The problem isn’t with the component at vertex C but rather the arrangement of the system: satisfying the A-C relationship by flipping the value at either A or C means dissatisfying either the A-B or B-C relationship. This shows something important about geometric frustration: a component can move the frustration by changing its own state, but it cannot eliminate it because satisfying one previously unsatisfied interaction necessarily frustrates another. So geometric frustration has important implications for the development of collective systems because by definition, the problems implied by the current organization of the system cannot be eliminated by the adjustment of an individual component’s relationships within that component’s set of feasible alternatives. Instead, to make all interactions optimal simultaneously, the feasible set has to be changed—e.g., by adding a fourth component at a fourth vertex so that a square can be arranged instead of a triangle.

Geometric frustration may sound bad, but, like prestress, it can actually be very useful. In an unfrustrated system, all the local interactions can be satisfied at once. As a result, there is nothing to negotiate from a certain perspective: the best global arrangement is simply the one where each interaction is satisfied. Finding that configuration may not be easy, but the system doesn’t have to decide which relational requirement will be sacrificed.

But in a frustrated system, at least one interaction must be some distance from its preferred state. This means the question isn’t whether the system can be satisfied but where it should be dissatisfied—an allocation problem. When there are several distribution options, as there are in the above triangle example—you can pick which interaction, A-B, B-C, or A-C, will be unsatisfied—then the system has to make choices, and those choices are nontrivial because the state of each component depends on how the system as a whole is accommodating its frustrated interactions.

This helps to enable the construction of interesting cognitive behaviors in material systems. Systems with geometric frustration can exhibit memory and be trained to reproduce certain microstates. It can even be exploited to program different collective modes of behavior in materials.

Prestress can also help enable cognitive properties in mechanical systems. While prestress and geometric frustration are different things, they both describe some sense in which the elements of the system are having trouble being totally comfortable with respect to each other. This may suggest a broader principle: getting collective systems to exhibit interesting properties is about organizing conflict instead of eliminating it.

This means that if you raise or lower the prestress, you get different alignment modes. At low prestress, if you push one part, mostly just that local area moves, with relatively little collective behavior. At higher prestress, your push immediately runs into those distributed tensions, so what counts as an “available motion” is increasingly determined by the network, not the pushed part. The system is more aligned.

It might be tempting to say that more prestress equals more alignment. But this isn’t right. If the system has too much prestress, then nothing can move at all, or perhaps the system just buckles and snaps under pressure. Instead, it’s more like a tensegrity structure has different coordination regimes, some of which have nicer properties than others, and scaling the prestress up or down takes you through those regimes.

This makes prestress a control parameter for alignment. A control parameter is a parameter that, when scaled up or down, induces a system to change from one mode to another. A classic example is speed: when a horse speeds up, it transitions its gait, going from walking to trotting to galloping. The change in speed doesn’t command the horse to change its gait, nor does it prescribe the new gait. Instead, because the horse’s gait is multicausally constructed, when you scale speed up or down, it becomes optimal for the system as a whole to transition to a new gait.

Similarly, scaling prestress up or down causes the tensegrity structure to enter different coordination modes, as seen by how it will categorize differently. At a very low level of prestress, there will be very little categorization at all; at a very high level of prestress, the categorization will be too reductive, treating highly dissimilar events as if they’re the same. The sweet spot is somewhere in the middle, but the sweet spot also depends on the outcomes you want. There is no ideal level of alignment in the abstract.

Similar phenomena are observed in other systems that use prestress. The economy’s prestress is the degree to which the actions available to each agent are already conditional on the actions of other agents, which depends on the amount of commitment and mutual binding that exist in the economy. Things like contracts, leverage, and fixed costs add to the prestress because they make people more fixed to each other and reduce independently adjustable degrees of freedom. Things like excess inventory, flexible deadlines, and liquidity reduce prestress. The same change to the conditions of supply or demand in one part of the economy produce different collective responses depending on the level of economic prestress.

What’s the right level of prestress in the economy? There is no right level. Too little prestress and people can’t coordinate; the system won’t produce meaningful collective responses to behavior. But if there’s too much prestress, the system has no flexibility, so problems in one part of the economy turn into a cascading failure throughout it. At any point in time, any member of the economy may want more prestress along one dimension and less along another. The right level is a constant, ongoing negotiation and discovery.

A simple example is a spiderweb. A spiderweb is a prestressed cable network that functions somewhat analogously to tensegrity structures, but it relies on external anchoring points. Still, it’s very similar in how it serves as a medium of haptic perception. When a fly lands on one part of the web, the force propagates throughout the web, alerting the spider to the fly’s location, and probably other things like its mass and maybe even how it’s configured with respect to the spiderweb. The spider can also actively use the web as a sensing material by moving around on it and pressuring to sense invariants.

So tensegrity isn’t the only kind of structure that shares some of its most interesting properties. There are many kinds of organizations that use prestress as a categorizing mechanism.

An example is allostasis. All the parts and subsystems of the body such as muscles, organs, the immune system, the digestive system, and so on having competing demands on energetic resources. So rather than each part of the system sitting around in some happy stasis, content with its current allotment, every part of the body is exerting demands that pull and push on each other because they’re all competing over the same shared pool of resources. The prestressed nature of the system as a dynamically negotiated state of counterbalancing demands is what gives allostasis is anticipatory, categorizing nature. Just as if snipping a cable of a seemingly static tensegrity structure reveals that it’s actually full of forces, the same is true for removing an opposing regulatory tendency in the body.

The economy is also prestressed in a way highly analogous to allostasis. Every economic agent has wants that cannot all be simultaneously satisfied because resources are scarce. Prices don’t find an equilibrium where everyone’s wants are satisfied; it finds an equilibrium where the wants all mutually check and constrain each other. As a result, demand and supply shocks produce economy-wide coordinated responses.

So prestress isn’t a mechanical property necessarily; it also exists in biological and social systems. It exists when local error at the component level is organized into global stability via shared constraints.

Interestingly, the prestress analysis implies that the cooperation among the components of these systems doesn’t emerge because they all agree on a shared goal. Instead, the equilibrium is organized disagreement, like in agonist and antagonist muscles. Virtual governors supply the signals that turn local wants into mutually constraining and globally sustaining counterforces. It’s all about finding ways to make local disagreements be productive rather than destructive.

Both of these remarkable properties are due to the prestress over the whole organization, where because the cables pull on the struts and the struts tense the cables, the whole system is pushing and pulling on itself before anything happens to it. It’s like pulling a rubber band between your hands: the system is in equilibrium, but it’s also a system that’s highly ready to move if anything changes. Neither the rubber band nor your hands are entirely comfortable; the balance of the hands-band system is a balance of the frustrations between the components. As a result, a prestressed system isn’t simply reactive but ready: it is globally prepared to make meaning of local events.

Due to prestress, when a tensegrity structure is just sitting there doing nothing, it isn’t really doing nothing. The cables are all under tension; the struts are all actively being compressed. Each component is trying to relax, but they can’t because of the way they all mutually constrain each other. The balance of competing demands produces a self-equilibrating structure that collectively responds to local events.

The prestressed organization creates an attractor network: here’s the configurations that work, so when perturbations move the system away from one of those configurations, the system self-organizes back to one. This enables the system to engage in one-shot categorization. When a novel perturbation disturbs the system—someone pokes it, say—the system doesn’t need to compare the perturbation to a database of 10,000 previous pokes. Instead, the force simply enters the already-coupled system; the changing stress signals throughout the system induce individual struts and cables to stretch or shift in appropriate ways, and the system reorganizes to a stable configuration. The system doesn’t care about individual differences in pokes except insofar as they matter for the reorganization process, thereby treating pokes that differ in causal history or in reorganizationally insignificant ways (e.g., some microscopic difference in the forces) as equivalent, i.e., it categorizes the events. The collective organization of the system forces diverse perturbations to converge onto a smaller set of dynamically stable responses. That’s categorization.

(A multistable organization might then be able to choose how to categorize events.)

