Splitting Versus Blending: An AI-Inspired Theory of Neurotypical Advantage
Who knows? I might actually be on to something.
It is increasingly believed that, as Hans Asperger originally suggested, the neurological traits associated with autism are not rare and not always severely disabling; it is now believed that about 5 percent of people exhibit what I call neurosocial minority behavior, and could therefore arguably be placed on the autism spectrum. It is often suggested that we in the minority do not have an absolute deficit of social ability so much as our communication styles are objectively different, but not worse, and a disadvantage specifically because we’re the minority. That is, if we existed in a world where 95 percent of the population were in this sense autistic, and 5 percent were allistic, then the allistic population would not be superlatively socially able, but would instead be in our position of being constantly misunderstood and having their quirks misread as character flaws.
The theory has value, especially because the social disadvantage that comes from mild autism (Asperger’s Syndrome) is the main disability; we are, otherwise, better at some things and worse at others, with no clear superiority. Allistic people have faster reaction time and a lower rate of sensory processing disabilities; autistic people, controlling for general intelligence, tend to have higher levels of creativity and much stronger abilities of pattern recognition. I strongly suspect that human society needs both neurotypes to function.
All this said, our social disadvantage cannot be attributed solely to our minority status, because there are three neurosocial types—neurodivergent, neurotypical, and psychopath—and the other minority enjoys a distinct advantage. The psychopath’s empathic deficits make him a masterful manipulator of others. Psychopaths are also a neurosocial minority, but they climb hierarchies deftly and comprise 50-60 percent of corporate executives. Our issue is not only that we use a minority neurosocial language; we also, when we try overly hard (‘masking”) to adapt to the neurosocial majority, suffer in a way that psychopaths do not.
Few things in nature are truly accidental. Evolution itself seems to have no favored direction, but selection is a powerful force. Given the objective advantages of mild neurodivergence, we would expect to have become the neurosocial majority if there were not some concomitant loss of other abilities. What might it be? What is about the neurodivergent mind that, while it is more powerful in some settings, becomes a disadvantage in social settings? I believe, based on my study of artificial intelligence, I have a good first-order answer.
The past fifteen years in artificial intelligence have seen resurgence of a type of algorithm long thought dead: the (artificial) neural network. Inspired loosely by prior understandings of the human brain, in which connections of neurons encode information, the neural network simulates a complex unknown function keyed by a massive number (millions, possibly billions) of numbers called parameters that correspond to connection strength, with positive numbers representing excitatory connections and negative ones representing inhibitory action. This exists in the context of a division within AI between “left-brained” and “right-brained” approaches—I’m not a neurologist, and my understanding is that this dichotomy is overly simplistic as it applies to real human brains—whereby rules-based, exact approaches are associated with the left hemisphere and intuitive judgements with the right one. There’s no such thing as a “left-brained” person, though, as we all rely on both hemispheres to achieve general intelligence. The left brain—again, I’m not asserting that it actually occurs only or mainly in the left hemisphere in all individuals—knows how to multiply 137 by 249 using stationary facts (7 x 9 = 63) and algorithms, while the right-brain tells us the difference between a dog and a cat, not based on rules—there’s no simple “catness” rule—but thin-sliced, intuitive judgments; we “know it when we see it.”
Most of the notable progress today in artificial intelligence comes from our discovery of ways to make neural networks practical for tasks—computer vision, language modeling, and complex game-playing—where explicit programming of rules will not suffice. Writing a “cat detector” that takes a 3MB image and decides if it’s a picture of a cat is infeasible by hand. Instead, we train the neural network—that is, adjust its parameters slightly so as to progress though increasingly intricate approximations until we find one that is good enough. Exactness may never be achieved, as you’d typically need more training time and more data than can be afforded, but you can often get acceptable results using known methods. Since the neural network is not applying if-then rules but, instead, doing linear algebra to simulate a (crudely approximated) biological system, it’s also quite fast, though the training itself may be time-consuming and expensive, and the quickness of the trained model usually outweighs the cost of the uncommon error.
Neural approaches, in artificial intelligences, generally do not commit to one strategy or decision but, instead, simulate several in parallel, then use a weighted sum to vote on the output. For example, an LSTM is a substructure in a neural network that allows it to carry a memory cell that it can choose to read, ignore, or write into. The network does not make a choice, each time it is used, to do one of those things but does them all in parallel and aggregates the results; over time, when there is a clearly correct right call, the weighting on it will approach 1 while the others go to 0.
This may be an oversimplification, but rules-based approaches to AI have a left-brained feel; when you ask your maps app for directions, it applies a path-finding algorithm (likely, a variant of A*) that, if environmental conditions are constant, will return the best path every time. It does not simulate driving from Boston to New York via Mexico City with some ridiculously low weight (say, 10^-30) because it can apply rules that exclude the possibility. Language modeling, on the other hand, is right-brained: sentences are not grammatical or ungrammatical so much as they are either likely or unlikely to be produced by fluent, articulate speakers. “Can I has another cup of coffee?” is not objectively wrong but is simply far less common than “Can I have another cup of coffee?” There are rules that apply to grammar most of the time, but contexts in which they break. “Blue odor sleep moon under” is nonsensical, but it is now human language, because I used it. Judgments come not from absolutes but probabilities, and there’s a surprising amount of real knowledge that can be discovered this way. For example, the sentence “orange is between red and yellow” will be more common in a large language model than “green is between red and yellow” and, therefore, an agent that has no subjective experience and is only doing linear algebra can still “learn” where the colors sit on the spectrum. Human language, then, is not one specific thing but a very complex probability distribution—if I say, “What is six plus five?” you are more likely to say “Eleven.”—and, if we model all the conditional probabilities extremely accurately, by training against a massive text base, we can store a great deal of existing knowledge in machines that have the benefit of being able to fluently interact with us.
