kortina.nyc / notes
4 Sep 2018

Social Systems are Computations that Minimize Uncertainty

Chaos dream parade, Paprika.

In Consciousness as Computation, I discussed how you can build up succinct explanations for some of the “most human” behavioral phenomena (like creativity) from the premise that consciousness is just a computational process.

Since then, I’ve read more research that uses a similar computational view of consciousness to explain various psychological disorders as failures in the computations individuals perform when matching perceptual data to prior beliefs.

In this essay, I’ll explore how adopting a similar view of social systems–as computational processes that minimize uncertainty about the behavior of other group members–might lead us to a similar diagnosis of some of our current failures of social and political coordination, namely, as problems of over-fitting our perceptions of others to imperfect world models.

Recap: Consciousness as Entropy Reducing Computation

There’s a well known problem in probability theory or computer science called the multi-armed bandit, in which

a fixed limited set of resources must be allocated between competing (alternative) choices in a way that maximizes their expected gain, when each choice’s properties are only partially known at the time of allocation, and may become better understood as time passes or by allocating resources to the choice.


[At each point in time, a gambler must choose] between “exploitation” of the [slot] machine that has the highest expected payoff and “exploration” to get more information about the expected payoffs of the other machines.

When I say a “computational” description of consciousness, I mean framing life as a multi-armed bandit problem and consciousness as the set of computational processes we run (i) to “explore” various opportunities, developing more accurate predictive models of the world (which “reduce entropy” in perceptual information, ie, minimize uncertainty about the world) and (ii) to “exploit” the information from our predictive models in pursuit of goals and rewards.

In Consciousness as Computation, I shared my excitement for how Jürgen Schmidhuber was able to explain some of the “most human” human values like creativity, art, music, science, humor, etc, using a very minimal definition of consciousness as a computation–specifically, he argued that these phenomena were the result of a reward mechanism for the “explore” process and the way that humans are able to use simulated data to efficiently test hypotheses without running experiments in the physical world.

In The entropic brain, Robin Carhart-Harris explains some of the important functions of the other (“exploit”) mode of consciousness, which active during goal directed cognition in a region of the brain called the “default mode network.”

[T]he mind has evolved (via secondary consciousness upheld by the ego) to process the environment as precisely as possible by finessing its representations of the world so that surprise and uncertainty (i.e., entropy) are minimized.


It is argued that this entropy-suppressing function of the human brain serves to promote realism, foresight, careful reflection and an ability to recognize and overcome wishful and paranoid fantasies.

[The regions of the DMN] are consistently activated during goal-directed cognition.

– Carhart-Harris, The entropic brain [emphasis mine]

Here, Carhart-Harris describes how predictive models of the world that reduce uncertainty help us efficiently make decisions in pursuit of goals. The rest of the paper goes on to discuss research that uses psilocybin treatments to switch the brain from a state of (overly strong) goal directed consciousness into a deep mode of exploratory consciousness.

World Models that Distort Perception

I came across Carhart-Harris’ research via Michael Pollan’s new book on psychedelics, How to Change Your Mind. Much of Pollan’s examines how overly strong predictive models can distort our perception of the world. As I noted in Consciousness as Computation, you can witness this phenomenon during exercises like “drawing with the right side of the brain,” which reveal that much of what feels like raw processing of perceptual information actually involves the application of prior beliefs, abstractions, and inferences. Top down application of priors is both a matter of computational efficiency (reducing the cognitive bandwidth required to process the massive amounts of visual and other sense data we are constantly barraged with) and is also useful for error correction (enabling us to “read” signs that are far away or correct and interpret sentences like “The qiuck brown fox jumped.”).

Pollan surveys the research about a set of disorders in which our prior beliefs cause us to get stuck in a rut of rumination or misperception:

David Kessler, the physician and former head of the FDA, recently published a book called Capture: Unraveling the Mystery of Mental Suffering that makes the case for such an approach. “Capture” is his term for the common mechanism underlying addiction, depression, anxiety, mania, and obsession; in his view, all these disorders involve learned habits of negative thinking and behavior that hijack our attention and trap us in loops of self-reflection.

