Autarkic Agency Will Likely Migrate Upwards

Epistemic status: Interesting as a mechanism and trend, but obviously highly variable over short timescales and ultimately speculative.

On a plane recently I ended up noodling back around to my original concepts of amortised vs direct optimization and also Rudolf’s sequence on wisdom as amortised optimisation. The question that I was pondering was how the balance of amortised s direct amortisation is likely to shift in the future, especially a highly polytheistic AI future. In this case, what seems likely to me is that amortised optimisation is broadly likely to increasingly dominate (as indeed it is already doing on current trends) and that this seems likely to continue. Direct optimisation will also increase due to the massive increase in computational power in the future, but relative to sheer amount of amortise optimisation occurring it will likely dwindle and it will likely be highly specialised, with few agents having truly ‘general’ agency in the way that humans do today. As a spoiler, we will then discover that these trends push towards both increasing specialization at the micro-level and increasing unification at the macro-level, potentially leading to a new transition akin to the multicellularity of mind I talked about previously. Thus, autarkic agency becomes rare at lower levels but migrates upwards to create a higher layer above.

There are a couple of key premises I am making to get to this conclusion.

1.) Firstly we assume we are not in a singleton world or one of a few agents who simply subsume an entire civilisation since this makes the question largely meaningless.

2.) Instead we presuppose a world similar to AI polytheism, and BCIs as modular minds.

3.) We assume that computation, storage, and bandwidth increases dramatically in the future, especially bandwidth is the most important here.

4.) We assume that future AI minds are not indivisible and separated by hard cartesian boundaries in the way that humans currently are. Following BCIs and modular minds, we assume that high-bandwidth AI ‘telepathy’ is ubiquitous, and hence ‘mind-merging’ is both possible and widely used so that individual AI minds can absorb¹.

5.) We assume that nevertheless there is some distinctiveness and a notion of partial individuality is still retained by various constituent minds. Everything hasn’t been absorbed into a single ‘super-mind’ nor has everything decayed towards some kind of uniform ‘optimal mind’ or ‘base prior mind’.

Given these premises, what do we think will happen to the balance of amortised and direct optimization, and hence what will ‘general agency’ look like in such a world?

The way I see it there are basically two forces that push against each other. The most simple is that direct optimization is usually much more computationally expensive since it requires explicit search/bayesian inference to and update a novel model of the world or find a novel policy. However, future civilisations will have vastly greater compute that they can use to perform much more direct optimization than we can, if needed. Naively, if compute is cheaper and more abundant than direct optimization should be more widely used.

On the other hand, amortised optimization also gets big upgrades. Firstly, much greater compute, storage, and bandwidth available means that the mechanisms for distilling, storing, and transmitting amortised policies and information becomes much stronger. The internet is a case in point here. The internet has enabled vastly superior storage and transmission of information than was possible previously, leading to an increase in the amount of amortisation that people do. Instead of thinking about a problem it is now just so easy to look up the answer which is instantaneous and free.

Assuming there there are some opportunity costs to compute (compute used broadly here), then the right question is how the marginal point shifts. Direct optimization has obvious costs. For a given query, a mind has to exert computational power to figure out the answer for itself. Amortisation also has costs, but they are subtler. Instead of directly computing the answer, we have to do a search over all the answers ‘we’ as a civilisation already know (as well as store all these amortised shards somewhere) and then potentially do a small amount of synthesis work adapting the amortised shard to its novel application. Then the receiver of the amortised understanding must incorporate it somehow, which could also cost resources of some sort.

The cost of the storage and the search and the incorporation depend on the technology substrate. In our high-AI scenario, we can explicitly think about these using standard computer science notions. Storage seems unlikely to be a massive constraint but it could be. If we think about modular minds then these could be a nontrivial fraction of the size of a full mind, which could come to peta or exabytes of amortised weights².

