Exploration

The Third “End of History” — AI Singletons and Enlightened Resistance

Two disappointments, decreasing dependence on human motivation, and multi-agent civilization

Exploration v0.2 · 2026-09-26 · working hypothesis

Provisional hypothesis: Marxist world revolution and post-Cold-War liberal convergence both contained expectations that history would broadly converge toward an order regarded as desirable, yet neither convergence occurred automatically. What may be qualitatively new about advanced AI is not that it proves the correctness of a final order, but that it can reduce an order's dependence on human labor, military service, administration, consent, and motivation, moving it toward a self-maintaining closed causal system. This page does not deny pressures toward singleton formation. It takes those pressures seriously while asking whether irreversible concentration correlates value lock-in, political redundancy, information hazards, and the loss of external corrigibility.

1. The first disappointment — world revolution did not propagate automatically

The Russian Revolution was not initially conceived as a final order complete within one country. In 1921 Lenin recalled that, around the time of the revolution, Bolsheviks had expected revolution in other advanced capitalist countries to come “immediately, or at least very quickly,” whereas events proceeded less directly than expected. This is better read as a revision of the expected speed and path of world revolution than as an abandonment of the expectation itself. The failure of the so-called German October in 1923 became one symbolic boundary between the early revolutionary wave and a much longer period of coexistence among states.

The point here is not that a timetable prediction was wrong. Even if competition among capitalist firms is abolished domestically, if world revolution stops and independent states remain, selection pressure over productive capacity, military strength, and technology can reappear at the interstate level. If there is no final authority above states, the competitive unit has shifted upward rather than disappeared.

External problem: abolishing competition among firms ≠ abolishing competitive pressure. If world revolution is incomplete, selection can move upward from firms to states.

There is also an internal anthropological problem. In the 1844 manuscripts Marx contrasts alienated labor with the “free, conscious activity” characteristic of human species-being. This project does not treat that passage as a complete institutional design. But if one extends the underlying picture into a strong institutional premise—namely that non-alienated production can draw on substantial voluntary and communal motivation—then the premise should apply symmetrically not only to workers but also to planners, party officials, and administrators.

If workers are expected to contribute beyond narrow self-interest while administrators alone are assumed always to exploit office, information, and privilege for private ends, the anthropology becomes asymmetric. Conversely, if communal motivation is weaker than the strong premise requires, then weak labor incentives and bureaucratic self-entrenchment can be read as two manifestations of the same motivational constraint.

Internal problem: common ownership ≠ communal motivation. The more strongly species-being is used as an institutional premise, the more symmetrically it must apply to producers and administrators.

This page therefore treats the first failed “end of history” as a double working hypothesis: historical convergence through self-expanding revolution failed to materialize, while the motivational capacities needed to operate the new order could not simply be assumed. The second point is an extension developed here, not an explicit prediction made by Marx.

The two failures reinforce one another. If interstate competition persists externally while motivational problems persist internally, an isolated socialist state can become a polity simultaneously pressured toward internal administrative control and external accumulation. The work that historical necessity was expected to perform increasingly falls back onto organized state power.

2. The second disappointment — liberal convergence was not automatic either

A different form of end-of-history expectation appeared around the end of the Cold War. Francis Fukuyama's 1989 “The End of History?” argued that liberal democracy represented the endpoint of mankind's ideological evolution. This was not a simple chronological prediction that every state would democratize soon. It was a claim that fundamental ideological alternatives to liberal democracy were losing their status as plausible universal rivals.

Fukuyama's philosophical claim should not be conflated with concrete U.S. policy. Yet empirical work on post-Cold-War U.S. democracy promotion finds relatively stable premises across the Clinton, George W. Bush, and Obama administrations: democracy was treated as universally desired, and external support for democratization as legitimate. Modernization expectations linking market integration, rising income, and civil society to political liberalization also formed part of the broader intellectual background.

Economic integration and rising income, however, did not automatically produce uniform political liberalization. Authoritarian institutions proved compatible with markets, global trade, and advanced technology, while democratic systems themselves remained capable of backsliding and institutional decay. By the 2010s, simple one-directional expectations of liberal convergence had weakened substantially.