– Pollan, How to Change Your Mind

He cites several studies in which guided psychedelic sessions help patients suffering from disorders like nicotine addiction, PTSD, depression, and anxiety temporarily return to a mode of consciousness that is more purely perceptual, free of the filters that tightly map perceptual data to prior beliefs. Although a psychedelic session provides only a temporary view of a different state of consciousness, these studies demonstrate remarkable long term results as potential treatments. The hypothesis is that temporarily removing a habitual pattern of thinking can help you understand that a particular set of beliefs or priors is a just pattern constructed by the mind (and that you are free to release this set of priors permanently, if you choose to):

“There are a range of difficulties and pathologies in adults, like depression, that are connected with the phenomenology of rumination, and an excessively narrow, ego-based focus,” says Gopnik, whose research explores the consciousness of children, which she believes bears a similarity to psychedelic consciousness. “You get stuck on the same thing, you can’t escape, you become obsessive, perhaps addictive. It seems plausible to me that psychedelic experience could help get us out of those states, create an opportunity in which the old stories of who we are might be rewritten.”


The psychedelic experience appears to give people a radical new perspective on their own lives, making possible a shift in worldview and priorities that allows them to let go of old habits.

– Pollan, My Adventures With the Trip Doctors

Explaining Social Norms as Entropy Reducing Computation

In The Philosophical Baby, Alison Gopnik (quoted by Pollan above) describes the same “explore” / “exploit” modes of consciousness that Schmidhuber, Carhart-Harris, and Pollan all observe in individuals–she labels them “lantern” and “spotlight” consciousness respectively, noting that children spend more time in the former, adults in the latter.

But she takes this computational framework even further and explores the implications for group behavior: just as heavy reliance on priors during spotlight consciousness helps the individual reduce uncertainty and operate more efficiently toward a goal, groups of actors trying to coordinate towards a shared goal can use rules to reduce uncertainty about the behavior of individuals in the group:

Normative reasoning depends on rules. Making choices is hard—it means weighing all the complex information about what we want and what might happen and then making a single decision. Following rules makes that decision-making process much easier. It also lets me coordinate my decisions right now with my past and future decisions. Instead of calculating each time whether the benefit of doing my exercises outweighs the pleasure of Web surfing on the couch, I have a rule—it’s yoga every Monday, Wednesday, and Friday, and I don’t get dressed until I’m done. Rules also let us coordinate our own decisions with the decisions of other people. When all of us follow the same rule, I can predict how you will make choices and coordinate your choices with mine.

[T]he basic assumptions of learning allow us to make radical changes in our theories of the world, but they protect us from knowledge relativism. We choose theories that lead to good predictions, or rules that lead to good outcomes.

– Gopnik, The Philosophical Baby [emphasis mine]

Just as I found the computational description of consciousness profound when I first encountered it in Schmidhuber’s paper, I find it similarly profound to consider all social coordination as an attempt reduce entropy. Gopnik derives an explanation for social norms and moral behavior from a barebones, computational description of humans as prediction making agents who reduce uncertainty about the world by using a mental representation of the self and of other independent agents whose behavior they also need to predict.

If a computational understanding of consciousness can illuminate the causes of (and potential remedies for) disorders that occur when individuals cling too strongly to priors, it is worth considering what social pathologies result from the equivalent failure mode, namely, when entrenched or outdated rules lead to more harm than good.

Let’s consider a few social frameworks, or models for predicting group behavior, and some potential failure modes for each.

Capitalism as a Predictive Model

Perhaps one of the most robust frameworks ever developed to help people model and predict the behavior of everyone else they interact with is capitalism. It is really quite incredible how much coordination, allocation of resources, distribution, and innovation result from a relatively small number of basic principles (voluntary exchange, a price system, competitive markets, private property, wage labor).

What makes capitalism such a powerful model is that the basic mechanics are pretty easy for an individual to learn and use to govern their interactions with other actors within the system. There is built in accommodation for diverse perspectives–in fact, one of the core principles of exchange is modeling how your partner values to world in order to set the price of exchange. Price gives everyone a common mechanism for communicating their values.

While the great power of modeling other humans as self-interested agents is that it can be used to predict anyone’s behavior, independent of what their actual values are, this ability to accommodate an infinite number of diverse perspectives is also a limitation, computationally. This model can tell you with very high certainty that the person you next encounter will be acting to maximize what they value, but it does not give you deep precision as to what exactly it is they value. You can have fairly high confidence that basics like food and shelter will be relatively high on everyone’s list, but there’s a significant chance ‘curveball’ values–perhaps certain religious beliefs–may be even stronger predictors of someone’s behavior than the desire to satisfy their basic needs.