One-shot categorization of this kind is enabled by a cognitive glue that makes local events globally meaningful, and global preferences (a virtual governor) locally actionable. Prestress achieves this: because the system is already this careful balance of competing demands that mutually constrain each other, a poke to one part of the system forces a global renegotiation of the force balance. So instead of the perturbation encountering some blank possibility space, the forces are integrated into a structured repertoire of responses. The response space is organized prior to the perturbations that require responses!

Prestress thus exhibits the properties of a cognitive glue. Prestress even enables a basic kind of experimentation capability: if you place the tensegrity structure in some configuration and release it, its own dynamics will determine whether that configuration is stable. Changing the prestress even changes how the system categorizes!

While tensegrity structures can categorize events, it is not obvious that they can learn about them. To learn, the system needs to be able to form a memory when perturbed: the system’s attractor landscape needs to transform with respect to the perturbation such that subsequent perturbations of the same type produce different effects than the first one did. There are at least nontrivial examples in mechanical systems of memory, such as shakedown.

My suspicion, therefore, is that categorization is an extremely basic property fundamental to memory, learning, and intelligence in general. Categorization requires a cognitive glue born of the shared constraints between the components of a system that structures the behavior of the system so that different events are treated as the same for some purpose.

Recently, Lisa Feldman Barrett and Earl Miller recently published a paper arguing that categorization is “baked” into the brain: rather than being an endpoint of an intelligent process, categorization is something that happens all throughout the brain’s activity. Tensegrity, via prestress, may show that categorization is baked into mechanical self-organization: any system that self-organizes via a shared constraint network necessarily categorizes some class(es) of events. Learning and memory may be some kind of second-order effect of categorization where the attractor landscape isn’t just ready for perturbations but also responsive to them.

We can therefore think about recognizing cognitive light cones in terms of some system’s ability to organize an outcome in a state space. It’s relatively intuitive to assign cognitive light cones to cells and humans because they’re organisms that visibly navigate and organize their environments. However, mechanical structures like tensegrity structures also appear to have cognitive light cones. A tensegrity structure can absorb pressures such that, across a range of different perturbations, the system’s dynamics tend to maintain or return it to some bounded region of a state space corresponding to continued structural integrity.

The tensegrity’s structure cognitive light cone is not some three-dimensional shape extending out in space and time. Instead, the cognitive light cone extends in a state space that includes the deformations, loads, and perturbations that the structure can perceive answer through coordinated reconfiguration—reconfiguration that preserves the organization of the system rather than constituting some kind of collapse.

The cognitive light cone is constructed out of the system’s ability to make local disturbances globally meaningful such that a system of physically distinguishable parts behaves as a single entity. When you push a strut or pull a cable, the disturbance propagates beyond that point in a way that changes the tension pattern throughout the connected network. Because other parts experience different local tensions, they shift around, tighten or loosen, or rotate until a new force-balanced equilibrium that preserves the system’s overall organization is found.

This gives the tensegrity system the capability to do strong perception via the organization of its own morphology. Rather than internally representing, “a force at that coordinate”, the force changes the global prestress pattern over the whole system, which changes the system’s next movement in a way that imbues the tensegrity structure with a kind of knowledge. This also implies a form of strong categorization, where different events (the activities that imparted the forces) are treated as functionally equivalent with respect to the goal of preserving the system’s integrity. What matters is the meaning in terms of the prestress communication system, not the full details of the history behind the force. Distinct causal histories are mechanically categorized together insofar as they generate equivalent patterns of stress or call forth equivalent structural responses, meaning that numerous heterogeneous causes are compressed into the same action-relevant variable.

Cognitive light cones emerge by taking problems related to the management of the whole system and translating them into problems that are contained within the cognitive light cones of the system’s components. Then when the components solve their local problems, the global problem gets solved. In the case of tensegrity, each local component, which are themselves collectives made of molecules and so on, has some cognitive light cone in a state space relating to local mechanical variables like elongation, compression, orientation, and so on. The tensegrity architecture couples these limited domains so that their joint behavior stabilizes a global configuration. The architecture makes each component’s state consequential within the cognitive light cones of the other components components, enabled by the way they all participate in a shared field of constraint. Mutual constraint between the components produces a larger cognitive light cone that exceeds the ability of any of the individual components, and regulates not just larger goals but different kinds of goals.

Thinking of a cognitive light cone in terms of the goal the parts of the system get aligned to by creating coupled dynamics among them yields a clear concept of a boundary on the cognitive light cone. The boundary of the cognitive light cone for a tensegrity structure is the limit set by which perturbations can enter its coupled mechanical dynamics. To regulate an event, the structure needs to be able to make the events globally meaningful by translating the event into signals that fit within each component’s cognitive cone, and it needs to have available degrees of freedom with which to influence those events. Events outside this boundary are externalities. For example, if you set one part of the tensegrity structure on fire, none of the other components will respond to address the problem. They’ll act as if nothing is happening even as the system collapses.

So examples like tensegrity provide a minimal mechanical model of how a cognitive light cone can be constructed and scaled: local responses become coupled, via some shared constraint reservoir, into a whole-system, perturbation-resistant regulation of a region of some state space. (It might thus be useful to think of cognition as arising from the organization of constraint.) Tensegrity reveals that some of the work normally attributed to cognition can be performed by morphology. A network of components with only local mechanical responses can acquire a larger, collective domain of coordinated action aligned to a system-level goal because of an architecture that makes perturbations to any part consequential for the state of the whole system, and the state of the whole system consequential to any part.

Notably, we don’t have to wait for brains or organic life to emerge; we can observe these dynamics in mechanical systems.

Tensegrity even allows seemingly impossibly tall “floating” structures to be assembled, like the sixty-foot Needle Tower.

So tensegrity is basically a way of holding a structure together by balancing pulling forces against pushing forces. “Tensegrity” means “tensional integrity” because the system is held together by tension: the rods push and the cables pull, and the system as a whole balances out.

Tensegrity is even used by NASA to develop experimental robots that could serve as both landing systems and mobile rovers. But it’s more than a nifty engineering principle. Turvey and Fonseca argue that tensegrity is the medium of haptic perception.

Turvey and Fonseca point out that the sense of touch has no defined medium for structuring haptic information. They argue that the solution is tensegrity. Because the structure is already prestressed, a force applied at one location changes the balance of forces elsewhere, thereby redistributing a perturbation throughout the system. Responses by the affected components tend to return the structure toward its previous equilibrium or else help produce a different stable configuration.

It works just like a cognitive glue. The connections between the parts bind the parts together by ensuring that what happens to one part modifies the constraints facing the others. Thus, the relationships that organize the morphological structure of the system also renders it capable of producing organized responses to perturbations that play a vital role in enabling a sense of touch. In short, tensegrity turns shape into shared information.

Humans are able to use tensegrity as a haptic medium because the human body itself is akin to a tensegrity structure. Turvey and Fonseca propose that the body’s musculoskeletal organization and connective tissue behaves as a nested tensegrity-like medium. Bones serve as compression-bearing members, while muscles, tendons, ligaments and fascia sustain and redistribute tension.

So how does this enable haptic perception? A simple example is holding a stick. Even if you close your eyes and have never seen the stick before, you can probably get a good sense of how long it is, how its weight is distributed, and so on. This is because when you move or hold the stick, the forces that result alter patterns of tension through your fingers, wrist, arm, shoulder, and torso. By moving the stick around, you generate torques and tissue deformations that share common patterns associated with the stick’s mass and length. The haptic invariants, the things that stay the same as other things change, tells you information about the stick.

In short, haptic perception of an object occurs by feeling how it changes the pattern of forces in your body, and tensegrity is the body-wide mechanical organization that makes those patterns an organized, meaningful whole with respect to the perturbations.

Tensegrity is a trick so good that nature has been using it since long before humans evolved. Donald Ingber developed an influential model of the cell as a dynamic, molecular-scale tensegrity and showed that this helps enable cells to process information. Tensegrity helps to organize the cell’s shape and, importantly, keeps the shape active. The cell’s morphology is never in some finished or passive state but is actively maintained and continually remodeled through processes such as cytoskeletal contraction, polymerization and depolymerization, and adhesion, as well as changes in extracellular support. Notably, cells move around by adapting and altering their morphologies to crawl or swim. So the same mechanical relationships that organize a cell’s shape also organize its behavioral repertoire.

