And What It Means for the Algebra of Change
Every case in the preceding essay shared a constraint that none of its protagonists named, because none of them could see it. The Chagga built their furrow network over five centuries. Yacouba Sawadogo spent four decades turning 27 hectares of hardpan into a forest. Wangari Maathai’s trees grew one at a time across thirty years. The algebra worked slowly enough that the people who began it could not see what they were building, slowly enough that the results arrived after the patience had been tested to its limit, slowly enough that most people, watching the early stages, saw nothing happening at all.
This was structural, not incidental. The accumulation took as long as it did because the minimum viable act required a human body, human time, and human knowledge transmitted person to person at the speed of conversation. A pit dug by hand takes the time it takes. A furrow cut by a clan requires the seasons it requires. A tree planted by a woman grows at the rate trees grow.
We are now watching that constraint dissolve.
The cost of intelligence, the marginal cost of applying reasoning to a problem, generating a design, diagnosing a condition, teaching a concept, synthesising a solution is collapsing at a speed without precedent in the history of technology. This is not an incremental improvement in a familiar tool. It is a change in the nature of a resource. And it changes the algebra. The question is how.
There is a distinction in economics between rivalrous and non-rivalrous goods. A rivalrous good is one whose use by one person precludes its use by another. A doctor treating one patient cannot simultaneously treat a second. A lawyer advising one client cannot advise another at the same moment. A teacher explaining a concept to one child is not explaining it to the thirty children in the next village. For the entirety of human history, expertise has been rivalrous. This is why it has been scarce, why it has been expensive, and why it has been unequally distributed in ways that compound across generations.
A trained AI model is non-rivalrous in the purest sense: an information good, copyable at near-zero cost, usable by one person without depriving another. A model approaching diagnostic accuracy of specialist physicians can, in principle, be applied simultaneously to every patient who presents a query. Not sequentially. Not with a waiting list. Simultaneously. Assuming that capacity will exceed demand, the marginal cost of the millionth application will equal the marginal cost of the first. Intelligence, as an economic good, will therefore split into two layers: a non-rivalrous core — the weights — wrapped in a rivalrous substrate of compute and energy. Everything in this essay turns on that split, because enclosure does not need to capture the core. It only needs to own the substrate the core runs on.
Even so qualified, this is a structural transformation in the economics of knowledge. For the first time in history, the capacity to diagnose, design, advise and teach at the level of the best-trained practitioners is becoming a resource that could, in principle, reach everyone with access to the network. The Chagga built a water system that served a million people across five centuries of accumulated labour. A model could serve a billion people by Tuesday — provided someone has built the substrate, and on whatever terms its owners set.
That sentence requires pausing on — both halves of it. The first half names a change so fundamental that we have not yet developed the vocabulary to think about it. The second half is where this essay parts company with its predecessor, and where honesty requires a harder argument than enthusiasm permits.
The most durable transformations in the preceding essay were not programmes but conditions — self-sustaining systems that reproduced themselves because every participant had both the incentive and the knowledge to maintain them. The mfongo worked for five centuries because every family that depended on the water also held the knowledge to maintain the furrow and a governance stake in the system. The zaï spread because any farmer who saw it work could replicate it with materials already at hand. The knowledge was held locally. The system was owned by its users.
AI, as currently structured, is the opposite of this.
The computational infrastructure required to train and run frontier models is held by a handful of companies — four, perhaps five, at the level that matters. The capital required to build a single frontier model now exceeds the GDP of many of the countries whose populations would benefit most from abundant expertise. The data on which those models were trained was drawn — on terms still being contested in courts and, more slowly, in conscience — from the accumulated intellectual labour of humanity. And the models that result are licensed, not owned: accessible by subscription, revocable by terms of service, dependent on continued investment by entities whose primary obligation is to their shareholders. This is not the architecture of a condition. It is the architecture of a programme — controlled by fewer decision-makers than any development intervention in history.
A doctor trained in a village and embedded in that community cannot be switched off. A subscription to a medical AI model can be. But the contrast runs deeper than revocability. The village doctor is not merely rivalrous expertise; she is an accountable node in a social system — liable for her errors, present for their consequences, trusted because the trust was accumulated the way everything else in this argument is accumulated, one consultation at a time. Expertise without accountability is advice. What turns advice into care, or counsel, or teaching, is the institutional fabric around it and that fabric is precisely what a subscription does not carry.
The history of scarcity is also, inseparably, the history of institutions designed to manage it. Property rights, legal systems, markets are at their core mechanisms for allocating finite resources among competing claimants. They were built for a world where expertise was rivalrous and scarce. They are not equipped for a world in which its core has become non-rivalrous and abundant and this creates what might be called the problem of artificial scarcity. It is possible for institutions built to manage genuine scarcity to manufacture scarcity where none need exist. Intellectual property regimes that restrict the use of models. Compliance frameworks so expensive that only large incumbents can afford them. Platforms that gate access behind subscription fees calibrated to extract maximum surplus from users with no alternative. Each is individually defensible on grounds that are not dishonest: property protects investment, regulation prevents harm, platforms need revenue. Their aggregate effect is to rebuild scarcity in the very domain where the technology has eliminated it; external legibility for internal coherence; artificial scarcity for the coherence of a commons. The result is the same: a system that works for the people who designed it and not for the people who need it.