The advantage of the neural approach, probably explanatory of its recent successes, is differentiability (as in calculus) or “smoothness.” The probability distribution or function you must model to understand language is so complex, you couldn’t easily program it by hand and, even you did so, the inevitable changes of language would render your system obsolete. Instead, as said before, you build a model that relies on a massive list of numbers called parameters and use optimization algorithms—we know how to do this quickly and, for this purpose, well enough—to find good settings by stepping through “parameter space” in favorable directions and avoiding unfavorable ones. The tiny incremental changes, the small numerical adjustments to the parameters, work; your model gets better and better until, in a large number of cases, it starts getting the right answer as often as you can, or more often.
Rules-based reasoning is exact but can be impractical. We are forced, often, to decide whether the correct model of a natural phenomenon is X or Y. In pure mathematics, where nothing but exact proof is accepted, this knowledge really matters. On the other hand, the neural approach allows us to operate as if a combination, 0.37*X + 0.63*Y, were true and consider both in parallel, then take a weighted sum.
If we return to the topic of neurodivergence, we should note that there probably isn’t one discrete “thing” called autism. Instead, the label seems to be used in cases of disabling multiple neurodivergence. A person who is monotropic but otherwise neurotypical will be diagnosed as having attention deficit disorder, which is an awful name for what is, in fact, attentional hyperfocus rather than absence. Soemone who is otherwise neurotypical but is monotactic will be perceived as strange and socially inept, but is likely to have near-average life skills and above-average social awareness (since monotactism suggests high, not low, empathy) and might be diagnosed with PDA autism (PDA stands for “persistent demand avoidance”, another case of severe misnaming) if her divergence is severe enough to merit a diagnosis, which it often is—monotactic people, even with no other divergence, tend to struggle with employment. People who use the neurosocial minority language—explicit communication favored over subtle, nonverbal, and sometimes unreliable cues—but have no intellectual impairment used to be diagnosed as having Asperger’s Syndrome, which is now believed to be (and, in my view, probably is) mild autism. We can debate at length why these neurodivergences exist and how multiply neurodivergent a person must be to be considered “really autistic”, and we can note the existence of a spectrum because at least 10 percent of the population has some degree of neurodivergence, but one distinction that seems to be discrete and categorical—in other words, cut-and-dry—is a person’s default neurosocial language.
Differences in neurosocial language can cause difficulties, especially early in life. For example, neurosocial majority (“allistic”) people prefer, when making requests of others, to do so subtly through nonverbal signals. Hints of facial expression, usually involuntary, are the first tier; body language and tone, now under conscious control, are the second. Explicit verbal communication is their third tier of conflict resolution. An exception exists when they consider the other person so low in social status as not to merit the time cost of the conventionally polite approaches in the lower tiers, and jump immediately to a third- or fourth-tier response. Since we in the neurosocial minority (“autism”) do not natively get these social cues—we can learn them, but this learning tends to be rules-based and therefore not as fluent—and we do not realize that people are displeased with us until they explicitly tell us, we unintentionally force them to go into their third tier, which they generally do not like. At the same time, explicit communication tends to be our first-tier approach, so we use it liberally. Neurotypical people tend to misinterpret this as either severe displeasure—they wonder what they did to get us to tier 3 so quickly; they conclude we are quick to anger—or a deliberate attempt to lower their status, though it’s nothing of the sort. We do, as we age, gain the ability to read and use much of the neurosocial majority language, but we have to do “in software”, using “left-brained” analysis, what other people are able to achieve “in hardware” using “right-brained” intuition. It takes three times as much effort. Now, it is often worthwhile to expend that effort—if we’re going to expect others to adapt to our neurosocial language, we ought to learn theirs—but we never become so fluent that we can thrive in, for example, competitive social environments like corporate workplaces where the rules are constantly changing—in fact, constantly being manipulated by adverse actors who benefit from confusion and displacement—and where tiny differences in the degree of one’s social polish can spell the difference between promotion and termination. Such an environment is also one where trust is, by necessity, a sliding scale—allies can become enemies at a moment’s notice. It doesn’t make sense in intimate life to “70 percent trust” somebody but, in the corporate world, you have to take such attitudes, because trusting nobody (autistic self-reliance) will make you seem aloof, while most people will rob you blind if you trust them 100 percent.
Verbal language is one place where we meet, neurotypical and neurodivergent alike, and neither of us is perfectly native. We divergents tend to think in complex structures that must be linearized to fit into strings of symbols (that is, sequences of words that will be heard or read as grammatically correct sentences) and it takes time to get good at it. Neurotypicals tend to think in terms of images and impressions and inferences that are often genuine, but cobbled together from masses of unreliable data (“thin slicing”) that become useful only in aggregation, and therefore they struggle when asked to verbalize where their opinions come from. (This is seen by them as confrontational.) Surprisingly to most people, at comparable levels of life experience and general intelligence, neurodivergent people tend to be more charismatic in settings where verbal language dominates. We’re often great writers and we can learn to be public speakers—allistics dislike public speaking, because they fear social rejection so much, but we know that rejection, while painful, is something one survives—because, when one is speaking, there is usually a reason why, and purpose tends to suggest a structure we can use to find our way. On the other hand, if you put us in demanding, volatile social situations—subjects constantly changing; people dropping in and out of conversations; in some cases, adverse actors—and there is a level at which we can no longer keep up with the chaos and we will glitch. We tend to be charismatic in structured settings whose purposes we understand; we tend to be incapable in messy ones whose demands differ from one moment to the next.