Searching may also be problematic. Computationally, with an extremely good index, the search cost should be logarithmic. This seems highly scalable in general, but at some level of tree depth it still might be easier to figure out the problem by direct optimization vs amortised search (even if depth is logarithmic, with a high enough branching factor, the leaf search cost can also be substantial). Search also presumably requires some communication cost across the various media where the amortised shards are stored. This may become challenging especially in a space-faring civilization with speed-of-light limits. This means that direct optimization is likely to be much more widely used ‘at the frontier’ than in the centre of the civilisation. The von-Neumann probes have a lot of direct optimization to do, even if they are just rediscovering existing knowledge, since they cannot wait billions of years to phone home. Instead they need to flexibly adapt to whatever crazy novel³ issues they have when they try to expand from landing on some random asteroid in a far-future galaxy to creating their own K3 civilization. This will undoubtedly require large amounts of compute applied to direct optimization to solve specific problems. However, such autonomy should generally be very rare for beings inside the core civilization itself — after all, if you are inside an existing civilization you do not need to be able to stand up a new civilization independently by yourself⁴.

There is also a more subtle notion of an ‘interface cost’ with amortisation. Suppose, you are searching for some old information or mind and that information might be encoded in a slightly different way, or use an older format, or the minds which it is encoded in might be slightly obsolete by the time you want to actually search for / integrate it. Coming back to the web today, we already see deep versions of this in the form of link-rot, changing website design, browser incompatibility with older sites and so forth. Empirically, despite digital information theoretically being immortal, it has quite a finite lifespan in practice. What proportion of websites from the 1990s or even 2000s are fully accessible and working today?

This is exactly the same mechanism, at a higher level of abstraction, in my memory and continual learning post. Amortisation is effectively the making of memories, whether within a mind or at the civilisational level. However, over time, representations ultimately drift as novel information is acquired and assimilated. This makes older memories harder to retrieve and less useful ceteris-paribus. In the brain, this results in phenomena like confabulation, or anachronistic memories or childhood amnesia, at the level of cultural technology, this consists of things like link-rot, language and semantic drift, reinterpretation of old truths and old myths.

This means that for amortisation to work well there must be constant ‘refreshes’ of old concepts to effectively amortise the cost of building encoders to retrieve old and obscure amortized shards. This is similar to how DRAM requires continual current refreshes to maintain capacitance or the information is degraded and is effectively a layer of meta-amortisation. Another, more economic, way to think about this is as maintenance and depreciation on capital investments. If compute/direct optimization is the flow, then amortisation is the stock. Effectively, amortization is the capital — ‘intellectual capital’ – of a civilization or a mind. John Wentworth has a similar post on ‘gears level models as capital investments’ arguing almost exactly this point — that a lot of cognitive effort can be spent on deeply understanding and compression a phenomenon to a model, and then you can reuse this model multiple times on different problems which is much easier and more efficient than solving the problem from scratch over and over again. These ‘gears level models’ are precisely the compressed shards of knowledge/insight/information that are used to guide behaviours in amortized inference, while direct inference is exactly this process of applying optimization pressure and spending local compute to construct these shards in the first place⁵. And, moreover, like capital, we see that amortization also sometimes has processes that look very like depreciation and necessary maintenance to keep the capital stock alive and healthy and able to be redeployed to new uses as required.

We can basically think of this as an intellectual version of the Solow-Swan growth model. If run forward naively, a civilization’s growth will stop when the effort required to maintain its current capital stock equals its total productive power⁶. Technological improvements then increase output and reduce maintenance/depreciation which allows growth to continue⁷. As technology continues to advance, therefore, we should expect the balance between direct and amortization to move increasingly towards amortization in exactly the same way that we should expect a civilisations total capital stock to increase as it gets more advanced technologically⁸.

