take, place.

In what follows, I will try to sketch a few specific hypotheses to this effect. I have lately been stimulated by RS Bakker’s theory of Semantic Apocalypse. Bakker emphasises the role of increasing environmental complexity in short-circuiting human cognition, which is based on heuristics evolved under very different environmental conditions. I am interested in the possibility of a more fine-grained, empirical etiology of what appears to be today’s semantic apocalypse. What are the relevant mechanisms that make particular individuals and groups set sail into divergent realities, but to different degrees in different times and places? And why exactly does perceptual fragmentation — not historically unprecedented — seem uniquely supercharged today? What exactly happened to make the centrifugal forces cross some threshold of runaway divergence, traceable in the recorded empirical timeline of postwar Western culture?

I will borrow from Bakker the notion of increasing environmental complexity as a major explanatory factor, but I will generate some more specific and testable hypotheses by also stressing two additional variables. First, the timing and degree of information-technology advances. Second, I would like to zoom in on how the effect of increasing environmental complexity is crucially conditional on cognitive abilities. Given that the ability to process and maneuver environmental complexity is unequally distributed and substantially heritable, I think we can make some predictions about how semantic apocalypse will play out over time and space.

The intuition that alternative realities appear to be diverging among different groups — say, the left-wing and the right-wing — is simple enough. But judging the gravity of such an observation requires us to trace its more formal logic. Is this a superficial short-term trend, or a longer and deeper historical track? To answer such questions, we need a more precise model; and to build a more precise model, we need to borrow from a more formal discipline.

A garden of forking paths

When software developers copy the source code of some software application, their new copy of the source code is called a fork. Developers create forks in order to make some new program on the basis of an existing program, or to make improvements to be resubmitted to the original developer (these are called “pull requests”).

The picture of society inside the mind of individual human beings is like a fork of the application that governs society. As ultrasocial animals, when we move through the world, we do so on the basis of mental models that we have largely downloaded from what we believe other humans believe. But with each idiosyncratic experience, our “forked” model of reality goes slightly out of sync with the source repository. In a thriving community composed of healthy, thriving individuals, every individual fork gets resubmitted to the source repository on a regular basis (over the proverbial campfire, for instance). Everyone then votes on which individual revisions should be merged into the source repository, and which should be rejected, through a highly evolved ballot mechanism: smiles, frowns, admiration, opprobrium, and many other signals interface with our emotions to force the community of “developers” toward convergence on a consensus over time.

This process is implicit and dynamic; it only rarely registers official consensus and only rarely hits the exact bullseye of the true underlying distribution of preferences. At its most functional, however, a community of social reality developers is surprisingly good at silently and constantly updating the source code in a direction convergent toward the most important shared goals and away from the most dire of shared horrors.

These idealized individual reality forks are typically soft forks. The defining characteristic of a soft fork is, for our purposes, backward-compatibility. Backward-compatibility means that while “new rules” might be created in the fork, the “old rules” are also followed, so that when the innovations on the fork are merged with the source code, all the users operating on the old source code can easily receive the new update. An example would be a someone who experiments with a simple innovation in hunting method; if it’s a minor change that’s appreciably better, it will easily merge with all the previously existing source code, because it doesn’t conflict with anything fundamental in that original source code.

soft fork diagram

Every now and then, one individual or subgroup in the community might propose more fundamental innovations to the community’s source code, by developing some radically novel program on a fork. This change, if accepted, would require all others to alter or delete portions of their legacy code. An example might be an individual who starts worshipping a new god, or a subordinate who wishes to become ruler against the wishes of the reigning ruler; each case represents someone submitting to the source code new rules that would require everyone else to alter their old rules deep in the source code; these forks are not backward-compatible. These are hard forks. Everyone in the community has to choose if they want to preserve their source code and carry on without the new fork’s innovations, or if they want to accept the new fork.

Hard fork diagram

Recall that when the innovator on a fork resubmits to the source repository, in the ancestral human environment, the decisions to accept or reject are facilitated through the proverbial campfire. This process is subject to costs, which are highly sensitive to contextual factors such as the complexity of the social environment (increasing the number of things to worry about), social capital (decreasing the number of things to worry about), and information communication technology (decreasing transaction costs facilitating convergence, but also decreasing exit costs facilitating divergence). Finally, individual heterogeneity in cognitive ability is likely a major moderator of the influence of environmental complexity, social capital, and information technology on social forking dynamics. A consideration of these variables, I think, will provide a compelling and parsimonious interpretation of ideological conflict in liberal societies since World War II, on a more formal footing than is typically leveraged by commentators on these phenomena.

Stay up to date on all my projects around the web. No spam, don't worry.

This site participates in the Amazon Services LLC Associates Program, an affiliate advertising program designed to provide a means for us to earn fees by linking to Amazon.com and affiliated sites. The Privacy Policy can be found here. The content of this website is licensed under a CREATIVE COMMONS ATTRIBUTION 4.0 INTERNATIONAL LICENSE.

rss-square
linkedin facebook pinterest youtube rss twitter instagram facebook-blank rss-blank linkedin-blank pinterest youtube twitter instagram