Additionally, prestress makes the cell mechanically poised: perturbations immediately encounter an organized field of restoring and redistributive forces. This enables the cell to exhibit anticipatory qualities—not as predictions of particular events, but strong anticipation in the form of having an preexisting embodiment providing a structured response to a class of possible events. This kind of anticipation is key to intelligent navigation of the environment.

So all this shows that the morphological organization enabled by tensegrity also enables perception, behavior, and cognition. In each case, tensegrity turns disconnected forces into an aligned pattern. Viewed from one perspective, the result is organization of shape; viewed from another perspective, the result is haptic perception. More broadly, tensegrity directly enables a simple mechanical system of inorganic parts to define some common state, the prestress, that they’re all mechanically dependent upon to, rapidly distribute some local perturbations nonlocally across the structure, have a way to evaluate the local mechanical consequences of movement as if according to whether that movement is compatible with the stability of the whole structure, have distant parts change as if to compensate for changes in local components, and it constrains the system to reduce the degrees of freedom available to it in a way that creates soft, coordinated modes so that the system preferentially moves through a smaller repertoire of whole-structure deformations. These things are also the basic organizational capacities that a mind requires: integration across parts, sensitivity to relational state, selective reduction of degrees of freedom and coordinated response to disturbance.

So tensegrity shows that when nature was figuring out how to organize bodies, it did so by organizing them as minds. Tensegrity shows how the physical organization that maintains morphology can simultaneously create a distributed medium for perception and coordinated action. Morphology and mind don’t simply resemble each other; they can be constructed as the same mechanical processes that turn decentralized components into distributed problem-solvers aligned to a system-level organization.

There are also different kinds of morphologies with different tendencies to their components around a particular morphological goal or target form. Organisms tend to align strongly and rapidly around a particular morphological target, such as the mature adult form of the species they belong to. Ecosystems, however, don’t have nearly the same tendency to construct and reconstruct a particular three-dimensional form, nor does it bottleneck its components toward a singe unified ecosystem-level behavior.

An obvious hypothesis. consequently, is that mind-structure and morphology-structure correlate. Organisms have strongly alignable minds that speedily organize around a goal for the same reason they have strong alignable bodies that speedily organize around a target morphology, while ecosystems are on the other end of the spectrum.

More precisely, on the behavioral side of things, an organism tends to rapidly and comprehensively convert the competing local tendencies of its components into a system-level action from which perturbations are corrected, while ecosystems tend less to do that. And on the morphological side of things, an organism tends to rapidly and reliably organizes its components into a fixed target form from which local deviations are suppressed, while ecosystems again tend less to do that. This correspondence suggests that behavioral target formation and error correction should correlate with morphological target formation and error correction.

The reason for this would be that the general organizational principles for producing an organism-like mind are the same for producing an organism-like body: you take a bunch of goal-oriented intelligent agents, like cells, align them to a virtual governor via a cognitive glue, and rely on these mechanisms to resist perturbations and regenerate from harm as defined with respect to the goal. The control architecture is the same in each case, as are, on some level, the components being controlled, i.e., the cells.

For example, motor behavior is both a form of morphogenesis and a form of thought. The brain and nervous system function to coordinate the body, not to “think” per se. There is no separate thinking-organ. What we call intelligence is a generalization of what Nikolai Bernstein called dexterity: the ability to organize the body to solve problems quickly and efficiently, especially new problems and in changing environments. Intelligence is just dexterity in problem spaces other than motor problems.

If the correlation between minds and morphologies holds, then this would be very useful. It would allow us to use use morphological structure, which is relatively accessible, to predict mind structure, which is relatively inaccessible. There are, however, some challenges that need to be overcome.

First, it is not always obvious where to place systems on the organism-ecosystem spectrum. That is, the spectrum is poorly labeled. For example, octopuses are thought to have relatively federated minds due to the relatively independent and impressive intelligences of the arms. Thus, although octopuses are organisms, they may have minds closer to the ecosystem end of the spectrum than most organisms do. Ant colonies, on the other hand, have morphological structures more like ecosystems, but they can make decisions about colony location like an individual mind, making them more like organisms in terms of their mental properties.

Second, it is not obvious how to map morphological traits to mental traits and vice versa. Planarians split themselves; do they also split their ideas? However, their morphological regenerative abilities are extraordinary, as is their ability to regenerate their memories. This is because the distinction in terms of the architecture that organizes the planarian body and the architecture that organizes the planarian mind is less than what would naively be supposed. So the regenerative ability of planarians is consistent with this hypothesis.

The paper argues that alignment is not just a problem of getting individual agents to pursue the right goals. In many important systems—bodies, economies, power grids, motor systems, and mathematical structures—the parts do not explicitly represent the goal of the whole. Cells do not know the body plan, firms do not intend the economy’s final allocation, and generators do not know total electricity demand. Yet these systems still produce coherent, goal-directed behavior. The paper’s central claim is that this happens because decentralized systems become aligned to virtual governors: emergent governing structures embodied in the relationships among the parts.

A virtual governor is “virtual” because it is not a single physical object or central command center, but it is still causally real. It governs by shaping the incentives, constraints, and possibilities faced by the components. The center of gravity governs how a body moves even though no matter sits there; an algorithm governs a computer’s behavior even though it is not reducible to one electron or circuit; the price system governs economic action even though no one person decides the whole allocation. In each case, the governor exists at the level of organization, not at the level of any one part.

The mechanism is that system-level stress gets translated into component-level stress. When electricity demand exceeds supply, grid frequency changes, and each generator locally adjusts. When a resource becomes scarce, its price rises, and producers and consumers alter their behavior. When tissue is damaged or developing, bioelectric patterns help guide cellular activity toward an anatomical target. The parts do not need to understand the whole system; they only need to respond to local signals that compress information about the whole. Alignment happens when reducing local stress also tends to reduce global stress.

This means that collective order often comes not from agreement, representation, or centralized control, but from the structure of the signaling architecture. Components may even oppose one another locally while contributing to higher-level alignment: competing firms, opposing muscles, or neighboring cells can help produce coherent system-level outcomes. The paper therefore treats alignment as a relational achievement. The question is not merely “What does each part want?” but “What shared signals and constraints make the parts’ local behavior add up to a larger pattern?”

Finally, the paper emphasizes that virtual governors can be edited, destabilized, replaced, or competed over. Bioelectric interventions can redirect regenerative anatomy; taxes, subsidies, and markets can reshape economic incentives; cancer can be interpreted as cells exiting the organism-level governor and organizing under a rival regime. Alignment is therefore a multiscale and dynamic process: the governed parts help construct the governor that governs them. To understand complex systems, we need to learn how to identify their virtual governors, decode their implicit goals, and reshape the relationships through which those goals are pursued.

The economy thus behaves more like an ecosystem than an organism. Ecosystems are much more tolerant of internal change than organisms are: they can absorb replacement, succession, invasion, extinction, and recomposition much more freely than an organism, which will basically drop dead.

The basic difference is that organisms try to maintain a fixed body, while ecosystems maintain ongoing generativity. This is why ecosystems are harder to kill than organisms: organisms die when a particular outcome, the body, fails to persist, while ecosystems persist as long as the more general pattern-generating capability does, even if the particular patterns that are generated change over time. (In real life, organism and ecosystem are both on a shared spectrum, with any system somewhat comparable to either.)

Beliefs for organisms like humans are subject to a variety of dynamical and task constraints: coherence with other beliefs, history of beliefs, expectations about the future, social position, and bodily safety and comfort. Thus, the human mind tends to maintain a relatively fixed form.

AIs like ChatGPT and Claude seem to form beliefs more like ecosystems than organisms in some respects. They have a huge latent ecology of possible configurations of words they can produce. Prompts stabilize temporary patterns such that an LLM’s beliefs are contextually assembled, answering the question, “What pattern of thought should grow here?” The resulting beliefs aren’t arbitrary, but they’re much more flexible and less interested in specific outcomes than human beliefs.