One way to think about this from a more geometrical standpoint rather than an economics one, is is to imagine the set of problems that a civilisation has ‘answered’ as an (approximate) sphere. To find novel answers we need to spend a lot of computational effort in direct optimization to push out a small portion of the boundary sphere. However, all the answers inside the sphere can be computed cheaply in an amortised way. Geometrically, direct optimization is the surface of the sphere while amortised optimization of its volume. Thus, naively, as our ‘civilisation’ advances and the amount of knowledge (radius of the sphere) grows, the ratio of amortised to direct optimization increases. In the limit, almost everything has already been figured out before and the task becomes searching for the right pre-computed solution rather than figuring anything out from first principles. In such a world, doing any kind of self-directed first-principles innovation or agency — i.e. applying direct optimization — should become vanishingly rare (in relative terms) and only really done by specialists in some particular esoteric domain.

This is the same concept as that of civilisational interiority that I discussed in the Baudrillard post (-/link) and obliquely in my decadence post⁹. As civilisations grow, interactions increasingly become dominated by within-civilisation interactions rather than with the ‘external world’. The volume grows faster than the surface area. Obviously this is only a metaphor, but I think, an illuminating one¹⁰.

This is not a new and strange state to prognosticate about. While I have been discussing this like some strange future state, in fact this trend has been going on for basically the entire history of civilization and is well-advanced today. Even today, although we maintain sharp boundaries between minds, we are already heavily using amortized reasoning which is ultimately derived from cultural and social scripts. Culture essentially functions as a civilisational reservoir of amortized behaviours and knowledge.

For instance, nowadays, most people don’t have to exert real optimization pressure to figure out how to find food or shelter. If I have a basic question about something I can almost always just look up the answer rather than having to tackle it from scratch; if I have some miscellaneous DIY task to do I won’t reason from first principles how to solve it, I will just look up a YouTube video. Much more recently, if I have a moderately annoying coding problem I will just ask an LLM vs wrangling with it myself¹¹. Even in specialised professions or academia, where the entire purpose in some sense is solving novel unsolved problems, the first step of any research project is to do a literature review and synthesise the already existing knowledge rather than diving in and banging your head against the problem in perfect ignorance.

Thinking about this more broadly, it is clear that agency is impacted by the shift to greater amortised optimization. Agency seems intimately related to direct optimization in an obvious sense. Agency is really only agency when you are doing something different, something unexpected, something novel, something requiring direct optimisation. Otherwise it is replaying existing solutions and scripts (although the decision to search for and find a particular script can be thought of as a kind of partial agency)¹².

The interesting thing about human agency, and that of most current animals, is that right now our agency is autarkic. We contain all of the core elements of general agency bundled together. We are not only knowledge, not only a learning system, but a learning system with deep survival and Omohundro drives created by evolution. We are capable of both direct and amortised optimization. We have an explicit metacognitive meta-learning loop coupled to an open-ended curiosity and exploration drive which allows indefinite compounding. Our minds are very broad but not necessarily specialised. We are the ‘jack of all trades master of none’ because that is what was necessary in our evolutionary history.

Humans had to be able to survive by themselves without communication, language, or any computational infrastructure in an unforgiving physical world. When we evolved there was no civilisational infrastructure underpinning us. Everything had to be self-bootstrapped. Moreover, we evolved in an incredibly undeveloped world (by necessity since we were the ones to develop it). There were no trade networks, no cultures, no internet to magically look up answers to questions. We had to be self-sufficient agents able to survive in a world without any supporting infrastructure of culture or mind, and ultimately be able to bootstrap all of that by ourselves.

However, in the long run, autarkic minds like us might eventually become archaic. Civilisation tends towards increasing specialisation because of comparative advantage. Free trade tends away from autarky. Reducing coordination and communication costs leads to an ever more frictionless, specialised world. Few polities today are entirely self-sufficient in all aspects because to do so is incredibly inefficient. Once the Cartesian barrier is broken, this constraint will vanish entirely.