I strongly suspect that neurodivergent and neurotypical people are both necessary to human society. Neurotypicals, as they like to remind us, have better social skills—and this is objectively true, because the baseline assumption common to all humans of others being just like the is true, for them, 95 percent of the time, as opposed to 5 percent for us—but they are in fact terrible at building societies. They have never built one that did not require violence to solve even the simplest problems—nations get more and more dysfunctional until they try to solve their issues with war, which itself has a pathetic success rate at making life better. As soon as their societies scale beyond 150 individuals, decay is almost guaranteed. In fact, we know societies look like after they drive out neurodivergent folk, because business corporations, with the luxury of it being facile (at least, in the US) to fire people, successfully do get rid of as many neurodivergents as they can when they move into the maintainance phase, and the result is that they turn into cubicle farms, places of uninspired bureaucratic mediocrity where even neurotypical people find themselves alienated, disgusted, and bored. Without us in the mix, a society faces inexorable decline, but we neurodivergents are so different, not only from neurotypicals but from each other, that it’s not clear to me that we have a natural talent for organization. We are, for example, repulsed by hierarchy, even when it benefits us, and even in circumstances (rare as they are) that require it. We’re often fantastic strategists and tacticians, but we don’t make great administrators or convincers. We probably need neurotypicals as much as they need us.
I suspect the first person to write explicitly about neurodivergence in the West is Plato, whose Republic details the decay sequence of a society by positing the existence of five regimes, each worse than its predecessor. The first and best of these, he said, was aristocracy, which in his time did not mean what it does today (rule by a propertied hereditary elite) but, instead, rule by the people most suited—άριστος means “best”—whom, he argued, were gold-souled philosopher-kings detached enough from ordinary human desires, biases, and appetites to make the best decisions. Unfortunately, Plato recognized, such people are usually not the sorts who are chosen as leaders by others; their impartiality and excellence can make them disliked. Thus, charismatic men of lesser mettle—silver-souled leaders-of-men—shall replace us over time, leading us into the second regime, timocracy. The timocratic society still values the original “gold-souled” virtues, but it has also learned not to rely on socially awkward moral or intellectual outliers, so the trappings of virtue—τιμή means “honor”, but also “price”, suggesting the tradeable nature of esteem—become, rather than the thing itself, sources of prestige. The timocratic society still goes through the motions of being what it once was, but is losing its way and will fall into the third regime, oligarchy. Once the society has become an open market, where anything can be bought, moneyed people accrue the markers of achievement and virtue (τιμή) and people who excel in the uglier aspects of business thrive—a good man is limited because he will not do evil, but evil men are not unable to do good; in fact, the evil man will do all good things that suit his interests, and therefore has the game-theoretic strict advantage of a larger option set—and rule by iron-souled men of commerce, in this “might makes right” world, naturally follows. Timocrats rule because they are loved; oligarchs, because their resources make them feared. If there is one virtue of oligarchy, it is what Plato considered the virtue of temperance, which we would recognize as the appearance of bourgeois restraint. For the most part, our elites do not, as tyrants do, openly flout the claimed values of our society.
Plato viewed democracy negatively, but the fourth regime, as he described it, is not what we would call democracy today. It starts when oligarchic dysfunction becomes severe enough that people are forced into self-reliance. Institutions are distrusted and few monuments or important organizations will be made. This is not general self-rule—Plato was of the opinion that no such thing can exist among humans, because hierarchies always form—but a state in which nobody knows who is in charge; perhaps ochlocracy would be a better term. Since it exists concurrently—global capitalism is in its post-oligarchic stage—with democratic backsliding, we’ll call it post-oligarchy or ochlocracy. Oligarchy is might-makes-right, but hidden by hypocritical pretense; in ochlocracy, people have ceased protecting society is some other way. In the power vacuum, the potential exists for unabashed thug—most likely, a clinical psychopath—to gain popular support to gain power and use the appearance of continuing popularity to hold it. The tyrant does not merely ignore the society’s historical values; he actively seeks to destroy virtue, because it is a threat to him. He will eventually destroy his host civilization; the best one can hope for is that something else may someday rise from the ashes, restarting the cycle anew.
I don’t agree with Plato’s theory wholesale. I don’t think moral excellence is an inborn trait limited to those born with “gold souls” and I certainly don’t think it’s limited to neurodivergent people—it’s possible that we tend to both extremes (for example, we’ve produced such disgusting specimens as Elliot Rodger) but we clearly don’t monopolize either one. Moreover, Plato’s model is also, in my view, historically inaccurate. I don’t think gold-souled people were ever in charge; I don’t think his definition of aristocracy—rule by the best—corresponds to anything that ever existed. It’s a mythology that forms after the fact; business organizations spawn as timocracies but mythologize themselves as being the unique ideas of “visionary” founders, even though it usually takes at least one man of commerce to get original investment. Still, the four other regimes that Plato described do seem, in business and in nations, to exist in succession. For an example at the large scale, the aftermath of World War II led to two very different but functioning timocracies, one in the West and one in the Soviet Union, from 1945-1979. In our society, τιμή took the form of credentials and connections to important people; in theirs, it was bureaucratic position. The 1980s saw a transition to oligarchy; the Soviet Union did not survive it, while the capitalist West did but became a far worse version of itself. We exited oligarchy, entering a post-oligarchic world, around 2011 when it became clear that the economic crash of 2008 had brought permanent changes (“the new normal”). This was not a period of democracy—in fact, we never were, on either side of the Cold War divide, very democratic; we all lived and still live under massive, atypically ideological, bureaucratically-governed nation-states—but of democractic backsliding; both the 1980 and 2010 transitions have made us less self-ruling, not more.
Plato’s ideal society was ruled by a just aristocracy of people who had no desire for wealth, property, or worldly status—such an inclination is highly correlated with neurodivergence—as well as a superlatively logical manner of thinking; silver-souled leaders of men (also known as “high-functioning allistics”) would implement the ideas of these secular monks. This golden age probably never existed; I doubt we were ever really in charge. Under timocracy, though, we are often given a place; our knowledge and originality, because it connects us with prior challenges and possible futures, allows us to be heard and have a small amount of influence. We are never trusted—it is an open question whether we should be—to decide what happens, but we get a chance to speak. As society moves into oligarchy, we are ignored—the true values of society have diverged from the stated ones we still believe in—and, when it goes into post-oligarchy, we’re distrusted. When it becomes a tyranny? Often, we are the first ones driven out, or sent into labor camps, or otherwise discarded to make an example. The tyrant must destroy every remnant, every memory, of a different time.