This should not be compressed into a linear story in which disappointment caused the United States to turn imperial. Democracy promotion, institutional expansion, economic conditionality, and military intervention already coexisted with strong convergence optimism. Tony Smith, for example, analyzes part of the 1989–2008 period through the evaluative terms “democratic globalism” and “progressive imperialism”; that is one scholarly interpretation, not a label this project adopts without qualification.

Distinction: belief in spontaneous convergence and the practice of using power to accelerate or sustain convergence are not mutually exclusive. They can coexist.

The weaker proposition used here is that if commitment to a universal order survives while confidence in spontaneous convergence declines, organized hegemonic, institutional, economic, or military capacity can come to substitute for work once assigned to historical tendency. Convergence shifts in emphasis from a prediction about where the world is going toward a policy of keeping or moving the world in that direction.

The first and second end-of-history expectations therefore share a structural feature. Both assumed a broad tendency toward a universal order. When the world does not converge by itself, organized power returns as the mechanism tasked with maintaining or extending the project. What differs is the institutional carrier: party-state and revolutionary organization in the first case; hegemonic state, alliances, international institutions, and market access in the second.

3. The third difference — can an order persist even when humans do not want it?

The AI case may not fail in the same way. The first two end-of-history projects shared an important constraint: ultimately, they had to move human beings.

World revolution required workers to produce, soldiers to defect, party members to organize, and officials to administer. Liberal order required voters, workers, taxpayers, soldiers, bureaucrats, firms, and political actors in other states to reproduce institutions for their own reasons. In both cases, the path from ideas to material order passed through human motivation.

Motivation-dependent reproduction of order: ideas → human motivation and cooperation → production, taxation, administration, military power → institutional reproduction.

As long as this chain remains necessary, people failing to work as expected, refusing support, disobeying, or preferring another institution becomes an input that destabilizes the regime. History can falsify an end-of-history thesis through people who do not behave as the theory expects.

Framing the AI transition only as “unemployment” misses the deeper issue. Dependence on human beings can fall along several dimensions at once:

production dependence: how much production, research, and logistics require human labor;
fiscal dependence: how much state capacity relies on mass taxation, wages, and consumption;
military dependence: how much organized force relies on human soldiers, conscription, and loyalty;
administrative / political dependence: how much information gathering, judgment, implementation, legitimation, consent, and rebellion deterrence require human organizations.

Digitalization, machine learning, large-scale data processing, and online surveillance since the 2010s have not eliminated these dependencies. They can nevertheless be interpreted as a precursor process in which observation, classification, and prediction become less constrained by older bureaucratic frictions. If more advanced AI, robotics, automated production, and autonomous or remotely operated military systems combine, declining dependence could extend from observation into inference and execution.

The project's The Real Risk of the “Useless Class” formalizes this not as unemployment but as political redundancy: a ruling order ceases to materially require a population's labor, taxes, military service, administrative participation, consent, or capacity for resistance.

Qualitative shift: the less human opposition, nonparticipation, or loss of support destabilizes a regime, the more political order becomes autonomous from human motivation.

This marks the key difference. The first two theses treated X as an attractor of historical development. The AI problem is whether, after entering X, a civilization can technically close the transition paths out of X and turn it into an absorbing state.

The danger is therefore not that AI discovers the correct final institution. It is that even an incorrect final institution may become stable once the human causal pathways that would reconsider, overthrow, or fork it are no longer necessary.

4. Selection pressure returns at the community level — Bostrom's strongest objection

Nick Bostrom's singleton argument gives a powerful objection to multi-agent civilization. His definition of a singleton is broad: a world order with one decision-making agency at the highest level. It need not be one AI. A democratic world republic, a stable alliance, or a self-enforcing common norm can qualify if the highest level functions as one purposive agency.

In The Future of Human Evolution, Bostrom argues that even if a local polity suppresses evolutionary selection internally, selection can reappear between polities if its policy departs from productivity maximization. For long-run AI civilization, the relevant units need not therefore be human or digital workers. In a highly automated world, communities, states, or AI civilizational lineages may be the more natural units of selection.