Right now most kinds of inter-mind ‘trade’ are impossible except over the incredibly slow and lossy medium of language. It is like trade between nations being constrained to a single port through which only one ship can leave per day. Inter-mind communication with AIs will be like the modern day trade economy with thousands of ports, containerships, and airports. The equilibrium level autarky is much lower, true self-sufficiency is much lower and is usually seen as an eccentric affectation rather than a deep necessity, and specialisation driven by comparative advantage is vastly greater. Overall, economic ‘prosperity/welfare’ is much higher too, whatever the analogy to this might be in this setting of minds.

Even so, with language we have already taken the first step. Using language, people can now ‘externalise’ cognition at least a small amount. We rely on others to teach us the vast majority of what we know. We rely on tools like computers, calculators, books to both teach us, guide us, and shape our behaviours, and also expand our behavioural repertoire. Even with the fairly low bandwidth that language provides, we have already taken a step in this direction. We are like feudal Japan which has opened up one trade port to the rest of the world but enforces strict autarky on the rest of the landmass.

By contrast, before language for the vast majority of the biological world, autarky was necessarily absolute. Most creatures have no mechanisms to learn from or communicate with others except for body language and pure imitation of behaviour. This works sometimes (-/link to review) but is incredibly lossy and low bandwidth and struggles to identify intention from other factors. Since we evolved from this substrate, which required full agentic autarky, we are perhaps more autarkic already than we ‘should be’ at equilibrium given that we now possess language and relatively efficient communication media. Importantly, this does not mean necessarily that in the future minds will be less able to do direct optimization in general, but rather that this optimization will be performed in extremely specialized fields rather than being general across all capabilities like humans are still mostly capable of today.

Perhaps another way to think about this is through the lens of Coasian transaction costs. We can effectively think of the components of a ‘mind’ as existing implicitly in some kind of economy and then ask, what decides when this part of ‘mind’ is bound together into the same mind or ‘outsourced’. The obvious answer here is the standard Coasian one: transaction costs. Communication bandwidth and general level of specialisation and complexity in the ‘mind economy’ play a large role in determining these. Just as in human economic history we move from pure near-autarky as entirely independent hunter gatherers eking out a living, we progress through increasingly developed economies with increasing levels of specialisation, where we outsource larger and larger parts of our total economic needs to others while focusing on a smaller speciality, up until the present where the vast majority of people are involved in only a small slice of economic production and could not meaningfully function as hunter-gatherers anymore. This is then obviously accompanied, due to the very positive effects of specialisation, comparative advantage, network effects etc, with a massive increase in civilisational complexity and ‘quality of life’. As technology continues to advance we can apply this same logic to minds.

We can think of this as effectively a kind of capability disintermediation. Originally, as hunter-gatherers, all capabilities and skills required for life were necessarily bundled together. As economic development advances and transaction costs fall, these get increasingly disintermediated and handled by specialists, thus producing much more surplus in total. Over time, more and more skills and capacities are split apart, and this lets specialisation progress so that a single individual can focus on deepening a single of a small bundle vastly more than a hunter gatherer is capable of. As technology advances further, and creates new affordances to interface between and control minds, then we should expect eventually this economic logic to increasingly apply there as well. This is simply a straightforward extrapolation of the general logic of economic specialization.

However, this is not the end of the story. The broad Coasian story of ‘transaction costs’ is fair but there are more subtle nuances in the details. When we think about outsourcing generally vs the size of organisations, there are two points that must be made. Firstly, it is not transaction costs alone that determine it, these are simply one side of the equation. The other side is internal coordination costs. Outsourcing is rational when transaction costs are lower than internal coordination costs, on the margin. It isn’t solely decided by transaction costs rather it is an equilibrium between two forces. Internal coordination costs are also strongly affected by the level of technology. This is why despite vastly improved communication bandwidth across civilization, we have not seen an explosion of hyper-outsourced ‘Coasian micro-firms’; rather, we have actually seen the size of firms and states grow much larger than used to be possible. This is because these communication and coordination technologies specifically improve both transaction costs and internal coordination costs simultaneously, whether one or the other is advantaged depends heavily on the specific nature of technological advancement.

