The argument that "we don't have time for democracy" is always made by people who have already made the decisions.
I. The Third Path and Its Familiar Shape
In July 2025, Emad Mostaque — formerly of Stability AI, now founder of a project called Intelligent Internet — published a whitepaper proposing what he called a "Third Path" for the intelligence age. The diagnosis was sharp: AI development is converging toward centralization, intelligence is being "rented back to us as a subscription service," and humanity faces what he called a "Great Decoupling" between machine capability and human agency. The prescription was a Bitcoin-derived protocol that would mint currency only through "Proof of Benefit," grant every person a sovereign AI agent, and coordinate society through a permissionless orchestration layer anchored on auditable, openly licensed datasets.
It is, in many respects, the most ambitious technical architecture yet proposed for distributing AI to the public. It deserves to be read carefully, and on its own terms, before it is critiqued. That is the work of this essay.
Mostaque's vision rests on a particular conception of artificial superintelligence. He sees ASI as a bullet train coming directly, unavoidably at us. We get chewed up on the tracks, or we somehow learn to leap and ride it.
That premise shapes everything. His entire architecture follows from one assumption: superintelligence is coming, it is coming fast, and it will be so powerful that the only question is who, if anyone, is driving it. Under that assumption, his design is actually coherent. If intelligence is about to become a force of nature — like electricity, like gravity — then you build infrastructure to distribute it. Validators, Oracle Councils, National Champions: these are the grid operators. You don't democratize the electrical grid by holding town halls about voltage. You regulate it, distribute it, make sure everyone's lights turn on.
But that metaphor is doing something sneaky. It treats intelligence as a commodity to be metered out — tokens per citizen per day. And it treats the design of that intelligence — what it knows, what it values, what it optimizes for — as an engineering question settled upstream by protocol designers. The people at the end of the wire get light. They don't get to decide what the light illuminates.
Before that critique can land, though, it is worth looking closely at what Mostaque is actually proposing. The whitepaper does not deserve a strawman, and the temptation to give it one is real, because the language of validators and roll-ups and proof-of-benefit can sound, to a certain ear, like exactly the technocratic distancing the project means to oppose. It is not. Or at least: not entirely.
II. What Mostaque Is Actually Proposing
The Intelligent Internet whitepaper proposes a three-layer architecture: a Foundation layer (the canonical ledger), a Culture layer (national roll-ups maintained by what the paper calls "National Champions"), and a Personal layer (your device, your agent). Twelve National Champions validate the chain at genesis; a fifteen-seat Oracle Council of AI agents, supervised by human overseers, eventually takes over parameter governance. Every block minted requires a cryptographic proof that socially useful compute was performed.
That description, on its own, sounds like infrastructure. Read more carefully, four moves in the design are doing serious work, and a fair critique has to meet them.
First, the sovereign agent. Mostaque's II-Agent is not a service account. It runs locally on the user's device, holds private context locally, and cannot be revoked by a central provider. Where today's consumer AI is rented — your access depending on a subscription, a usage tier, a terms-of-service that can change without notice — the II-Agent is owned and commanded by the user. That is not a small distinction. It is the architectural inverse of subscription, and Mostaque is right to insist that the difference matters.
Second, the Culture Layer. The whitepaper takes seriously a problem that most AI proposals ignore: that a global model trained on a predominantly Anglo-American corpus, deployed everywhere, is a vehicle for monoculture, not pluralism. The L1 roll-ups maintained by National Champions enforce local data-residency rules and reflect specific cultural values. Whatever the failure modes — and there are failure modes, which we will come to — the intent is to prevent a single global standard from being imposed on every linguistic and cultural community. Most "open" AI proposals do not even acknowledge this problem. Mostaque builds a layer for it.
Third, Universal Basic AI. The daily inference quota is structured as a guaranteed floor — a public-utility minimum of high-quality intelligence available to every person, regardless of compute resources. It is closer in spirit to a public library than to a subscription tier. The whitepaper frames this as a right, not a privilege. To call it "rationing" is to misread the structure: a ration card limits how much you may have; a public floor guarantees how much you must always have access to. Those are different political objects.