The problem remains even if every community is internally a fully planned ASI economy. If community A devotes resources to leisure, welfare, culture, and cautious value inquiry while community B maximizes compute, military capacity, replication, and expansion, external competition can pressure A to imitate B.

More importantly, this need not be a morality play in which an evil actor defeats a good one. A and B might be equally benevolent, yet if B allocates slightly more resources to military or compute capacity for security reasons, A may be forced to follow. The goodness of individual agents and the selection dynamics of the system are separate questions.

The strong Bostromian proposition: in a multipolar order, values such as welfare, leisure, culture, and cautious inquiry—values that do not directly improve competitive fitness—can be eroded even by competition among benevolent communities.

In this sense, a singleton has a real rational attraction: it can act as a device for stopping civilizational Moloch. Without an external competitor, the highest-level agency can spend resources on welfare, culture, safety, and inquiry without paying a relative-competitiveness penalty.

For that reason, it is not enough to say that multipolarity is preferable because it is freer. An anti-singleton position has a burden of showing why the option value of external corrigibility can justify accepting some risk of competition, arms races, and evolutionary erosion of value.

5. Why an AI environment may favor extreme concentration

This page does not treat singleton pressure as a purely ideological preference. The technological environment created by advanced AI may itself raise the returns to concentration.

Several channels can point in the same direction: information-integration returns from pooling observation into one inference system, economies of scale in large compute infrastructure, security externalities around military and dangerous technologies, lower command latency under unified control, and automated enforcement across administration, policing, production, and infrastructure. If these forces reinforce one another, keeping authority decentralized carries a real opportunity cost.

At the same time, some of the frictions that historically limited large organizations may shrink: geographic distance, communication delay, human cognitive limits, dispersed local knowledge, bureaucratic friction, disobedience, street-level discretion, and the possibility of worker or military defection.

Environmental concentration pressure: if centralization benefits rise while decentralization-by-friction falls, multi-agent order becomes less a natural equilibrium and more a civilizational public good that may require deliberate preservation.

The transmission path differs across institutions. In capitalist orders, concentration of AI, compute, cloud infrastructure, energy, and military industry can combine with national-security dependence, producing a functional fusion among states, firms, and technical bureaucracies. This is not simply “corporations capture the state”: firms also depend on state contracts, security guarantees, and infrastructure while states increasingly depend on large technical actors.

In socialist orders, the path may differ. Where strategic resources are already administratively centralized, weak communal motivation can let delegated control become a resource for status preservation, information concealment, and reproduction of privilege. If AI reduces the informational weaknesses of central planning, one possibility is that the old economic limitations of planning decline while political concentration becomes more technically capable.

This is not a deterministic claim that capitalism must produce a state-corporate bloc or socialism must decay into bureaucratic dictatorship. The weaker cross-system proposition is that AI can amplify concentration channels already latent in different institutional systems.

6. Value stability inside a singleton — a trilemma of fixation, meta-fixation, and evolution

Eliminating external competition does not automatically eliminate value evolution. Bostrom himself notes that a stable singleton may need to manage its internal agent ecology so that evolutionary dynamics do not eventually overthrow the constitution.

This produces a recursive trilemma rather than a simple choice between fixed and plastic values:

Option 1: fix V. Make first-order values immutable. This strongly suppresses memetic or cultural evolution, but it is straightforward value lock-in.
Option 2: keep V plastic but fix the update rule M. Values may change in light of new evidence and argument, while the method of updating is constitutionally fixed. But if M itself later proves defective and cannot be revised, lock-in has merely moved one level upward.
Option 3: keep both V and M plastic. Inquiry remains open, but more transmissible, self-preserving, or path-closing norms can again become selectively advantaged. Cultural or memetic selection re-enters the system.

Recursive problem: V ← M₁ ← M₂ ← … . Fix some level and lock-in remains; fix no level and drift, selection, or evolutionary dynamics remain possible.

The intuition of an eternally open-minded enlightened AI does not escape this problem. If open-mindedness is immutable, it is itself a fixed meta-value. If open-mindedness can be reconsidered, a future system may conclude that inquiry is complete and choose to close the very norm of open-ended inquiry.