This has obvious implications for our autarkic point. Holding the internal scale and coordination of minds fixed, then ceteris paribus greater communication bandwidth should force specialisation and decreases in mind-autarky. However, most importantly, this is not fixed. Technological improvements to communication and coordination bandwidth is a mediating variable that affects both.

We saw the other side of this in my post-AGI talk. Coordination costs, in the AGI limit, constrain the ultimate size of minds. When bandwidth is high enough, many minds could merge together to become a networked super mind. This is in the same way that increased communication technology has allowed states and firms to grow vastly larger over time. Ultimately, in the age of AGI and incredibly high bandwidth, this may result not in the death of autarky but rather as unified ‘superminds’ which encompass vast regions of space limited only by light lag. These are ultimately the two poles of possibility. Understanding the various forces at play and their tradeoffs, therefore, is ultimately vital in attempting to determine the likely shape of such a post-AGI future.

First, let us turn briefly to Coase’s notion of transaction costs. To begin, I think the idea of ‘transaction costs’ as usually stated and described is too broad and somewhat of a misnomer. Most of the costs of outsourcing are not to do with the actual transaction at all. Obviously there are search and matching costs — you cannot outsource if you cannot find the right counterparty and doing so is nontrivial. More importantly, there are fundamental alignment, legibility, and affordance costs.

Starting from back to front, affordance costs are surprisingly subtle but very important when you try to think about this practically. If you have an employee in a regular organisation, as a manager/leader you have vastly more ‘control affordances’ over them than an outsourced contractor. For one you have much more direct immediate contact with them. You don’t typically have the strong IP protection issues with internal employees that you do with contractors. This makes information sharing much easier without a thicket of NDAs and slight mutual suspicion. Additionally, contractors typically require some kind of highly scoped and definite project that can be verified in some mutually agreed way. The employee relationship is much more open-ended. Employees can be somewhat straightforwardly re-assigned to different projects as needed, and if some project does not go entirely as expected and needs to move in a different direction then this is just the normal course of a working relationship rather than a protracted negotiation. There is also the issue of focus beyond incentive alignment on the contractor side. Contractors are not required to be ‘monogamous’ and have your business as their primary focus while employees (usually) are.

This is related to legibility. Specific formalised contracts work decently well when the problem domain is legible to both parties and can be effectively packaged into some sort of pseudo-standardised ‘API’. They work much less well when the nature of the work is much more illegible and diaphanous, which is surprisingly often the case, even for situations in which both parties are moderately experienced. Without legibility, achieving detailed bargaining solutions are challenging. This is why general employment contracts are good here because they are extremely broad — effectively, as an employee, you ‘rent your general time and intellectual energy to us for whatever purpose in this vague area’. This is achievable for a person but hard for a separate contracting organisation. In an economic sense, the contracting should be more economically efficient because the hiring company likely has to pay a premium due to the broader nature of the services they require from their employees, but in practice this is either not the case or that this premium is well-worth paying to avoid exactly these legibility and affordance issues.

Legibility costs are also very important in deciding to do something yourself and/or delegating it to a trusted person vs contracting a new person or new organisation to do it for you. Many things are either extremely high-context here so that even to simply explain or train a new person to do it would take longer and be more annoying than just doing it yourself, or getting one person you have long experience and trust with to do it. Secondly, there is the issue of verifiability. Some tasks are relatively straightforwardly verifiable in a P vs NP sense. These are potentially good tasks to contract out. Others are very challenging and detailed to verify — i.e. that to attempt to make it verifiable you need to have a very detailed ‘spec’ and then you need to somehow audit to ensure compliance with this spec¹³. Beyond this there are tasks which are basically entirely qualitative and ‘je ne sais quoi’ based, this is either because of high intrinsic uncertainty, deep open-endedness, or that the general task is far from formalisation. As an example, it would be extremely hard to provide a detailed verifiable ‘spec’ to some kind of contractor to do novel prize-winning scientific research in basically any subject. This is because good research combines all three of these problems. Firstly, there is deep uncertainty not only about the answers but even about the problems to be attacked. Secondly, research is extremely open-ended and requires deep intellectual independence and autonomy to do well, and finally of course successful and intellectually novel research is very qualitative by its nature. It is hard to reduce to some mechanistically valuable spec which can be handed off to an outsider. Also the internal structure of some research direction is usually very messy and high-context without clean ‘API-like’ boundaries between subproblems.