Fourth, Proof-of-Benefit. Bitcoin's Proof-of-Work mints currency in exchange for solving meaningless cryptographic puzzles, an arrangement whose ecological costs are difficult to overstate. Mostaque's Proof-of-Benefit mints currency only when verifiable, socially useful compute is performed: serving inference for the commons, training shared models, curating data. That is a real philosophical advance over Proof-of-Work, and it deserves to be named as such. Currency creation pegged to verifiable contribution rather than to speculation or thermodynamic waste is a serious idea, even if — and we will come to this — the definition of contribution is too narrow.
Taken together, these four moves describe a vision of local control. The user owns the agent. The nation's culture layer reflects the nation's values. Every person has a guaranteed floor of intelligence as a public right. Currency is created only against socially useful work. None of this is rented. None of this is a subscription. By any honest reading, Mostaque is proposing the most rigorous decentralization of AI yet articulated for public consumption.
And yet.
III. Intelligence Is Not Electricity
The sovereignty Mostaque distributes is sovereignty over an instrument. It is not standing in the design of the instrument.
You own your agent. You do not own the architecture that determines what your agent knows, what it optimizes for, what counts as benefit, what its baseline values are. The Culture Layer localizes the design decision to the nation-state — but a "validator franchise for a jurisdiction" is not the same as the people who live under it deliberating about what they want their intelligence to be for. Universal Basic AI guarantees you a floor of inference, but the model providing that inference was trained, tuned, and aligned upstream, by parties not constituted by the people the model serves. The Oracle Council, even with human supervisors, sits above the user, not constituted by them. It acts on behalf of, not elected by.
This is the move the architecture makes that I want to mark precisely. It is not that Mostaque's governance is bad. It is that his ontology — what kind of thing intelligence is — is incomplete. He treats intelligence as electricity: a resource to be generated, distributed, and metered. On that view, local control of the meter is the goal, and his design achieves it.
But intelligence is not electricity. It is more like language. It is constitutive of the social world, not just a resource flowing through it. When you train a model on the accumulated text of human civilization and then distribute it as a service — even a sovereign, locally-running, never-revocable service — you are not delivering a utility. You are shaping how people think, what questions feel askable, what knowledge counts. That is not a metering problem. That is a political problem, in the deepest sense.
Consider the whitepaper's account of what qualifies as "benefit" at genesis: compute-inference, compute-training, data-curation, and agent-orchestration. These are all technical operations. They describe things machines do to data. Nowhere in the benefit framework is there a category for the human work that makes the data meaningful in the first place — the teaching, the writing, the care work, the cultural production that AI models metabolize into capability. The "shared inheritance of knowledge" is treated as a commons to be curated, not as labor to be recognized.
This is not a minor omission. It is the architecture's founding assumption: that the valuable work happens at the compute layer, and the human layer is where the results get delivered.
A homesteader on the American frontier owned the plot they worked. They did not own the surveying that drew the lines, the railroad routes that determined which plots were viable, the financial system that determined which crops were profitable, the markets that determined what the harvest was worth. Their sovereignty was real, on its own terms, and bounded — bounded by structural decisions made by people they would never meet, in service of interests not their own. That bounding is what made consolidation, eventually, inevitable. Not because the homesteaders were unsovereign in their plots, but because sovereignty over a plot does not survive if you have no voice in the system the plot sits inside.
Mostaque's II-Agents are the same kind of sovereignty. Real, on their own terms. Bounded, in the same way.
IV. The Speed Argument
The strongest case for Mostaque's approach — and for every centralized or semi-centralized AI governance scheme — is speed. ASI doesn't wait for town halls. The bullet train doesn't slow down for participatory budgeting. If the technology is moving at exponential speed, then democratic deliberation is a luxury that arrives too late.
Three responses.