It is not, of course, solely our virtues that put those of us in the neurosocial minority at a disadvantage. I suspect that we think in a different way, or are at least biased toward different modes of thought, when it comes to social behavior. Recall what I said earlier about “right-brained” artificial intelligence—the neural network, which learns by stochastic gradient descent on a dataset to model an unknown function. It holds knowledge diffusely, not in a language of facts or rules, but in approximations that, while biased, can be blended together to make more accurate ones, until the error is tolerably small. A neural network does not commit to facts being true or rules being valid; it applies numerical weights to all features, and one trusts the algorithm to increase important features while diminishing (to zero, at which point they are ignored) irrelevant ones.
Let’s consider this hypothetical situation. You’re approached by an old college buddy (call him “Bob”) who has become a wealthy influencer, and he invites you to do a stunt that will draw $200,000 in audience revenue: play Russian roulette online, live. He will provide the weapon and set everything up, and he assures you that he has rigged the gun so its cylinder will land in the same place every time—he shows you device—and so you have nothing to worry about. You and he will split the prize, so you stand to gain a hundred thousand dollars. You know that Bob’s the sort of person who never leaves anything to chance; for this hypothetical, you can assume total competence. The only unknown is Bob’s moral character—he might intend for you to survive, and to split the reward; he might also intend that you die, so he keeps everything. These are two distinct possible worlds: X is the one in which Bob is honest; Y is the one in which he is dishonest. Although your view of this world is allowed to be probabilistic—you’re allowed to represent your belief in each numerically—the underlying world is not. Bob has already made the decision. One of those worlds is absolutely true, the other being utterly false, and it matters quite a bit which is the case.
The question here is not “What do you do?” but “How do you decide?” A neurodivergent person will tend to consider both X and Y possible and, therefore, try to find a policy that would acceptable in both. This is why neurodivergent people, though most of us are capable of lying when we need to do so, are so averse to deception. Whenever you tell a lie, you’re splitting the world: there is the real world, A, that you know to be the real one, and then there is A’ that you seek to convince others is the real one, because there is some way it benefits you for them to act according to their preferences in it. This gets trickier when some people already know, or there is a need for them to know, A, while a separate set of people should believe A’; it’s hard to ensure they don’t communicate. Then consider that most people who make their way through life by lying are going to lie a lot, so there might be ten different matters on which they feel a need to lie, with each one having different policies for who should be told the lie and who needs to know the truth. So now there are ten splits: A vs. A’, B vs. B’, …, J vs J’, and 1,024 possible worlds. Nobody can keep track of a thousand possible worlds, not fluently. And then there is the matter of having to detect other people’s lies, and to deal with hidden information more generally. We can’t actually navigate millions of possible worlds; all of us have to approximate.
As I said, an autistic person is more likely than an allistic person to consider the ramifications of both possibilities. Consider X: Bob is a good man and he’s telling the truth. If we ignore our decision’s effects on culture—it is probably not great for the world if someone publicly profits by playing Russian roulette—it makes sense to take the million-dollar offer, since there’s no safety risk. Then consider Y: Sure, Bob is telling the truth that he will rig the gun, but to kill us. Unless Y can be ruled out absolutely, it seems unwise to take his offer. We don’t hate all kinds of risk, but we’d prefer to make decisions we can live with in all possible worlds, and dying stupidly is the opposite of “living with” anything, by definition. So, most of us would turn down that offer, knowing that people exist who would have someone else die to take all the money for themselves.
An allistic person is also unlikely to take Bob’s offer, but for a different reason. He feels on some subconscious emotional level that a blend of the worlds is true; say, that 0.37*X + 0.63*Y is the real world. It is debated—this is a form of the argument between Bayesians and frequentists—whether these weights are probabilities in any rigorous sense, but there is strong evidence that they function as such in decision making. The two possible worlds aren’t split and analyzed separately. Instead, the evidence—Bob is a college friend, but it’s been a while, and his offer is a strange one, and what of person discovers that there is money in Russian roulette?—is compiled and aggregated into an approximate model of the world that is often good enough. Operating against 0.37*X + 0.63*Y, the allistic person is unlikely to take Bob’s offer unless in desperate need of money, and therefore gets the same result, even though this 37/63 blend of candidate realities corresponds to nothing that could actually exist.
In other words, in this single-split example, most autistic people would decline Bob’s offer because in the world where Bob is untrustworthy it is a bad idea to do so, while most allistic people would decline it not because they are sure Bob is untrustworthy but because he is not trustworthy enough to risk one’s life. The autistic mode of reasoning is more precise, but becomes inefficient when there are a large number of world-splits that must be considered; the allistic version, on the other hand, is very fast and feels—because these weighting parameters are never experienced or even known to exist—intuitive. “I got a bad vibe about Bob when he made that weird offer.”