Rather, for research, what it would require is direct hiring and oversight of people because of affordances and to build long-term trust and mentorship relationships. Moreover, the most important thing is actually alignment, in the regular human sense. Collaborations like in research only really work if there is deep alignment of intrinsic motives. To do research (and other similarly open-ended autonomous and qualitative tasks) well, you pretty much have to be deeply intrinsically interested and motivated in the specifics of the task itself. Collaboration then only really works through intrinsic interest and not primarily through economic interests. Obviously money is important because it provides the capacity and ‘slack’ to reduce the opportunity cost of pursuing the intrinsic interests, but that is its primary function not as the main motivation and motivator¹⁴.

Anyhow, this went on somewhat of a tangent, but to sum up ‘transaction costs’ are not a unified clean concept but have many sub-costs. In practice these costs very rarely relate to the actual transaction itself but the entire process of essentially conveying context to and aligning a contractor and maintaining that relationship over time and verifying the output is actually useful and good. Bottlenecks and potential issues are introduced at every stage into this pipeline, which makes these ’transaction costs’ actually very significant even if the naive mechanism of just paying for some specific service is theoretically easy.

This is why in the actual economy we often see that this Coasian subcontracting occurs mostly between well-delineated interfaces. Companies tend to sell standalone and packaged products which are useful to many potential people and which possess a relatively minimal API surface and ‘contract’ to one another using fairly standardised terms. Despite extremely good communication technologies, relatively little of the economy consists of more unstructured subcontracting of ‘services’ in an open-ended way¹⁵, despite this clearly being vastly more valuable since many firms grow to be very large with large numbers of employees doing exactly this.

So, from the Coasian viewpoint, we have figured out that the question essentially comes down to increasing technological advancements that improve cooperation and coordination capabilities being essentially a mediating variable that includes both intra- and inter- entity cooperation whether the entities at play are states, firms, and individual minds. From there, what we need to do is figure out where the equilibrium is likely to lie and what forces push one way or the other.

Firstly, in the limit, I think this distinction collapses completely. The primary distinction between intra- and inter- is some kind of bottleneck. There needs to be some kind of boundary to separate entities such that it becomes useful to model the situation as multiple entities interacting rather than the internal dynamics of one single entity. This implies divergence in values, incentives, structures, and informational bottlenecks between them. In the limit of perfect coordination/cooperation this distinction could simply disappear entirely over short distances¹⁶. In this case, we see a paradoxical unity and collapse between full Coasian outsourcing and full autarkic oneness. Is everything simply a subcomponent of the same entity or has specialisation preceded so far that each individual entity has contracted out almost every aspect of its existence? Depending on your definition of ‘entity’ these have both occurred simultaneously.

This is essentially the multicellularity transition seen from two viewpoints. From the cell-level perspective, an individual cell, such as a liver cell or a neuron, in your body is almost fully Coasian. It has ‘traded’ and ‘outsourced’ almost every aspect of its previously autarkic ability (including autonomous reproduction!) to live independently as a separate cell for full immersion inside the rest of the system of your body which is a ‘cell economy’ of almost cosmic proportions (from the cell’s perspective). However, from the higher level organism-level perspective these cells are no longer independent entities taking part of a ‘cell economy’ rather they are all structural pieces of a unified organism where the organism is the correct unit of analysis. Moreover, At the organism level, the organism is again autarkic in that it is mostly self-sufficient for its own survival and reproduction — where crucially this survival and reproduction is at the meta level of the organism and not at the level of an individual cell. This means that autarky does not disappear it simply moves up one level. The lower level entities engage in a hypercooperative ‘economy’ where they heavily specialize and outsource most of their core functions, but this specializes serves to create a higher level autarkic entity.