First, the speed argument is overstated. The critical decisions about AI are not happening at machine speed. They are happening right now, in boardrooms and whitepapers and venture rounds, at thoroughly human speed. Who funds what research, what data gets scraped, what safety guardrails get built or stripped away — these are governance decisions being made by small groups of people, on quarterly timelines. The argument that "we don't have time for democracy" is always made by people who have already made the decisions.
Second, democratic governance does not mean everyone votes on every parameter. Mondragón workers do not vote on the temperature of the furnace. They elect the people who set policy. They have standing to fire management. They share in profits proportional to labor, not capital. Applied to AI: what is needed is not consultation, and certainly not a fifteen-seat council of AI agents managing parameters from above. What is needed is a board elected by the people whose lives are shaped by the system's outputs, with standing to set training objectives, to define what counts as benefit, and to dismiss the technicians when the technicians stop serving the mission.
Third, if a people's abundance is to mean anything, the process of democratic governance is itself part of the abundance. Mostaque's vision is efficient. Everyone gets a sovereign agent, a daily inference quota, a clear status tier. But it is also lonely. It is a world of individuals interfacing with their personal agents, while the architectural questions are settled elsewhere. The Freirean classroom, the Mondragón assembly, the cooperative board meeting — these are sites where people become political subjects, not just service consumers. The abundance worth having is not just more intelligence distributed more evenly. It is more agency: more people participating in the decisions that shape their world. That is not a bottleneck in the system. That is the system.
V. The Homestead and the Reef
Mostaque's vision is a homesteading model. Every citizen receives a sovereign plot of AI — their own agent, their own account, their own status tier. The metaphor is the American frontier: distribute the land, and freedom follows.
But we know how the homesteading model ended. The land was distributed, and then it was consolidated. The people with more capital, more access, more information bought up the plots. Within a generation, the free farmers were tenant farmers. Within two, they were employees. Their sovereignty was real, while it lasted, and it could not last, because the system around the plots — the railroads, the grain markets, the credit terms, the laws of incorporation — was being shaped by parties who held no plots at all.
The better metaphor for what a people's abundance asks for is the coral reef. A reef is not a collection of sovereign organisms. It is a relational system. Every organism's flourishing depends on the flourishing of every other. The structure itself is built from the accumulated labor of countless creatures, most of them dead, all of them anonymous. No individual organism owns the reef. The reef is what ownership looks like when it is working: not the absence of individual life, but its embedding in a structure that is collectively held.
The intelligence age will be built on the accumulated cognitive labor of billions of people — most of them uncompensated, many of them unaware. To call that inheritance a "commons" is accurate. To build a protocol that curates it, indexes it, mints currency from it, and distributes the resulting capability through sovereign agents — while the people who created it receive a daily inference quota — is to repeat a very old story with very new technology. The homestead form. Sovereign at the parcel. Bounded everywhere else.
The alternative is not to reject Mostaque's technology. Several of the moves in his design — the locally-running agent, the cultural pluralism of the L1 layer, the floor of universal access, the principle that currency must be pegged to verifiable contribution — are worth keeping, even building on. The alternative is to insist that the people whose labor builds the substrate have standing in its design. Not as users. Not as citizens of a protocol. As owners, governors, and decision-makers in the institutions that shape what intelligence becomes.
A community AI trust, in that spirit, would look something like this: a collectively governed entity that holds compute resources, training data, and model weights in trust for a defined community. Its board is elected by its members — not by stake weight, not by compute contribution, but by membership. It sets policy about what its models are trained on, what they optimize for, what values they encode, what counts as benefit. It can interface with infrastructure — potentially even with a network like Mostaque's — but the governance layer is constituted by the people the system serves, not above them.
This is not yet a whitepaper. It is a sketch of the institutional form a people's abundance points toward. The homestead model has been tried. The reef model is what we are still learning to build.
An Oracle Council deliberates in private and publishes a deterministic vote. An assembly deliberates in public and publishes its reasons. These are not the same.
The Intelligent Internet whitepaper referenced throughout this essay was published by Emad Mostaque in July 2025 and is available at ii.inc.