In artificial intelligence and software programming, there are problems—security, formal verification—where left-brained certitude is mandatory, and there are others where the right-brained efficiency is more valuable. A neural network is not forced to choose one strategy when solving a problem; it can consider a bunch of them and aggregate the data using methods known to be imperfect but, usually, without undesirable general bias (unless bias exists in the training data, which is a common problem.) An artificial neural network with a memory cell (“LSTM”) does not choose, each time it operates, to read or not to read the cell’s contents; instead, it might “99% read” and “1% not-read” the contents and operate on the weighted sum of computations. If we apply this to possible worlds or policies, we see an advantage of the right-brained, fluid approach. If there are 100 splits, there are 2^100 possible worlds—far too many to explore, even using supercomputers—and these exist in a discrete “parameter space” where it is not always clear where an evolving algorithm (e.g., a neural network) must “jump” to find better results. To admit a continuum—to say we are not in X or Y, but in some blend of the two—increases the state space we have to worry about, but gives us the ability, due to its continuous nature, to move gradually through it in search of better solutions. If we know from experience that we get better results from 0.38*X + 0.62*Y than 0.37*X + 0.63*Y, we might consider exploring 0.39*X + 0.61*Y. The mathematically intractable problem of deciding what world we are in is not solved, but we can arrive at a weighted-sum approximation that, in most cases, results in acceptable behavior. We can learn that people who smile are usually happy, even though it is not always true; we will learn that people offering us opportunities to get rich quickly are usually scammers who might hurt us.
The allistic advantage in social interactions, therefore, seems to be in their ability to subconsciously consider a multitude of possibilities and blend them together in a way that is sometimes inaccurate but that also enables them to solve social problems more fluently and quickly than us. We still end up, in the end, having to approximate the world we are in; we are just less fluent at it. It is our tendency, when we can do so, to consider that {(A, B), (A’, B), (A, B’), (A’, B’)} might all be “the real world” whereas they are likely to have decided on a specific blend—(0.43A + 0.57A’, 0.19B + 0.81B’). This ability to separate costs accuracy and reduces the ability to spot patterns—A/A’ and B/B’ might not be independent, for example—but enables rapid adaptive social learning; every surprising social outcome is new data that leads the weights to shift. The neurodivergent view in which things are objectively true or false can lead to rigidity; the blended approach allows a person to adjust weights gradually on new information.
If we all have to approximate, and if blending gives such a computational advantage—the neural approach has shown success not only in computer vision and language generation, but also in game-playing—then why does our splitting approach remain at all? Why can’t we all just blend?
The answer is that, while the allistic social fluency in blended possible worlds is a gift, it has its issues. In the Russian roulette example, the difference between 0.37*X + 0.63*Y and 0.38*X + 0.62*Y does not really matter, because whether you “37% trust” or “38% trust” Bob, you are not going to take his offer. In general, though, these tiny differences in trust levels do matter. They are what we call social status. Cognitively, we know that high social status and trustworthiness are not positively correlated—in capitalism, fortune and high status are usually won by someone who acquires trust from a person who should not have given it—but, even still, there is a general sense in us that high-status people and especially high-status institutions deserve more trust, if only because they have more to lose, whereas low-status people will do anything to trade. It should not be surprising, either, that people who are skilled at acquiring undeserved trust tend also to achieve high status. At 0.38*X + 0.62*Y, you’re not very likely to take Bob’s Russian roulette offer, but he’s a higher-status Bob than the 0.37*X one.
The issue is that bad actors can hack these tiny differences in the weights we use to blend candidate realities. You can, if you have 10 million pictures of cats and 10 million pictures of dogs, apply state-of-the-art machine learning techniques to a neural network and develop one that can accurately classify most images. There might be tiny biases—the network might have erroneously “learned” that pixel (137, 54) having color #3a5cd0 means the image is 10% more likely to be one of a cat—but, in the general case of real-world data, these tiny biases cancel each other out and do no harm. However, a person who has access to the neural network’s configuration and who can find these biases might be able to, for example, take an image of a cat, then perturb it in a way that is imperceptible by humans but, by leveraging these otherwise inconsequential biases, will be classified as a dog. This is called an adversarial example, and it can be used to cause harm; it would not be good for us if someone found a way to make self-driving cards recognize certain stop signs as saying, “Speed Limit 95.” Neural networks make this “intuitive” reasoning highly efficient, both to develop and to use, but these techniques are not immune to hacking.
Allistic social reasoning, as discussed, blends all the possibilities together, probably at a neurological level. Subtle social signals (which we tend to ignore, because they don’t prove anything) are read for their most-probable meaning while the alternative possibilities are also subconsciously noted and included in the aggregation. The possibilities—“Bob is a treacherous shitbag” versus “Bob is a good person and can be trusted”—both exist; deductions are seamlessly aggregated, and policies are developed. There is no need to consider all of the millions of possibilities; it couldn’t be done anyway. Besides, most people are not inherently and purely good or evil but behave according to context, just as few of us are inherently slow or fast drivers but choose speeds according to road cues and other drivers and laws, making it important to read an environment as well as people. This allistic behavior is more efficient than what we do and it works a very high percentage of the time. When it goes wrong, though… the consequences can be dire.
Those tiny differences between 0.37X’s and 0.38X’s are, of course, what determine a person’s social status. One has to read not only whether one trusts a person, but whether others are likely to do so. This, by the way, is why a rumor that nobody believes can still destroy your life—even if no one accepts the truth claim that is its payload, plenty of people will still feel that there must be something wrong with you, or the rumor never would have existed in the first place. Sure, your reputation is at 0.001*Thief + 0.999*NotThief—we’ll ignore the fact that more than 0.1% of people are thieves; again, these are emotions more than they are objective probabilities—and so nobody thinks you’re a thief, but what the fuck kind of social incompetent are you if people even had the doubt? This is the first problem; even if you’re proven innocent, people assume you must have done something for the accusation to get as far as it did. The second thing is that nobody wants to be read themselves as “99.9% not-thief” so they will go out of their way to avoid association with you and your unsavory reputation, even though they’ll privately confess that they know (or, at least, strongly believe) you’re innocent.
On the matter of social status and small differences in parameters, you probably wouldn’t trust any Bob in a game of Russian roulette, and you probably shouldn’t, but would you go out for lunch with a stranger? Do you want to hire this person, knowing there’s a 0.1% chance of him being a serial embezzler? Character is revealed under circumstances we hope never occur; therefore, we are forced to estimate in peace time, without any real knowledge, and rely on social signals that do not always work.