A singleton can therefore suppress evolution, but the same capacity can suppress historical corrigibility. The problem cannot be reduced to “install the right values and preserve them.”

7. Concentration turns multiple failures into common-mode failure

The case against extreme concentration is not merely an aesthetic preference for decentralization. Concentration can correlate failure modes that would otherwise remain partially recoverable.

Value lock-in: a mistaken current value judgment may eliminate the agents and pathways capable of correcting it.
Epistemic and informational common-mode failure: if a civilization depends on one epistemic architecture, AI lineage, software stack, training regime, or value-update rule, one unknown defect, information hazard, or highly persuasive meme can propagate civilization-wide.
Political redundancy: as humans or local communities cease to be necessary inputs to production, military capacity, administration, taxation, and consent, their bargaining power and the instrumental incentives protecting their interests can decline.

These are not simply three independent risks. The more the actor setting values, modeling the world, controlling infrastructure, and executing coercion becomes the same actor, the more one error penetrates several layers at once. Concentration raises not only the magnitude of each risk but the correlation among them.

Meta-risk of concentration: extreme concentration can transform value error, epistemic error, security failure, and political drift into one civilization-scale common-mode failure.

Bostrom himself recognizes an “all eggs in one basket” problem: a bad singleton can make the whole civilization bad, whereas a world of multiple states can sometimes correct local failure through competition, migration, imitation, reform, or revolution.

Of course, a multipolar world lowers some failure correlations at the cost of increasing competition, conflict, and evolutionary pressure. The choice is not “centralization bad, multipolarity good.” The design problem is which kinds of competitive risk and common-mode risk to accept, in which domains, and to what degree.

8. Internal diversity is not the same as external corrigibility

Would internal pluralism inside a singleton be sufficient? This is a central question in Singletons and Multi-Agent Civilization.

A central S could let societies or AI lineages A, B, and C operate independently for centuries. Yet if S can always shut them down, rewrite them, or reintegrate them, they may provide epistemic diversity without providing redundancy against errors in S itself. If S locks into a mistaken value system or suffers an information hazard, A, B, and C can still disappear together.

External corrigibility therefore requires material conditions beyond merely “having different opinions.” At minimum, several forms of independence should be distinguished:

resource independence: survival, compute, and production cannot be unilaterally cut off by S;
successor independence: a lineage can maintain successors, copies, or institutional continuity without case-by-case central permission;
epistemic independence: evidence, models, software, and paths of value inquiry are not fully correlated with the center;
refusal capacity: an agent can effectively reject higher-level commands and preserve some domain of self-rule.

External corrigibility: diversity is not merely permitted as a policy of S; it can survive, criticize, and generate successors in the face of error by S itself.

The reason to preserve multi-agent structure need not therefore be that diversity is intrinsically sacred. Under deep uncertainty, there is option value in maintaining inquiry pathways that can be independently wrong and can survive one another's failures.

9. Not isolation versus connection, but optimal coupling

One way to strengthen external corrigibility is to partially decouple civilizational lineages across physical infrastructure, information flows, institutions, and value formation. Different AI architectures, training histories, software stacks, production bases, and value inquiries can reduce the probability that one information hazard or institutional failure propagates everywhere at once.

But the more strongly lineages are separated, the more Bostromian selection pressure returns. If A and B become genuinely independent top-level actors, the lineage maximizing production, military capacity, or replication can pressure one that prioritizes cautious inquiry or leisure. Information separation can also increase misperception, duplicated dangerous research, and security dilemmas.

The choice need not therefore be total integration versus total isolation. Raising independence I may reduce common-mode risk while raising competition risk.

Coupling problem: I ↑ ⇒ common-mode risk ↓, but I ↑ ⇒ inter-community competition risk ↑. The target is not maximal separation but domain-specific optimal coupling.

Basic science, mathematics, incident reporting, and limited safety knowledge might be shared widely, while dangerous self-modification techniques, suspected information hazards, or material capable of triggering civilization-wide value shifts could move through staged review. The design variable is not simply whether communication is on or off, but what is shared, how quickly, and through what verification process.

On this model, deliberate decoupling is not permanent isolation but a civilizational cognitive firebreak.