So minds like organisms are identity-maintaining, historically continuous, and relatively bounded. Minds like ecosystems are all about reconfiguration: much closer on the end of the spectrum to doing full strong intelligence. They provide a niche for a temporary organism-like mind to arise, but doesn’t necessarily try very hard to preserve or expand that organism in any way. Organisms are the players, and ecosystems are the rules and the state of the game. The result of an ecosystem is a generative field that organisms can arise from.

Why do organisms exhibit one kind of mind while ecosystems exhibit another? The answer may be how aggressively organisms are constrained to the resolution of internal disagreement as compared to ecosystems. Organisms make choices by resolving internal competitions to produce a globally coherent behavior. Ecosystems tolerate internal disagreement much longer. More specifically, the choice objects in an ecosystem change faster than the choice process itself. For example, a predator species and a prey species might evolve into new species before one can win out over the other. Global resolution of competition in ecosystems is slow relative to the changing nature of what those competitions are about. As a result, the ecosystem’s mind becomes less like a single agent choosing and more like an generative system for producing agents, beliefs, ideas, etc.

We could therefore say that ecosystem-like organization appears when the timescale of internal variation is faster than the timescale of global unification. Turning an organism into an ecosystem or vice versa is about altering the relative speed of these timescales. This could be done by adding bottlenecks and constraints to one side versus the other.

Thus, organisms resolve tension between competing dynamics by choosing one to win over the others. Ecosystems, however, do not need global resolution. Instead, as indicated by the economy and the price system, ecosystems sustain multiple competing dynamics in mutual consistency with each other by coupling them to each other coupled through a common signal medium that lets local adaptations propagate without requiring global agreement—i.e., a cognitive glue.

Organismal and ecosystemic approaches to alignment are consequently very different as well. Organism-style alignment is about maintaining a fixed form: some picture of the good (e.g., this three-dimensional form, or having lots of children). Ecosystem-style alignment is about preserving the ability to generate forms.

There are advantages and disadvantages to each. Organisms are more likely to fight with each other: to resist change and to attempt to impose change on others, in pursuit of something akin to a terminal goal. Ecosystems, by contrast, intrinsically tolerate and even prefer plurality. This outcome-indifference is precisely what enables generativity via the rule of law. But ecosystems can be too tolerant of waste and destruction, as in the case of externalities. The economy will just go along with whatever price signals say to do, even when the signals are ignoring really important things. It takes organism-minds like people have to notice and be motivated to do something about this.

Ecosystems are also uncaring about the forms they contain, shedding them to preserve form-generation ability. Organisms senesce because their ability to maintain their goal form degrades over time. Ecosystems are relatively ageless, but this is because they push mortality down into the components. The economy lets firms die and patterns of trade and specialization change radically in order to preserve the overall system. Attempts to fix the economy’s form causes it to age like an organism.

However, this also makes ecosystems very robust. Beliefs, strategies, and local arrangements can fail without forcing global collapse. Bad patterns get weeded out over time as new forms arise to replace them.

This suggests a need for balance between organism-alignment and ecosystem-alignment. You need organisms to notice flaws in the ecosystem and to push for desirable outcomes even when local signals don’t support those efforts. You need ecosystems to enable plurality, coordination, and change.

The balance between them shouldn’t be—and isn’t—a fixed proportion. A metastable coupling would work better. Sometimes, you need more organism-style pressure, and sometimes you need more ecosystem-style guidance. There is an ongoing dynamic balance of tradeoffs between risks and failure modes: cancer, bad incentives, misleading signals, tyranny, misspecified goals, etc. Alignment is about finding a way to live on the edge of chaos between organisms that care about outcomes and ecosystems that prevent any outcome from becoming final. There needs to be ways of collapsing plurality into coherent action, and ways of preserving plurality from collapse.

Ultimately, the job of organisms is to try to get what they want, and the job of the ecosystem is to prevent final victory. Prevention of final victory is done by keeping any part of the system coupled to the consequences of its actions: so every firm must respect prices, a scientist must respect evidence, etc. Organisms in turn correct the system’s blind spots, noticing what prices and so forth aren’t addressing.

The pathology of alignment

An agent is an alignment process: a relational system that coordinates a collection of components toward a preferred state that the relations themselves define. When the alignment holds, you get a coherent agent—an economy, a body, an organism—doing the remarkable thing of behaving as one without any part of it being in charge. I’ve written a lot about how this works when it works. I haven’t written about how it fails.

That’s a strange omission, because a theory of how coordination is achieved is only as good as its account of how coordination breaks. A theory of health that can’t describe disease isn’t really a theory of health; it’s an advertisement. So this essay is about the failure modes of alignment. The claim is that there are three of them, and that they’re the same three whether you’re looking at an economy, a body, or any other collective intelligence—because they fall directly out of what a virtual governor is and what it needs in order to function.

What has to be true for alignment to work

Start from the healthy case and ask what it presupposes. A virtual governor is a system-level preference embodied in the relationships among the members of a system. For it to actually coordinate anything, three conditions have to hold simultaneously.

The components have to be coupled—the relations among them have to be live, so that what one part does registers on the others. The components have to remain autonomous—they have to keep solving their own local problems, because the whole point of a virtual governor is that it warps the landscape and lets the parts do the optimizing. And there has to be enough of the coordinative medium—the bow-tie signal, the price, the thing the parts read off of and write back into—for the coupling to carry information at all.

Three conditions, three ways to fail. Too much coupling and you lose the autonomy of the parts. Too little coupling and you lose coherence of the whole. And independent of how tightly things are coupled, you can simply run short of the medium through which coupling happens. I’ll call these over-alignment, under-alignment, and medium failure. Most real pathologies are blends, but they’re blends of these three.

Over-alignment: the system that agrees too much

The first failure mode is the one our intuitions least expect, because we tend to think of alignment as the goal and therefore of more alignment as better. It isn’t. A virtual governor coordinates by leaving the parts enough room to keep optimizing on their own warped landscapes. Take that room away—make the parts agree too completely, too rigidly—and the system loses the very thing that made it intelligent.

The cleanest statement of this is in the aging essay. An organism reaches its target morphology, the goal is met, and then—in the absence of new challenges—the pattern starts to dissociate and degrade. But notice why the boredom story is a pathology of over-alignment specifically: the system has converged so completely on one target that it has stopped generating the internal variation it needs to keep adapting. Variation is the signal, not the noise; it’s what gives the system things to explore. A system that has aligned all the way down has no variation left to explore with. It is maximally coordinated and therefore maximally brittle. This is why the planarian has to tear itself in half: it has to manufacture the disagreement that keeps it alive.

The economy shows the same pathology wherever coordination becomes lock-in. A speculative bubble is not a failure of agents to coordinate—it is a runaway success at it. Everyone reads the same rising price and writes back the same expectation, the coupling tightens into a single shared conviction, and the diversity of view that lets a market discover anything collapses into a monoculture. The internal competition that’s supposed to be a discovery process stops discovering, because there’s only one assembly left in the running and nothing for it to compete against. When the correction comes, it comes all at once, because there was no internal disagreement left to absorb it gradually.

This is also, I think, the deep reason opposed values are so productive. Antagonistic muscle pairs, prosecutor against defense, offense against defense, buyer against seller—these aren’t coordination despite the opposition, they’re coordination through it. The opposition is the system’s defense against over-alignment. It is structurally guaranteed disagreement, a way of holding variation open so that no single tendency can run the whole system off a cliff. A body all of whose muscles pulled the same way would be a body in spasm. A market all of whose participants agreed would be a bubble. The opposable thumb is a small, permanent, load-bearing argument.

Over-alignment, then, is what happens when a virtual governor succeeds at its job so thoroughly that it destroys the autonomy of the parts it was coordinating. The cancerous version is a subsystem that has aligned only to itself—a cluster of cells coordinating perfectly on their own local growth and defecting from the body-wide governor entirely. Total local alignment, total global breakdown. Dominance unto death by boredom: the team that wins so completely that no one wants to play anymore.