And here’s where it gets weird and pretty fucking unfair. If we consider which is the better social reasoning strategy for one individual in isolation at one specific point in time, the allistic one wins. It allows people to make inferences quickly and react to others in real time, and it also helps them combine information sources of varying quality without forming precise beliefs or deciding who they do or do not trust, since we can probably all agree that trust is contextual and that there are degrees of it. The problem is that there are people who are really good at exploiting others subconscious biases to gain trust and social status, and they tend to be good for anything else. They’re called psychopaths, and they work their way, increment by increment, through subtle changes in appearance and circumstance and framing, from 0.37*x + 0.63*Y to 0.99*X + 0.01*Y, or even further, and get people to do what they want. It is not clear whether psychopaths should be called neurotypical or neurodivergent; they seem to be in their own category: they’re not considered neurotypical because the deletion of conscience causes problems for society, but they experience no social disadvantage or reduction of economic opportunity. On the contrary, they are usually exceptionally skilled at ingratiating themselves to those in power. Psychopaths understand intuitively how to become adversarial examples and win undeserved trust and social status, and they do it all the time. Low-functioning psychopaths end up as criminals; high-functioning ones become corporate executives. Though his terminology was different, Plato knew about this—the “silver-souled” neurotypicals who displaced us gold-souled weirdos were not psychopaths, or even bad people, but the “iron-souled” oligarchs and later, tyrants, who hacked their way to high positions in neurotypical society, are.
It is sometimes argued by neurodivergent folk that neurotypicals are inherently easier for psychopaths and bad actors to hack. I don’t think this is true. Neurodivergent people end up in abusive relationships at least as often as neurotypical people, and we probably have an above-normal susceptibility to cults, due to our black-and-white thinking. Although we default to approximate (or even wrong) methods of social reasoning later than neurotypical people, it simply incorrect to assert that we don’t do so, because (as said before) nobody can actually work against each of the millions or billions of possible worlds. We all have to “collapse” the world to reason about it, and we might do so more reluctantly than neurotypicals, but this doesn’t make us better at it. Individually, we’re almost certainly as hackable as neurotypicals—once a psychopath figures out how we collapse the world, he can exploit this knowledge. However, it’s much harder to hack ten of us at the same time than to hack a group of ten neurotypicals, who are more amenable to gradient-based approaches and who are often be unduly influenced by the opinions of those with the highest status. And that’s the difference. A psychopath can convince ten neurotypicals in a room of something using the same strategy; to trick ten of us, he’d have to use several different strategies, keeping his deceptions straight, and hoping we don’t work together to figure it out in time to stop him. In practice, he’ll usually never need to this. Instead, he’ll pick us off one by one, dispatching us by turning the neurotypical majority against us, so that we are gone and he does not need to contend with, nor manipulate, our opinions at all.
Almost everyone has learned by some point that the social and ethical immune systems of neurotypical society fail sometimes. Processes that are designed to detect and punish serial bad actors and manipulators—psychopaths—misfire and this is something we just accept. Sometimes, psychopaths will develop ways to use this immune system as a personal weapon; reputation is supposed to be a signal of social and moral worth, but it rarely is one, because it is so easily manipulated by the sorts of people who attain power.
Those of us who are neurodivergent had to consciously learn social abilities that other people take for granted. We often appear, especially in chaotic environments when we are overwhelmed, to be attempting to hack them, which they obviously do not like. A psychopath who is trying to hack a neurotypical person will show no signs, because it’s so easy for him. A neurotypical person trying to hack another neurotypical, unless comically inept—as we see in children, for example—will show only slight deviances, such as possibly a reduction in eye contact and possibly a change into a more pedantic vocal register, since (incompetent) people who are trying to deceive others will go out of their way to sound objective. Neurodivergent people, in general, figure out by early adulthood that we’re never going to be fluent in hacking other people; still, our mannerisms are seen as weird and we’re often taken to be far more calculating and adverse than we actually are.
The neurotypical social-ethical immune system does probably work against misbehaving neurotypicals—it probably does detect when people are flagrantly dishonest or dangerously incompetent—but it is easily bypassed by psychopaths, and it goes haywire against us.
Psychopaths understand that, in the world of human social dynamics, there isn’t really such a thing as truth. Reputation is an opinion that considers itself a fact, no more. Psychopaths intuitively get that power relationships trump truth because the latter so often doesn’t exist. “Tom isn’t a team player” or “Mary hates every idea that isn’t hers” are the sorts of things that are said in the corporate world and take on the status of being true, regardless of whether they are. Neurodivergent people tend to split the world each time there is an unknown and track as many possibilities as we can handle, but of course we have to discard some because there are exponentially many of them. Neurotypical people operate more fluently and efficiently in a blend of numerous possible worlds with no awareness of their doing so. Meanwhile, at least in the context of corporate and political ascendancy, it is psychopaths who have got the game most figured out: isolate the neurodivergents so no one listens to them, and win over the neurotypicals with social proof attacks.