There’s another interesting breakdown of capitalism as a predictive model that I think is happening. Capitalism can enable extremely complex optimizations and coordination between a huge number of diverse participants that historically have not been possible through central planning. The very fact that these optimizations were not possible through central planning points to their inscrutability–the systemic behavior can grow beyond the ability of a single individual to plan or even understand.

Participants are more likely to fear or distrust the system when they cannot understand it, breaking down their confidence that their individual action will actually achieve the desired ends. If actors doubt their own agency, and imagine that other actors also doubt their own agency, then the predictive model completely breaks down, as you would no longer have confidence that others would indeed act to further their values in the ways you’d expect.

As the number of diverse actors whose behavior affects your economic opportunities, your political liberties, and your cultural values increases, the complexity of ‘your’ network increases–and, it becomes harder to reason about the world and make decisions towards achieving your goals in a larger, more connected world. As the scale and complexity of the global capitalist system grows, more and more often events opaque to each individual (housing market failures, changes in labor supply on another continent) will both impact outcomes for each individual and undermine each individual’s confidence in their own agency, weakening the predictive model for everyone.

Tribalism as a Predictive Model

As people lose confidence in capitalism as a predictive model, I think it is a natural reactionary phenomenon for them to look to a very different kind of model for predicting behavior. Where capitalism provides a high confidence, low precision model for predicting human behavior, tribalism provides a low confidence, high precision model for predicting human behavior.

The tribalist strategy for reducing complexity and uncertainty in a network of seven billion independent actors is to categorize these actors into two groups: “us” and “them.” There are a billion ways to draw the line (citizen vs foreigner, liberal vs conservative, man vs woman, black vs white, straight vs lgbt, etc), but no matter which way you draw the line, the strategy is the same: you can’t ever hope to develop precise models for predicting the behavior of each individual (perhaps any single individual can hold up to 150 precise models for other individuals in their brain?), but you can develop a pretty accurate model for predicting your own behavior. By dividing the world into two groups, those that share your very specific model and those that use an effectively opposite model, you can make very specific predictions about every other actor in the world. The problem is that while these predictions are very precise (and their precision is attractive because it feels like it reduces a ton of uncertainty), the model is not a very strong fit for most actors either as “us”, since most actors don’t share all of the sundry attributes, context, and experiences that make you uniquely you, or as “them”, since the generic opposite-of-you model hardly explains what makes every other person uniquely them. The only way tribal reductionism achieves an actual (vs perceived) reduction in uncertainty is by enforcing compliance with your “us” tribe’s social norms (via censorship, rule of law, military force) or eliminating all “them” tribe members.

Reducing Entropy via Cultural Homogeneity

Top down enforced compliance with an arbitrary tribe’s set of social norms is one way we see humans attempt to bend the world to fit a predictive model. Interestingly, there is some evidence of a similar drive towards a smaller possibility space emerging in a more bottoms up way within the capitalist system.

You might have expected an explosion of diversity as the world becomes more connected through efficient transmission of text, photos, and videos on the internet and the declining cost of international travel. More exposure to diverse cultural ideas, and then the remixing of this broader set of inputs, could exponentially increase the possibility space.

Instead, I think there’s strong evidence of contraction in taste and accelerating cultural homogeneity. Philly feels a lot like Seattle… and Austin and San Francisco and Portland, etc. The millennial apartment in Venice Beach looks a lot like the one in Williamsburg… and in Shoreditch and Berlin. You can find a Starbucks pretty much anywhere in the world, and you can find the hipster, unbranded, barista cafe in pretty much any affluent neighborhood.

Consumers, it turns out, do not value expression of their uniqueness as much as they value the ability to signal belonging to (and participation in) a social group. Belonging to a group whose values you understand, whose members have behavior you can predict, is a great way to reduce entropy.

Problems for the Attention Economy

On the one hand, people desire homogeneity and a strong match between a their own behavioral model and all the other people they encounter in the world. On the other hand, people crave novelty (for this is part of the reward circuitry that helps us develop new and better models of the world, cf. Consciousness as Computation). In a highly efficient attention economy–which is how I would describe both the media ecosystem we use to share cultural ideas and the democratic political system governing many nations in the world today–the desire for both familiarity and novelty can create a kind of toxic feedback loop.

It’s difficult to catch someone’s attention with stories about the familiar (and good) parts of our everyday lives, because they are banal compared to the excitement of a myth about a hero journeying out to slay dragons.