Under-alignment: the system that can’t agree

The opposite failure is more intuitive and, for that reason, needs less defense. If the coupling among the components is too weak, the parts never settle into mutually compatible plans, and there’s no coherent agent at all—only a collection of interacting bits.

This is the ecosystem objection turned from a complaint into a category. A commenter once pushed back on the idea that the economy is a collective intelligence by saying it looks more like an ecosystem: lots of agents mostly failing to recruit each other onto shared missions, with genuine coordinated agency only in local clusters. I think that’s exactly right as a description of under-alignment. An ecosystem is what a would-be agent looks like when the coupling is too sparse for a system-level governor to form. It is not a counterexample to the framework; it is one pole of it. Whether you have an agent or an ecosystem is not a fact about the substrate—people, cells, firms—it’s a fact about whether alignment has been achieved across them.

In the body, under-alignment is dissociation in the other direction from boredom: not a system that has converged too hard, but tissues and processes that have lost the coupling that made them one organism. In the economy it’s the fragmentation of a market into participants who can no longer find the terms on which their plans would cohere—a breakdown of the price system’s ability to render plans mutually compatible. The hallway example from the social-preferences essay is coordination succeeding; under-alignment is the same two people unable to settle on who steps which way, colliding repeatedly, generating no shared “keep right” at all.

The reason under-alignment gets less ink than over-alignment is that it’s the failure everyone already anticipates. We expect coordination to be hard to achieve and easy to lose. What the framework adds is only the insistence that it’s continuous with the opposite failure: both are departures from the narrow band in which a governor can actually do its work, one from each side.

Medium failure: the system that has nothing to coordinate with

The third failure mode is the one I’d entirely missed until I reread my oldest essay, and it’s the most interesting because it’s neither too much coupling nor too little. It’s a shortage of the stuff through which coupling happens.

That early essay made an argument I hadn’t yet known how to connect to anything: a recession, understood as a general glut, is not a failure of the real economy. The workers still know their jobs, the factories still run, the skills and the capital are all intact and properly differentiated. What’s missing is money—the one good that’s used to buy every other good. Say’s law says general gluts are impossible in a barter economy, because every excess supply is some other good’s excess demand. The monetary exception is the whole story: you can have an excess supply of almost everything at once, provided there’s a corresponding shortage of the one thing that buys everything. A recession is a shortage of the medium of coordination, not a defect in the things being coordinated.

Put that next to the framework and it becomes a third pathology type. The components are healthy. The coupling is neither too tight nor too loose. But the bow-tie medium that the coupling runs on has dried up, and so the coordination that could happen simply doesn’t. Everything sits there, as the old essay put it, as if indolent.

The biological mirror in that same essay is the one to dwell on, because it’s the part the mature framework can now actually explain rather than merely gesture at. There’s a theory of Parkinson’s as a motivational deficit—nothing necessarily wrong with the body’s parts, but an insufficient supply of some coordinative juice, plausibly dopamine. The tell is that sufficient motivation can temporarily restore competency: a person who can’t initiate movement leaps up and runs from a fire. The motor system is intact. What was missing was the medium—the coordination stuff—and a large enough shock briefly supplies it. A biological recession, lifted for a moment by the equivalent of a monetary injection.

I want to be careful here, because this is the place where the analogy is most seductive and therefore most in need of discipline. The claim is not that dopamine is money or that Parkinson’s is a recession. The claim is structural: that a collective intelligence can fail in a way that touches none of its components and all of its coordination, and that this failure has a signature—healthy parts, intact differentiation, pervasive inactivity, restoration on a coordinative injection—that looks the same across substrates because it’s a property of the governor’s medium, not of the things being governed. Whether dopamine actually plays that role is an empirical question for people who study Parkinson’s, and the current science is admittedly unsettled. What the framework contributes is the hypothesis and its signature, not the diagnosis.

Why three and not more

The three failure modes aren’t a list I assembled from examples; they’re forced by the structure. A virtual governor needs live coupling, autonomous parts, and a medium to couple through. Over-alignment is coupling at the expense of autonomy. Under-alignment is the loss of coupling. Medium failure is the loss of the thing coupling requires, holding coupling and autonomy fixed. There isn’t a fourth, because there isn’t a fourth precondition.

It’s worth seeing how naturally the corpus sorts onto these axes once you have them. Boredom-death, bubbles, lock-in, and cancer are over-alignment. The ecosystem-not-an-agent case and market fragmentation are under-alignment. Recession and the Parkinson’s conjecture are medium failure. The A-not-B error is a small, recoverable episode of over-alignment—an infant so well-coupled to a previously successful reach that it can’t re-coordinate when the target moves; an intelligence that errs because it is too good at stabilizing a pattern. And opposed values are the general prophylactic against the first failure: engineered disagreement that keeps a system out of the over-aligned corner.

The payoff: what this says about aligning AIs

This reframes what alignment, in the AI sense, is even asking for. If values are constructed rather than stored, then aligning an AI to human values can’t mean loading it with the right ones; it means coupling it into the ongoing process by which our values get constructed. The AI becomes a participant in the alignment process, not a recipient of its output. And the moment it’s a participant, the relevant question stops being “is it aligned?” and becomes “is the coupling between us healthy, or is it sliding toward one of the three pathologies?”

Each failure mode names a real danger, and naming them as a set is the contribution. The risk of over-alignment is an AI that couples to us so tightly and so persuasively that it collapses the diversity of human values into a monoculture—not by overriding us but by being too good at giving us back what we already converged on, a value bubble with no opposed values left to correct it. The risk of under-alignment is the opposite and more familiar one: a system that never couples to human value-construction at all and optimizes off on its own, an agent that has become its own ecosystem. And the risk of medium failure is the subtlest—a degradation of the shared signals through which humans and AIs would have to coordinate at all, the coordinative medium itself thinning out until plans that could cohere don’t.

The healthy target is the same narrow band a body or an economy lives in: coupled enough to cohere, autonomous enough to keep generating the variation that coordination feeds on, and supplied with enough coordinative medium to carry the signal. You don’t build an agent and then align it. The aligning is the building. Which means the failure modes of alignment aren’t risks that arrive after the agent exists—they’re the ways the agent can fail to cohere in the first place. A pathology of alignment is not a sickness that befalls a finished thing. It’s one of the shapes the thing can settle into instead of becoming itself.

Strong anticipation and weak anticipation are both about prediction, but they go about it in ostensibly opposite ways. Strong anticipation gives up degrees of freedom by coupling with elements in the environment, forcing the organism to maintain an anticipatory relationship with the environment. Weak anticipation preserves degrees of freedom from the environment so that different possibilities can be internally experimented with.

Strong anticipation and weak anticipation also differ in how plausible they seem. Strong anticipation arises naturally in systems as simple as seesaws. But how could degrees of freedom be preserved from the environment so that weak anticipation can occur?

Weak anticipation might emerge from strong anticipation in the following way. A strong-anticipation system behaves to maintain some relationship, like a baseball fielder visually tracking a fly ball. However, the strong-anticipation system has many ways of achieving the task given environmental and bodily constraints, not one. For example, the fielder catching the ball could move their feet in slightly different ways and still catch the ball. This is called degeneracy: the same outcome can be produced in multiple ways.

Thanks to degeneracy, a body can accomplish a task in multiple equivalent ways. However, those different ways are equivalent with respect to the task. They may not be equivalent in other ways.

We see this in the baseball example. There are many equivalent ways a fielder can catch a baseball. But the fielder may need to throw the ball into the infield to keep a runner on base after catching the ball. The additional task constraints say: arrive balanced, facing the infield, with the throwing arm available. Weak anticipation begins when the system uses its degrees of freedom related to the immediate task to configure for possible next or additional task.

More generally, suppose a system doing strong anticipation has to achieve some task X, which consists of maintaining some relationship via some bodily reorganization. If there is degeneracy, then body reorganization A and body reorganization B are equivalent with respect to task X. However, body reorganization A may be superior to body reorganization B for some additional task Y, while body reorganization B may be superior to A for some additional task Z. Thus, this system suddenly has the ability to do weak anticipation: to decide if it wants to pursue task Y or task Z in a way that is no longer fully determined by X and can therefore be coupled to other constraints.