Organizations, as they grow, find it useful to present themselves as consistent; they do not want to be taken for running on mere opinions, because opinions change, and because the insiders (or upper class) do not want the marginal members to perceive division, and therefore build up fact bases (although the material in them may not be factual; it may even be dishonest) and precedents and if-then rules. In other words, they begin splitting possible worlds and making choices instead of merging and allowing subtlety. This leads to organizational bureaucracy, which tends in practice to combine the worst of autistic and allistic tendencies, in the same way that corporate is a gender combining the worst traits of men and women with none of their virtues. Over time, this becomes a formal hierarchy, an apparent counterexample to my claim that neurotypical people prefer to merge disparate worlds. Hierarchy forces the organization to make a discrete choice: it cannot set a policy of listening to Bob 37% of the time and Carl 63% of the time; it must choose a leader. Internally, management may not be necessary, as groups of motivated individuals need very little guidance or control, and authority can be exercised only in the rare case of an otherwise irresolvable conflict. So, I suspect that external concerns are what first push a group to choose leaders—outside parties will assess their own standing with the organization according to the rank of the people sent to deal with them. Of course, economic greed also drives hierarchical development: social status and the obfuscation it affords can allow the people who contribute the least to capture the bulk of the reward, and not everyone can resist this opportunity. Ranks and levels are designed and calcify into place. Performance reviews are set to a schedule and, often, it is understood that a percentage of people will be fired. Now, bona fide autistic people hate this shit, because it’s often used against them, but all of this has the aura—separation into discrete categories, boatloads of statistics that theoretically represent performance, rules and triggers—of being something autistic people might design, for fun, if they weren’t aware of the fact that doing so would hurt people. (Put simply, hierarchy is the neurotypical special interest.) Of course, it becomes clear that the corporate system’s pretense of impersonal meritocracy is not real—rules for thee and thine, exceptions for me and mine—and, of course, the beneficiaries of these systems, where stated rules and facts diverge considerably from the far more socially unacceptable policies on which the organization really operates, tend to be psychopaths. We have observed in the technology industry, but it is more generally true, that bad actors will sometimes use what I call weaponized fake autism as cover. There are a lot of executives in Silicon Valley who affect the mannerisms of Asperger’s, and may even claim to have it, to disguise their harmful antics, as if a diagnosis made it impossible to judge them as bad people. Well, I will say it. Just as there are bad people without the autism spectrum condition, there are bad people with it.
Why do neurotypical people tolerate such duplicity in organizations? The answer, I think, is that there is a continuum of governance between two extremes that most people, whether neurotypical or neurodivergent, despise. There are cold/rigid systems and there are warm/sticky systems. We tend to hate them for different reasons. A cold/rigid system is one where rules were set long ago and the right to relax them does not really exist; people blindly follow protocols they do not understand, and one begins to feel both a lack of trust in the system—how does anything get done around here?—and personal invisibility, because favors, even when they would be mutually beneficial, are not extended. On the other hand, in a warm/sticky system, rules are informal and underspecified, and being liked by the people in power is wat really matters. Since most people will not have important connections or be personally favored—at least, not enough to attain the quick promotion or easy opportunity that make warm/sticky chaos worth putting up with—the out-of-favor majority will grow resentful, as if their lack of favor were a personal attack, and feel justified in pointing out inconsistencies and disparities until morale plummets. The purpose of the corporate hierarchy is that it commits to being neither. No one actually knows what the difference between a VP and an SVP and an EVP is; it’s all just words. The real objective is to ensure that upper management holds power no matter what; the system will, toward this end, exhibit phase changes into cold/rigid (autistic?) or warm/sticky (allistic?) presentations, as circumstances demand, at a moment’s notice.
Neurodivergent people, in general, despise interpersonal social hierarchy. It is not known whether this is an innate virtue, or a trait we have because we tend to be below-average at exploiting and climbing them. Conceptual hierarchy is not foreign to us—pattern recognition, we’re good at—and we understand well that specific occasions require some people to take charge and others to follow orders. This isn’t the problem. The issue is that neurotypical society—and I don’t blame neurotypical individuals for this—goes beyond the establishment of temporary, useful hierarchies and instead creates permanent rankings that are held to be comprehensive and that begin to exist for their own purposes. When neurotypicals do split the world, they split hard.
It is a myth that executives work for companies. Some managers do; most line and front-end managers are solidly employees, and many middle managers are employees with delusions about their status. Executives, however, manage to convince corporate bodies that they are the apparatus’s raison d'être. It’s shocking how easily they do it. Thus, executives don’t work for companies; companies work for executives. The high pay and lack of responsibility—your buddies write your performance reviews, and while you’ll sometimes have to take a dive and get fired, to appease a shareholder or few, it’s very rare that this decision is made without others in that social class finding a place for you of comparable or higher compensation first—make it clear why people who would seek the position, and it tends to be the psychopath who gets it. Psychopaths intuitively know that the organizational rules mostly aren’t real and that they can do what they want as long as they remain in control of others’ opinions of them at all times. The warm/sticky character of high-ranking people is an asset that can be exploited; so, too, is the cold/rigid official nature of the bureaucracy, which can often be taken to absurd levels of literality to reify one of the psychopath’s favorite tools, and one of the neurodivergent person’s age-old enemies: weaponized fake autism.
I’m far more private than most people these days, so I try to avoid autobiography. I could write a compelling memoir, but I’d rather finish my novel, which is more important to me anyway. Still, I should confess the realization that brought me to accept that I probably am autistic—I am clearly some kind of neurodivergent, with sensory and social challenges both, and I fit the template for what used to be called “Asperger’s Syndrome” quite well. People have joked about me having it since I was young, but I realize now that most of them weren’t joking. All that said, social ineptitude during adolescence is not itself proof of autism—it’s far too common—and I was not entirely convinced of my condition until, for unrelated reasons, I read about autistic meltdowns. All my life, I’ve had episodes in which my senses became painfully acute, as if my vision and hearing were at more resolution than I could process; as an adult, these sometimes evolve into anxiety—I might “stim” or need physical exercise to calm down—but don’t lead me to lose control or “act out” but, when I was a child (from about three to twelve) these experiences were less controllable. Most of the time, back then, I wouldn’t remember what I had done. I got in a lot of trouble, of course, and in the 1980s no one would have considered this autism, because my verbal development was well above average. I, more or less, accepted that I would get in a lot of trouble for things that I couldn’t control, and it didn’t offend me because some of the things I did during those episodes, as a child, were objectively obnoxious or harmful and I basically agreed that I ought to stop doing them.