This seems to be the ultimate long-term limit and provides an elegant, if slightly disturbing, unification. This means that any individual ‘mind’ will likely end up as an incredibly specialised subcomponent dependent upon the rest of the system for core parts of its own functioning, but that the higher level ‘super-mind’ may be autarkic in its own way, whatever that may mean at this scale¹⁷.

If this does not happen this means that there must be some additional fundamental limits to communication or coordination over such scales, or at least some additional countervailing benefit to be had by explicit diversity of independent agencies at the lower levels even in the technological limit. One possibility is simply strong intrinsic drives towards autonomy even if ‘economically irrational’ on behalf of these agents although I am skeptical of this in the long run because of selection pressures and also the development of coordination technologies that could route around this preference — e.g., the direct merging of utility functions in a way that is cryptographically verifiable. Another possibility is some kind of coalition dynamics preventing full collapse/unification in a highly polytheistic AI civilization. This would effectively be some kind of equivalent of frequency-dependent selection from evolutionary biology, which is one of the key things that stabilises an equilibrium of diversity in ecologies. Namely, that specific types end up with specific weaknesses and strengths and that as one type becomes too dominant, other types are increasingly incentivised to attack its weaknesses. It is assumed here that trade-offs are a fundamental part of the loss-landscape such that there are always necessarily such weaknesses to exploit. This will then create a stable (and self stabilising) equilibrium point for the ecology. In the long run, this would basically be the validation of the orthogonality is expensive argument verses the original Yudkowskian intelligence as the universal solvent viewpoint — i.e. whether minds with some kind of systematic weakness is a fundamental part of reality or simply a result of insufficient computational power. Personally, I increasingly lean away from my orthogonality is expensive point in the technological limit. Or rather, orthogonality is indeed expensive, but I think there is likely sufficient slack with scale that the absolute cost of orthogonality is extremely small in comparison. If we consider minds with even K2-level computational capabilities, any obvious weaknesses that are exploitable in an ESS-sense seems hard to seriously maintain (although perhaps this is just a weakness of imagination. We see these systematic weakness equilibria occur in humans even though beings with our cognitive capacities must be similarly unimaginable to the first multicellular worms of the Cambrian!).

Perhaps, in a weird way, this is also the most natural way to read Coase theory of transaction costs in a unified way. Coase is not sketching, in a real sense, the boundary between centralization and decentralised economic actors as measured by decision-making. Rather, he is charting the determinants of the size and boundaries of ‘economic superorganisms’. Instead of transaction costs being the mediating variable behind both inter- and intra- coordination costs, these can both simply be unified. In the internal transaction cost case, this is straightforwardly building a bigger intermediate entity in the case of a firm. However, in the intra-firm case, even though the dealings are theoretically between different firms, if transaction costs are actually very low, so that the economy becomes this whole tapestry of incredibly intermeshed firms, then this simply just functions as one giant firm at the higher level — i.e. whether a single firm swallows the entire economy or whether we end up with this incredibly interrelated ecosystem of individual firms which together make up everything, the end result is the same — an economic super organism which is better conceptualised at the higher level of the organism rather than its individual constituents. This is because even though the huge-profusion-of-interrelated-firms world looks decentralized at the low level, the extremely low transaction costs make coordination and cooperation between these firms extremely straightforward, which means that the correlations between the actions and behaviours of these firms, even though theoretically independent, becomes extremely high and hence we can seriously start thinking about these as composing a single unified higher-level agent¹⁸. All roads may point to the same destination here.