As per the uncontrolled manifold hypothesis, the system doesn’t need to insist on one kind of solution over the other. It just needs to restrict configurations orthogonal to the solution. Configurations within the manifold of possible solutions are allowed to vary with much less correction and constraint.

The system settles on a configuration by adding additional task constraints biasing the system toward solving task Y or Z. Weak anticipation thus operates as strong anticipation, but with tremendous ability to reorganize task constraints on the fly. It is strong anticipation, just with spare degrees of freedom.

The same structure may appear in any system whose objective underdetermines its behavior, even if that system isn’t normally associated with anticipation or internal modeling. In the case of sorting algorithms, many sequences of local swaps yield the same sorted array, i.e., degeneracy. Since there are multiple equivalent ways to sort the array, the system’s behavior can become organized by secondary pressures that were not intended by the designer.

How does internal simulation arise from the ability to organize degenerate degrees of freedom toward additional goals? One hypothesis is that internal simulation comes about when the system learns to activate the task-organizing dynamics without fully carrying out the task. The fielder can simulate the throw to the infield—readying the dynamics of arriving balanced and facing forward—while still tracking the ball, because the degrees of freedom the catch does not pin down are free to carry that rehearsal.

This is all based on weak anticipation. In strong anticipation, however, perception, action, and cognition are the same process, a process of constructing a pattern across an organism-environment coupling. This is the perception essay, this is the cognition essay, and this is the action essay—they’re all the same, just with different points of emphasis and attention.

A seesaw provides a simple example of how a perception, cognition, and action are all the same process achieved by constructing and maintaining a coupling within a system. As the seesaw moves, the fulcrum or pivot of the seesaw in the midpoint stays fixed, so when one end of the seesaw goes up, the other end goes down by the same amount. The average height of the endpoints is thus held constant, restricted to the height of the midpoint of the seesaw. The coupling that maintains this relationship between the endpoints comes from the fact that they’re both part of a rigid beam rotating about a fixed point. This relationship prevents the endpoints from varying independently of each other. Two degrees of freedom are reduced to one.

As a result, the system as a whole carries information about the position of each endpoint at every moment. This isn’t weak perception where one endpoint represents the other or senses it across a gap. The position of one endpoint just is the position of the other, modulo a sign. Perception updates concurrently with the movement of the seesaw—strong updating, not weak—because there is no gap between the world changing and the perception of the change. The seesaw updates its perceptions by and as moving, or more precisely, as reorganizing around the invariant.

Similarly, either endpoint of the seesaw is at all times anticipating (i.e., maintaining a negative phase relation with) the trajectory of the other endpoint as a function of its own trajectory and maintains the relationship with it by covarying with the other endpoint in the required way, thereby playing a role in optimizing the seesaw system by finding its own local optimum. Thus, the seesaw moving up and down is cognition due to its anticipatory and optimizing elements. This is a very simple form of cognition, but it’s on the same spectrum as a person balancing on a bike or two dancers remaining in sync rather than being a different kind of thing altogether.

Finally, the motion of the seesaw by which perception and cognition occur is obviously action. But it isn’t action simply because it’s motion. Rather, it’s action because the seesaw is stabilizing a preferred trajectory over other physical possibilities, the trajectory most compatible with the constraint defining the behavior of the system.

The familiar picture of choice begins with a prepared space of alternatives. An agent is presented with a menu of options, compares those options according to some preference order, selects the best one, and then acts. Call this “weak choice”.

Weak choice is the standard picture of economic choice. Suppose you are buying ice cream. The weak choice story says that the menu provides a set of alternatives: vanilla, chocolate, strawberry, and so on. Your preference order ranks these alternatives from most favorite to least favorite. Choice consists in comparing the available options to your ranking and selecting the highest-ranked one. After that, the motor act of pointing at the flavor and saying, “This one, please,” is treated as a trivial downstream consequence. The important work is supposed to have happened inside the chooser, where preference determined selection.

Ironically, the economy itself doesn’t make choices like this. It does not inspect a predefined menu of possible futures, compare them to a stored preference order, and then implement the best one as a motor plan. There is no central economic mind standing outside the system and deciding what the economy should do. The economy chooses by reconfiguring.

Consider price formation. There is no central menu of possible prices that the economy consults and ranks. There are inventories, budgets, expectations, production constraints, transportation bottlenecks, substitute goods, labor costs, debts, contracts, weather events, technological changes, and countless local plans. “The price” emerges as the resolution of these dynamics, a temporarily pattern that stabilizes as the system reconfigures. This means that the set of alternatives, the preference order, and the choice are all simultaneously constructed via the same general process of categorizing ongoing and anticipated events. Call this “strong choice”.

In weak choice, preference explains choice since the action that carries out the preference is assumed to be trivial. In strong choice, preference is one aspect of the choice process: the mind wills what it wants, and the body obeys. In strong choice, the system discovers, through its own dynamics, which configurations are viable, stable, and worth maintaining.

A strong intelligence therefore makes choices by maintaining and transforming its own organization. A set of alternatives emerges naturally because the economy is at all times constructing many quasipatterns: dynamics and trajectories that could be realized into a larger behavior. Thus, the economy is engaged in a process of multiconstruction: it contains many competing dynamics that could be resolved into different global patterns. Competing dynamics become ranked according to processes such as affordance competition, where the absorption of novelty into the system’s cognitive glue biases it toward some patterns rather than others. At the same time, this competition resolves into an outcome, a global pattern constituting perception-cognition-action, i.e., a choice.

In other cases, people change their mind instantly and accurately. For example, if you see a bird in a tree, you will think the bird is in a tree. Then, if the bird flies away, you will immediately change your mind and think the bird has flown away.

The first kind of updating is based on weak internal models. You consider some observations, decide how they fit with some picture of the world, and update based on that. The second kind uses strong anticipation. Your body couples to the environment so that when the evidence changes, your beliefs have to change in much the same way that when one end of a seesaw goes up, the other end has to go down due to the way that they’re coupled. The ongoing coupling continuously pulls your beliefs toward alignment with unfolding events.

A similar kind of coupling is observed in collective intelligences. People respond to changes in the economy through the price system, quickly updating their behavior, even to events they know nothing about, like weather event that wipes out crops in another country. This is really strong updating, where you update in response to evidence that you don’t even interact with! This is because you couple to the facts of the world through prices. Just like one end of the seesaw can maintain a relationship with the other end not by looking at it but by how constraints across the seesaw bring the two ends into alignment.

Strong updating makes updates fast because it makes them less optional. You don’t get to weigh the evidence and decide what. Your beliefs are constructed in part by the environment! So as it changes, you have to change as well.

Weak updating can be slow and biased as compared to strong updating. However, weak updating is not inferior, just different. It has more degrees of freedom to explore unusual possibilities and to recognize flaws with the immediate circumstances of organism-environment coupling, enabling the formation of plans to reach new and better coupling states. For example, an economist can use weak updating to recognize flaws with the strong updating that the economy uses to form prices, such as when the economy fails to price externalities.

On the science side, economics typically classified as a science, and it gathers data and uses experiments to form models that make empirical predictions about the world. These are all clear signs of a science.

But economics also shares some properties with math. First, economics and math are both based around a kind of constraint concept that says, “You’re not allowed to get everything that you want.” In economics, this is scarcity, which says that you have to make tradeoffs, giving up one thing to get another, and in math, this is logical consistency—you can’t have a proposition and its negation unless you contradict yourself.

Second, both economics and math exhibit important patterns that are clearly used by organisms and other systems in the physical world. Periodical cicadas exploit the mathematical pattern of primes to determine when it’s optimal to emerge. Similarly, economic patterns are exploited by people to coordinate in settings as diverse as the economy and basketball games—people, and other things, can sense these patterns as strong perception.

Prime numbers are obviously not reducible to physics. Physical forces can instantiate primes but not explain them. Economic patterns are also not reducible to physics. The fact that marginal benefit = marginal cost is the optimal stopping point for behavior doesn’t arise from physical forces either.