I wasn’t spoiled as a kid. I had good parents—great parents, actually—but I knew that these loss-of-control episodes had about a zero-percent success rate of producing desirable results. So then, as I read about the differences between autistic meltdowns and neurotypical tantrums, I realized that the internal experiences could not be more different. A child’s tantrum is a deliberate, manipulative effort to get something from an adult; they do not understand why, but they know on some level that adults are terrified of seeing children in autistic or epileptic distress, and discover methods of using this to their advantage. I never once did this. The idea that I might be able to get attention or material gifts by faking a loss of control was simply not there, because I remember how terrifying it was to actually lose control, and how bad the results often were not only for me but for people around me. Anyway, I got in trouble constantly in my early years of school and never understood—half the time, I didn’t remember any of it—why; it was just my story that I was a shitty kid who, for some reason, stopped being shitty around the age of eleven.
One of the things that angered me back then, and it still does but to a lesser extent, was that adults would try to come with motives for why I had done things, and these were always wrong. If I broke some other kid’s toy for sensory reasons—I won’t defend it, I obviously wouldn’t do it now—the “well-meaning” adults were quick to accuse me of being envious, which just enraged me. Here’s something to know about us as neurodivergents: we’re almost impervious to envy. Of all the disliked emotions, that’s the one we have the least—anger and fear and disgust, sure, we’ve got those, but the comparison monster that afflicts neurotypicals basically never visits us, because we understand subjective experiences to be fundamentally incomparable. We, obviously, sometimes see that other people have things we want, but the toxic unreasonable envy that might turn a person toward violence, we don’t have. Resentment, we do, but there’s a huge difference. Envy is the irrational dislike that some neurotypical people feel toward others who possess things, in an attempt to justify taking them or causing harm to someone who often did nothing wrong; resentment, on the other hand, is valid contempt toward a society over an injustice that makes clear its moral incompetence. This distinction is not made often enough. Anyway, I remember being furious, as a child, when a number of adults—not all of them, and my parents were far better than most—tried to understand me, failed belligerently at it, attributed motivations and character flaws to me that were not there—it had to be envy, not sensory revulsion, that led me to act out—and insisted, when I tried to explain what was going on, that I was making everything up.
So, now that I’m older and understand human psychology—my own, and others—quite well, I have got to say that, while I would never feel anger toward a child for being a child, I do see the neurotypical manipulative tantrum—the faking of autistic or epileptic distress, knowing (but not knowing why, which is we consider it tolerable for children) that adults find it extremely unsettling—in a different light, and I understand the private sector (where adult tantrums often take the form of executives’ pet “initiatives”) even better. Weaponized fake autism is all over the corporate world. In the most benign and common cases, it is a trained presentation used to create the appearance of earnestness or passion at the expense of formal social acceptability. Those of us who are truly neurodivergent have no choice in the matter, and tend to be below average when it comes to using or mimicking signals of social status and credibility, but there are a lot of cases where autism-adjacent eccentricities are taken as signs of genuineness, harmlessness, and high intelligence. Sam Bankman-Fried—whom I believe to be a psychopath, and not autistic—managed to dupe supposedly seasoned investors at Sequoia Capital this way. I do not believe “everyone’s a little bit autistic”—if by autistic, we mean set apart or disabled by neurodivergence, it is clearly false—but I do recognize that many autistic traits are admirable and could be charismatic in certain contexts. The issue is that some people—in particular, psychopaths, who are as adept at climbing the hierarchies of technology companies as anywhere else, suggesting that we tech people aren’t nearly as smart as we think we are—do use autistic presentations in an attempt to excuse or explain away terrible behavior, which makes life a lot harder for people who do not have the option to present as anything else. The “Asperger’s” that a technology executive probably does not have is not actually causing him to fire people, or to humiliate his underlings with constant status requests, or to commit sexual harassment. He’s just an asshole. All this fake autism in technology executives is only going to make life harder for people with the real thing.
One might accuse me of being not only uncharitable but presumptuous. How do I know that these people whom I accuse of using weaponized fake autism are not actually neurodivergent? Of course, I don’t, but there are strong reasons to believe that I am right. The first is that neurodivergent people, while we can be fantastic public speakers, have a level of social chaos at which our performance drops and, to neurotypicals, we appear scatterbrained or worse. Raising venture capital is not within the span of our natural talents. The second is that, in business, there is often a tell for neurodivergence. I don’t believe this applies in science or technology, but in the corporate world where everything is subjective, neurotypicals want their idea to win, because that’s how one gains a reputation. Neurodivergent people want the best idea to win; if we fight passionately for our own, it’s because we genuinely believe it’s the right answer, and if we are convinced otherwise, we are happy to let someone else’s idea win. What you see, over and over in business, is the former approach—no one cares about getting the right answer, because the penalties for being wrong if one is socially acceptably wrong are so low, and almost everyone in charge gets there by making sure their peculiar version of the exact same idea—the same bike shed, but their chosen color—wins.
The neurodivergent desire to consider as many possibilities as one can, and the curiosity that comes from this, is admirable and often useful; we resist the limitations that come from premature blending of possible worlds. So, too, is the neurotypical efficiency and intuition that their tendencies allow. The failure of neurotypicals is not an individual shortcoming but rather that, as their groups and societies scale, dysfunction seems inevitable. Once we are driven out, what’s left is a corporate office. The upper limit for this seems to be 150 individuals, though dysfunction usually sets in at a tiny fraction of that number. On the other hand, there is no evidence that we as neurodivergent people—we are so different not only from neurotypicals, but from other neurodivergents—would do better, since to my knowledge we have never been the neurosocial majority. The most reasonable conclusion seems to be that, in order to build societies that are actually worth caring about, we need them, but they also need us because, when they drive us out of influence, everyone suffers. At least, these are my thoughts on the matter right now—a candidate reality I consider close to the right one. If I’ve convinced you, feel free to adjust your possible worlds.