So, as a storyteller (or politician), you have two options.

You can try to inject the everyday / domestic / familiar with the thrill and novelty of the avant-garde, which is admirable among intellectual circles, but probably doesn’t get massive distribution.

Or, you can publish ever more novel–and more extreme–content that reinforces a familiar point of view. This, I think, is the route mass media and politics have gone, with a new story every hour that catches your eye, not because it challenges your model of the world and asks you to consider an alternative point of view, but because it’s an extreme and surprising instance that reinforces the tribal beliefs you already hold (either about “us,” or, more likely, about “them”).

This feedback loop seems like the worst of all possible strategies, as it lacks the stability of a more moderate homogeneity (which would be a local maxima, but at the very least, a maxima) or the useful exploration we might achieve by remixing of diverse points of view. Instead, it narrows the focus of every spotlight, driving people further and further apart, making them less likely to learn from and compromise with anyone with a different model of the world.

Rational Discourse and Empathy

As I mentioned above, one entropy reducing ‘solution’ would be for a single group to use military force to compel all non-members to comply with a single belief system. While this would achieve the desired effect of making it easier to predict the behavior of all other actors, it seems like a high price to pay, and is almost certainly suboptimal, as it would essentially cut off the exploration of other predictive models of the world. (It would be the same sort of problem that happens when an individual gets trapped by in a bad mental pattern, with an overly strong prior belief (mis)governing their perception of the world.)

Rational discourse is the enlightenment solution to this problem of myopia, and, given enough time and the willingness of parties with differing points of view to participate, seems to work pretty well. But given the extremist tendencies of our current attention economy, we may need a different strategy altogether for getting groups with different beliefs to willingly enter into any sort of civil discourse.

Gopnik suggests that empathy is one such tool helping groups break out of entrenched or extreme beliefs:

Rules allow us to perform complex, coordinated behaviors—they let us help other people in new and powerful ways. But intimate, emotional empathy is a force that can change even the most entrenched rules. If we discover that a rule leads to harm rather than good we can reject that rule. This is especially true if we experience that harm in the rich, intimate way that comes from interacting face-to-face with a real person in real life.

Often a return to the intimate empathy of infancy—that immediate sense of how other people feel—can be the most powerful way to change what people do. For example, we dehumanize people in the “out-group”—people who are not like us. This impulse is deep-seated and very difficult to overturn completely. One of the best ways to change it is to actually become intimate with the out-group—to recognize that those people are actually like me. People who come to know someone well who is openly gay are much more likely to support gay rights. Individual stories are powerful agents of moral change—often more powerful than rational arguments.

– Gopnik, The Philosophical Baby

When she says that empathy is “more powerful than rational arguments,” Gopnik is arguing that rational discourse is not possible without some sort of empathy as a prerequisite. The idea that group coordination is possible at all, that we can come to agreements and compromises with other independent actors, is based an understanding that other people perceive the world and act within in as we do. Although each person’s experience, context, behavior, and goals are different, there is some sort of shared experience common to everyone upon which we can build up methods of negotiation and attempt to align shared goals around.


As I conclude this essay, I fear the takeaway amounts to something like, “if we could all just see that we are all just computational processes, we’d recognize our shared humanity and resolve all of our differences.” I hope I illustrated in the section on the attention economy that I have no illusions about that message gaining any sort of distribution.

I suppose the takeaway is this: my own interest and excitement in learning about how the brain processes information is rooted in self-skepticism, a fear that the misperceptions resulting from my specific context and set of prior beliefs are preventing me from getting fully in sync with others.

When experiences–like drawing with the right side of the brain, the psychedelic experience, etc–call your attention to the fact that even what feels like objective / raw perception is actually the result of using prior beliefs to filter data and make inferences, this can be an incredibly important reminder of the subjectivity of experience, which is ultimately at the root of most miscommunication and social conflict.

My hope is that recognizing that this subjectivity is just a mechanism our brains use to process perceptual information will (i) enable us to have more empathy when we encounter individuals or groups with beliefs different from our own, (ii) help us remain open to the idea we are capable of updating even our most ingrained models of the world, and (iii) encourage us to develop a more broadly shared model of the world through rational discourse; such a more accurate shared model of the world should reduce uncertainty for everyone and help each person better achieve their goals, whatever they may be.

Tweet Like andrew.kortina@gmail.com