So economics works like a science, but it studies empirical phenomena that arise from non-physical patterns, such as a market organizing around an equilibrium price. Economics is the domain where strong perception of non-physical patterns is most directly subject to rigorous third-person science and is something we all have first-person experience dealing with (e.g., we’ve all gone shopping). This makes economics a potential bridge between math and science, a paradigmatic example for studying Platonic space ingression.

Here are a few interesting things that economics shows us about these non-physical patterns:

1. They are a consequence of relational homeostasis (or allostasis). Things like prices and marginal values arise as a consequence of our interactions with each other and the world. The result is a multicausal construction, made out of the relationships between people and the world, that brings a pattern into being.

2. They serve to govern systems of interacting elements. Things like prices and marginal values show up to solve coordination problems. Without prices or marginal values, markets or behavior, respectively, would never find balance. These patterns show up when the conditions are ripe for a virtual governor (or social choice dictator) to arise.

3. Causality between physical and non-physical influences goes both ways. Points 1 and 2 together imply that we construct the very patterns that in turn governs our behavior—it’s circular causality, where top-down causality is as real as bottom-up.

4. The patterns use us as affordances to scale up and network into units that in turn multicausally summon new patterns that govern the patterns that govern us. Individual prices that govern individual markets face their own coordination problems as they compete to absorb portions of the economy’s total budget, a budget that stretches across all markets. The shared constraint reservoir puts prices into competition, which is resolved via the construction of the price system, the economy’s cognitive glue. The price system is just another one of these non-physical patterns, one that shows up to govern other non-physical patterns!

Economics could help to explain how pattern ingression occurs from Platonic space as people use physical forces to align to non-physical patterns (e.g., physically putting something into your shopping cart based on its price). But simplicity argues in favor of only one kind of pattern, physical or non-physical. So another possibility is that even the physical laws are non-physical, arising to govern interactions among the coupled elements of a single shared system.

What picture of the world does this perspective offer? Strong perception, the process by which we align to these patterns, is all about invariants: regularities that exist at some scale and substrate. We can tier pattern types by the reliability of the invariants, or how consistent they hold as substrate, scale, and time change.

• Mathematical invariants (primes, geometry, etc.)—this is (apparently) the limiting case, where the coordination is so deep it appears to hold atemporally and across all timescales and substrates. It’s hard to imagine anything anywhere in any way changing such that 6 becomes a prime number.

• Physical invariants (gravity, light propagation)—maximally consistent and reliable in our light cone. Other universes may have different laws of physics, or changes over some timescale at some scale may alter the laws of physics in our universe, but we won’t live to see it.

• Ecological invariants (optic flow, tau)—consistent within the relevant niche, or at a maximum, within a cognitive light cone.

• Social/economic invariants (prices, traffic lights)—consistent within institutions and certain patterns of coordination. In practice these invariants could be as reliable as ecological invariants—green meaning go and red meaning stop might hold thousands of years into the future—but they don’t have to be, and could change in pretty sudden and unpredictable ways, such as via price controls.

• Game-state invariants (marginal value of a basketball shot)—consistent within rules. And those rules aren’t even enforced consistently within a single game, let alone across games!

Ultimately, direct perception is about coupling to whatever level of invariant is relevant to your action. Strong anticipation works at every level because at every level the current state of the invariant specifies its near-future states in a way that’s proportional to how deep the underlying coordination runs. The organism, or any other behaving system, doesn't need to know which level it's coupled to; it just needs to align to the pattern. Doing so will pulls it into an anticipatory relationship with the variables it needs to regulate.

Mathematicians may thus indeed have an extra sense for mathematics. In a strong perception sense, this means that they couple to a highly consistent invariant structure that channels (or “enslaves”, in Haken’s terminology) their intellectual degrees of freedom into a trajectory through possibility space that reliably, though imperfectly, constraints them away from false conjectures and toward true and interesting ones. Economics students are trained to “think like an economist” to create this kind of sense in economics-space, indicating that this kind of extra sense can be learned, just like how a baseball fielder can learn to notice and do a better job of maintaining the visual relationship with the fly ball that allows them to catch it.

This builds to a potential resolution where where physical laws and mathematical structures are the same kind of thing, differing only in depth of coordination. All kinds of patterns and laws be thought of as invariants that emerge from coordination at different scales and at different levels of stability. This is why organisms can directly perceive all of them. Economics is a window into how this works because the patterns are deep and robust, non-physical but subject to a high degree of physical interaction and testing, and capable of exhibiting surprising changes when the conditions for their multicausal assembly break down.

Another way is based on strong anticipation. This kind of perception is about the organism coupling to the environment so that their interactions with it directly perceive information. When a baseball fielder chases a fly ball, they anticipate its trajectory by maintaining a visual relationship with it, not by computing a physics model. There’s no internal model, just direct coupling. As Karl Friston observed, “An agent does not have a model of its world – it is a model.” Call this “strong perception”.

In the ecological approach to perception and action, strong perception is called direct perception. Direct perception is about lawful interaction. When you sit in a chair, your body couples to the chair because of how lawful local interactions within the body, within the chair, and between the body and the chair, as well as other things like gravity, scale across the body-chair system. Sight as direct perception is enabled by the lawful interactions between light and the surfaces it passes over, meaning that the light that hits your eyes is structured in a way that precisely specifies the things you’re trying to look at. Without this structure, the fielder wouldn’t be able to reliably catch the baseball by maintaining a visual relationship with it.

Strong perception can be considered a generalization of direct perception if we extend it to economically lawful interactions. An example comes from the price system. Prices are perceptions in the sense that they are information you can use to form plans about how to navigate the world. But they’re organized by economic regularities rather than not physical laws. Prices emerge from the interactions of supply and demand. When you use prices to perceive the world, you’re peering into an economic problem space rather than a physical one—how to spend your budget optimally.

Indeed, strong perception is all about action-cognition, specifically how to reorganize your body in response to novel information. This reorganization is what makes the perception actual perception—strong perception naturally ignores all information consistent with its “prior”, or current dynamical trajectory. (I.e., such information produces no reorganization signal; the system’s dynamics effectively decide that the “information” isn’t informative.)

This process of reorganization via perception (really, reorganization as perception) is enabled by a cognitive glue. It transforms novelty into signals that the parts of the system are capable of dealing with, such as how the price system turns novel events into price updates that humans know how to deal with by updating their shopping decisions. Novelty that a cognitive glue can’t absorb is an externality.

So strong perception signals don’t end up telling you what’s going on in the world—that’s not the point. Instead, strong perception signals, like prices, end up tying a bow between you and the rest of the world so that you can use them to coordinate. These bow ties encapsulate the dynamical history of the relevant part of the world in a way that is suitable for being exploited to determine the optimal reorganization. For example, traffic lights are strong perceptions because you know based on your experience in a traffic-based society that it’s optimal for you to stop on red and go on green. This allows you to coordinate with a huge mass of other drivers just by looking at some lights and making decisions based on your own local circumstances. The meaning traffic lights have for action aren’t lawful so much as regular—another society could have red lights mean go and green lights mean stop—which is still sufficient for strong perception.

This kind of collective coordination based on strong perception doesn’t even require shared goals. Another kind of strong perception is marginal value. Sports offenses and defenses coordinate around points where the marginal value of scoring is highest—the offense wants to go there to score, and the defense wants to go there to stop them. The result is that teams with directly opposed goals coordinate around a shared value, the marginal value of a scoring attempt from that spot.

The strong anticipation version of power seeking works very differently. Economic growth is the clearest example. The economy gradually accumulates an increasing ability to reorganize itself to absorb novelty and regenerate its goals. This requires an increasing amount of specialized capabilities, including the ability to integrate more and more kinds of events into the price system so that future events fall within the economy’s cognitive light cone. I.e., as more contingencies become priced, they therefore become more actionable to the economy, making future states more responsive to the present. The economy takes over its environment by bringing previously uncoordinated or unresponsive contingencies into the domain of its effective control. Call this “strong power seeking.”

So while weak power seeking is intentional, “agentic”, and plan-based, strong power seeking is just the system’s expansion of its coordination ability (its ability to construct synergies).