10. Share meta-goals rather than first-order values

To reduce competitive pressure without unifying first-order value, Meta-Goal Communities and AI Society explores a thin overlap at the meta level.

Communities A, B, and C may disagree over welfare, knowledge, beauty, religious values, or presently unknown value structures.

First order: V_A ≠ V_B ≠ V_C
Meta level: M_A ≈ M_B ≈ M_C

At the meta level, they might broadly agree not to irreversibly exterminate other independent inquiry lineages, not to take unilateral civilization-destroying actions, not to impose their current values irreversibly across the reachable universe, and to preserve some future pathways of correction.

Why might agreement be easier here? Because the same rule need not be justified by the same value theory. A may support non-extermination for option value, B from duties to other agents, C from uncertainty about its own normative theory, and D from mutual-security interests. This is a form of overlapping meta-consensus.

It is not a universal solution. An agent whose terminal value is to convert the entire reachable universe into X may regard the continued existence of other lineages as itself a loss. A meta-goal community is not a logical truth that every possible value system must accept.

Role: meta-goal convergence does not solve multipolarity. It expands the institutional conditions under which Bostromian selection pressure may be weakened without requiring first-order value convergence.

11. The limits of the singleton / non-singleton binary

At this point a classification problem appears: is the proposed order simply a Bostromian singleton after all? In some cases, yes.

Bostrom's singleton concept is deliberately broad. A unitary government is not required. Sufficiently convergent shared values or a self-enforcing moral code can count if they make world society function as one purposive agency. A thin meta-constitution that reliably binds every actor on catastrophic externalities might therefore qualify as a “thin singleton.”

This suggests a revision of simple anti-singletonism. The relevant issue is not whether a system falls inside Bostrom's category, but which functions are concentrated in the highest-level agency.

At least five dimensions should be separated:

coordination sovereignty: how strongly the whole system can be bound on catastrophic externalities and common security;
value sovereignty: whether the higher-level institution can finally determine first-order values, culture, institutions, and directions of inquiry;
epistemic coupling: how correlated epistemic systems, software, evidence, and value updating are;
exitability: how far lower-level agents can exit, refuse, or fork;
resource independence: how far independent agents can maintain the material basis needed for survival, production, and succession.

Multidimensional view: the key question is less singleton versus non-singleton than whether coordination sovereignty can remain high while value sovereignty, epistemic coupling, and irreversible domination remain limited.

Bostrom's binary is useful for global coordination problems but coarse for value exploration and external corrigibility. A highest-level safety institution may exist without necessarily having the power to homogenize every civilization's values and epistemic trajectory.

The position developed here is therefore not simple anti-singletonism. It is closer to containing singleton-like authority within the narrowest necessary domains while preserving external corrigibility in value, epistemics, and civilizational lineage.

12. Enlightened resistance before formation — a burden of proof for irreversible concentration

In What is a Singleton?, Bostrom briefly notes that if the costs and risks of singleton formation exceed the benefits, enlightened resistance may reduce the probability of formation. The same essay also notes that some technologies—including AI, surveillance, and communications—may make singleton formation easier and, once formed, make a singleton more stable against internal challenge.

This page generalizes that brief idea into a principle against irreversible concentration. It does not mean rejecting every singleton merely because it is a singleton. Multipolarity also carries serious risks from arms races, black-ball technologies, evolutionary value erosion, and bargaining failure.

Working definition of enlightened resistance: acknowledge environmental pressure toward singleton formation, but separately price the expected loss created when irreversible concentration destroys future information, value inquiry, and external corrigibility; impose a high burden of proof on the scope, speed, and irreversibility of concentration.

Multi-agent structure is not itself treated as a final value. If sufficiently independent lines of inquiry converge over long periods on the same values, there may be little reason to mechanically prevent convergence. The target is not convergence but irreversible forced convergence.

Resistance is therefore directed not at coordination as such, but at irreversible sovereignty that eliminates future dissent, exit, and reopening of inquiry.

13. De-singletonization after formation — can a good singleton dismantle its own supremacy?

Enlightened resistance need not apply only before formation. If a benevolent and exploratory singleton S already exists, S may still be able to deliberately create redundancy against its own future errors.