The counter-pressure is real. Open-weight models have demonstrated that the non-rivalrous core need not be confined to proprietary platforms: a developer in Lagos using an open model commands reasoning capacity that a decade ago required the infrastructure of a major research institution. The question is whether this openness is a durable condition or a transitional phase. The internet is the instructive precedent, and it is more precise than the usual lament. The internet’s protocols never closed — they are as open today as in 1995. The enclosure happened in the layers above them, where data and network effects accumulate: search, social platforms, marketplaces. Openness at one layer did not prevent enclosure at another; it determined where the enclosure formed. The prediction this yields for AI is uncomfortable in its specificity. Open weights may well persist, the way open protocols persisted, while the enclosure forms beneath them, in chips and energy, and above them, in the distribution surfaces through which ordinary people actually reach the capability. An open model is necessary for a commons. It is nowhere near sufficient.
The preceding essay identified four elements present in every durable transformation: a specific and legible privation, a minimum viable act, a network effect, and sufficient time. The question for AI is which of these it changes and which it leaves intact.
It changes time, dramatically. When the minimum viable act involves expertise rather than physical labour, it can be executed at near-zero marginal cost across millions of simultaneous instances — which means the network effect, the moment at which an act becomes more likely because others have already done it, can arrive within months rather than decades. The zaï spread across the Sahel over forty years of farmer-to-farmer teaching. An AI-enabled extension service could demonstrate the same technique to every farmer in the Sahel in a season.
It does not change the requirement for a specific and legible privation. The technology has no mechanism for identifying the right problem. It can optimise any objective function with extraordinary efficiency; determining which objective to optimise, which privation to address, which minimum viable act to enable, remains a human judgment. This is not a minor caveat. The history of development failure is largely a history of optimising for the wrong objective with great efficiency.
And it does not, on its own, create the condition. Ralegan Siddhi’s embankment worked because the community owned the result and understood the mechanism. An AI system that delivers a solution without transferring the understanding of how it works produces dependency, not capacity. A programme that uses AI is still a programme.
What AI can do, if directed with the structural intelligence Sawadogo brought to the zaï, is compress the time between the minimum viable act and the legible result. It can make the algebra faster. It cannot make it different. The direction still has to be chosen. The ownership still has to be local. The condition still has to be self-sustaining.
There is a version of the AI transition that reproduces every failure mode of development orthodoxy — expert-designed, externally legible, dependency-creating, answerable to the wrong stakeholders. In this version, the non-rivalrous core becomes the foundation of a new monopoly: not the East India Company accumulating a subcontinent through a succession of small decisions, but three or four companies accumulating a civilisation’s worth of cognitive infrastructure through a succession of licensing agreements. The enclosure would be invisible in the way the best enclosures always are. The field would still look green. The access would still feel free, for as long as the freemium model serves the acquisition strategy. The dependency would accumulate quietly, each step defensible, the aggregate catastrophic.
There is another version, in which the collapse in the cost of intelligence reaches the places that need it in a form they can own: open models running on local infrastructure, agricultural AI trained on local crop varieties and soils, medical AI calibrated to local disease burdens and languages, legal AI that turns the entitlements Sen spent his career identifying into something a landless labourer in Bihar can actually use. The non-rivalrous core translates into non-rivalrous access. The algebra accelerates.
The difference between these two versions is not technological — the technology is identical in both. The difference is institutional: the difference between conditions and programmes, between systems that reproduce themselves and systems that can be withdrawn. And an essay that insists on minimum viable acts owes a candidate for one.
Here is its shape. An open model adapted to a single district — its crop varieties, its disease burden, its language — by the institution that already serves that district, with the adaptation documented plainly enough that the next district can copy the method rather than licence the product. The act is small enough to complete: a district-level adaptation is within the reach of an agricultural university, a state health directorate, a teacher-training college. Its result is legible within a season: the advisory that arrives in the farmer’s own language, about the farmer’s own soil. It replicates with what the next institution already has, because the method travels with the output. And it passes the test that separates a condition from a programme: it transfers the understanding, not just the answer. This is the zaï pit of the new terrain. Whether anyone digs it — whether public money builds percolation tanks of compute the way it once built percolation tanks of water — is the institutional choice on which the two versions divide.
This is the second unlocking. Not the unlocking of capability — that is underway and cannot be reversed — but of the conditions under which capability becomes a commons rather than an enclosure. It requires, in every domain where it matters, the same patient structural intelligence Sawadogo brought to a field of hardpan in Burkina Faso: the identification of the minimum viable act, the creation of conditions for it to replicate, and the refusal to mistake the speed of the technology for the wisdom of its direction.
The algebra has not changed — only the speed at which it runs. But speed changes one thing. The question of who is doing the arithmetic used to wait decades for an answer. It will not wait now.