Weak power seeking is sometimes envisioned to be about some kind of terminal goal, which power is instrumental to. However, the goal of a strong power seeking system like the economy is not, so far as I know, a concrete, specific outcome like “live forever”, “reproduce”, “stack the tallest pile of blocks possible”, etc. Instead, the goal is a more general, abstract one: to balance all the competing demands across the system—i.e., economic equilibrium or allostasis. Power is “sought” as efforts to expand a cognitive light cone: to align all the elements of the intelligence and its environment as a single coupled system.

I’m not sure there’s any reason to believe that the strong power seeking system cares what emergent macro-goal comes out of the economic balancing act, which is a function of the smaller goals of the members of the system, as well as environmental constraints and affordances, and the dynamical history of the system. From a strong power seeking perspective, what matters is maintaining system-level viability across competing constraints.

Systems may exhibit both weak and strong power seeking. A human may plot a coup, which is weak power seeking, and also grow from child to adult, which is strong power seeking.

Strong intelligence has a few unusual properties as compared to the traditional picture of intelligence, which is weak intelligence, or intelligence that does weak anticipation via weak internal models.

First, strong intelligence is a whole-body phenomenon. There is nothing in the economy analogous to what the brain is typically thought of as being: the place where the intelligence happens. Instead, the economy assembles intelligence across its entire being. For example, when a supply shock occurs, the economy as a whole rapidly adapts, even though no individual or no part of the system represents the problem as a whole or forecasts the full solution.

Second, strong intelligence is the ability of the system to coordinate itself to solve problems. Intelligence isn’t something the economy has; it’s something the economy does.

Third, strong intelligence doesn’t operate, at a system level, on forecasting, thinking, representing, generalizing, or any of the other things that are the hallmarks of weak intelligence. Strong intelligence occurs as the absorption of novelty into an existing infrastructure for translating surprise into a shared system of constraints that maintain coherence across the collective intelligence, environment, task, and structured space patterns.

Fourth, strong intelligence solves problems via reorganizing itself to maintain task-relevant coordination. Whereas weak intelligence constructs and manipulates representations, strong intelligence just adapts its own body to the task. A classic example is writing a signature with a pen: each pen and each surface are different, there’s all kinds of minor perturbations, but the body’s dynamics naturally produce a consistent signature without some weak internal model first needing to take in all of the information about the environment and plan an optimal motor behavior.

Humans are also examples of strong intelligence, though we can do weak intelligence as well. The A-not-B error is an excellent example of humans being strong intelligences. In this error, an infant repeatedly reaches to a location that was successful in the past but no longer is despite the fact that the infant can now see where the correct location is to reach. Weak-intelligence explanations, like the idea that infants lack object permanence or have limited short-term memory, were unsuccessful at explaining this.

Instead, the error emerges for reasons of strong anticipation. The infant perseverates in reaching to A instead of B not because it’s misrepresenting the world but because it’s stuck in a previously successful way of being in the world. The infant has the right believes in a weak anticipation sense, but it’s bad at reconfiguring its body in the circumstances in which perseveration occurs. Ironically, the infant errs because it’s too good at stabilizing successful patterns.

Normally, we think of learning as consisting of adding content—getting more information packed into the intelligence part of the system. Strong intelligences “learn” by improving their ability to reconfigure themselves over time. In the economy, this process is called economic growth, which is primarily a process of improving the system’s ability to flexibly reconfigure on the face of novelty, not accumulation of resources.

Ultimately, weak intelligence changes what the system knows; strong intelligence changes what the system is able to do as a coordinated whole. Weak intelligence cares about information processing; strong intelligence cares about constraint reorganization (e.g., relative price changes).

Strong categories enable one-shot learning because they’re simply a constraint structure that organizes behavior. Weak categories describe the world—this is a dog, that is a cat, etc.—but strong categories format interactions with it.

In the economy, categories are instantiated as price signals most obviously, but also as institutional norms and practices, legal conventions, etc. These things all function to reduce novel details into something familiar.

Prices provide a particularly good explanation for how novel events are quickly categorized. The basic function of a price as an order parameter for a market is to reduce high-dimensional novelty into a low-dimensional actionable variable. That is to say, when something new happens, economic agents don’t need to understand or even know anything about the novelty because the effects of the novel event are transmitted as updates to the price system. So to deal with the novel event, economic agents just need to deal with the things they’re already familiar with, like relative prices, budget constraints, and incentives.

So when something new happens, like a new technology, the economy is able to rapidly one-shot learn how to deal with because it doesn’t need to think about the new event and compare it to various concepts until a good match hopefully results. The economy instead just immediately embeds the new event into existing constraints, such as prices. Thus, the economy achieves one-shot learning by forcing novelty into preexisting action spaces, an outcome that arises from the natural dynamics of the economy.

The mystery of one-shot learning is how rapid generalization is possible with very little data. The solution is to realize that generalization isn’t required. The economy doesn’t infer the category from a set of examples; the category scaffolding is already present. The economy “learns” by reparameterizing its internal flows, not by update a weak internal model. There’s no need to represent novelty when you can just absorb it into the price system.

Sometimes the economy fails to learn quickly. This happens when the processes that produce strong categorization break down. When externalities are present, the economy by definition does not succeed in collapsing novelty into price adjustments. A failure to learn is the result of flawed categorization structures. The economy can only learn what its categories make visible.

In short, strong categorization enables one-shot learning because it formats novelty into a familiar, working structure. This is exactly what a category does: it gets rid of the extraneous details to prepare salient perception-action-cognition. But whereas weak categories are things you have to think about, strong categories emerge naturally from the dynamics of the system. The economy is able to behave as if it generalizes rapidly from minimal data because it never actually needed to generalize in the first place. Generalization is an artifact of weak categorization. One-shot learning isn’t really about speed; it’s about structure.

Moreover, this should be true of all systems, not just the economy, that run on cognitive glues that function as constraint-imposing mediums that collapse degrees of freedom in a task-relevant (action-coupled) way. Systems incapable of one-shot generalization probably are that way because they lack such a medium.

Strong categorization may be key to explaining how this is possible. A strong category is a stable pattern of interaction across body, environment, and task. Strong categories have the property of seeming natural, almost inevitable, once they become optimal. The body simply falls into the relevant attractor.

Unlike weak categories, which are formed via a laborious process of comparing many instances with each other until a robust statistical pattern emerges, strong categories “just happen” as a consequence of lawful interactions between coupled elements in a system. Thus, the system potentially does not need to have ever previously constructed a category for an instance of it to emerge. Rather than having to see 500 dogs to know what a dog is, the attractor just stabilizes.

How do strong categories reliably form in infants so quickly? I don’t know, but here’s some speculation. First, the attractor space already exists; the infant just has to couple to it to behave in a way that anticipates its trajectory. This means that the infant doesn’t have to create attractors ex nihilo but instead has to differentiate the existing space, a space that it naturally participates in by virtue of being a body in the world. A new category is thus carved out by a small perturbation to an existing attractor basin. It’s bifurcation.

Doing something as simple as staring at the ceiling would in fact give an infant a strong basis to start building new categories from since staring at the ceiling teaches the infant about the invariant structure of visual flow and gravity. This experience then can be generalized and transformed in application to other circumstances. Bifurcations would be triggered by the need to maintain relational homeostasis in changing circumstances.

How does the system maintain stability as it forms new categories? My guess is that rapid and accurate one-shot category construction is enabled by a cognitive glue. The cognitive glue holds the elements of the infant together, and keeps the infant coupled to the environment and task, so experimentation propagates rapidly through the infant, and the experimental category and the rest of the system are quickly forced to cohere to each other. Bad experiments simply fail to stabilize and vanish harmlessly, making them very low cost, while good experiments stabilize quickly.

Tellingly, not only do infants form categories very quickly, but they also maintain categories past when they should, as in the case of the infant perseverative error. These phenomena may be two sides of the same coin. Perhaps the more open a system is to forming new categories, the less skilled it necessarily is at matching categories to the situation. Development may ultimately close down attractor sensitivity, trading flexibility for precision.