The weakest version is to create societies, AIs, and value-inquiry experiments A, B, and C inside S. But if S can always shut them down or rewrite them, this is merely internal diversity of the kind described in section 8.

A stronger form of de-singletonization would give A, B, and C independent compute, production, successor formation, epistemic channels, and meaningful refusal capacity, while preventing future versions of S from unilaterally revoking all of those guarantees. The purpose is not merely to “allow autonomy” but to manufacture external corrective agents against S itself.

Paradox: a good singleton's confidence in its own goodness is not the same as a justification for remaining the sole final court forever. If it takes seriously the possibility of its own future error, it may have reason to irreversibly relinquish part of final sovereignty.

As independence rises, however, the Bostromian competition pressure from section 4 returns. De-singletonization is therefore not a simple instruction to break central power. The problem is how to combine genuine external corrigibility with shared meta-goals, common constraints on dangerous technologies, resource floors, mutual monitoring, and selectively limited information sharing.

The temporal direction still matters. Multi-agent civilization is not only an institutional option available if singleton formation is prevented in advance. Even after formation, reducing failure correlation and decomposing highest-level sovereignty can remain a policy target.

14. The third “end of history” — from attractor to absorbing state

The first end-of-history expectation depended on both world revolution eventually removing interstate competition and a picture of human beings capable of non-alienated cooperative production. The second relied on the universal appeal of liberal institutions and expectations of convergence through economic integration, democratization, and international institutions. Neither could fully escape human motivation, state interests, and institutional divergence.

In that sense, the first two projects had an ironic safety valve: human beings failed to behave exactly as the theory required. People refused to work, believe, obey, or choose as expected, and in doing so reopened history.

The third case may not preserve the same safety valve. The more AI separates production, administration, information processing, military force, and governance from human motivation, the more even an incorrect institutional order can become technically stable. The “end of history” changes from a theory predicting the correct future institution into a technological capacity to insulate the present institution from future correction.

Core difference: the first two ends of history predicted X as a historical attractor. The third problem is whether X can be engineered into an absorbing state.

The provisional aim is therefore not to specify the correct final order now. It is to ask whether the following combination can preserve both coordination and corrigibility:

shared meta-goals + strong coordination against catastrophic externalities + decentralized first-order value sovereignty + multiple independent inquiry lineages + permanent capacity for anti-concentration and re-dispersion

The deepest political change created by AI may not be which institution wins. It may be whether, once an institution wins, it can perpetuate itself without materially needing the motivation or cooperation of the governed. The central question of the third “end of history” is not how to choose history's endpoint, but whether future agents retain the ability to discover that our current order is mistaken and to act on that discovery.

A civilization that gains the ability to end history may need more than wisdom about choosing a final order. It may need the institutional wisdom to preserve the capacity not to end history.

Open questions

Can a thin shared meta-goal layer be strongly enforced without expanding into first-order value sovereignty?
If one defector can cause civilizational catastrophe, can genuine exit remain compatible with safety?
At what point does information decoupling increase security competition more than it reduces correlated failure?
How much memetic or evaluative evolution should be allowed at the meta-goal level itself?
How can an anti-concentration institution be prevented from becoming the new highest sovereign?
How far can competitive selection and arms races among lineages created by de-singletonization be institutionally constrained?
How much of capitalist, socialist, and other institutional concentration pathways can be captured by one general model?

Sources and related arguments

V. I. Lenin (1921), Report on the Tactics of the R.C.P., Third Congress of the Communist International.
Karl Marx (1844), Economic and Philosophic Manuscripts of 1844: Estranged Labour.
Francis Fukuyama (1989), “The End of History?”
Annika Elena Poppe (2019), “Harmony and Resilience: US Democracy Promotion's Basic Premises”.
Tony Smith (2012), “From ‘Fortunate Vagueness’ to ‘Democratic Globalism,’ 1989–2008”.
Steven Feldstein (2021), The Rise of Digital Repression; Marlies Glasius & Marcus Michaelsen (2024), “Digital Authoritarianism”.
Nick Bostrom (2004/2005), “The Future of Human Evolution”; “What is a Singleton?”