#198 - The New Covenant: Individual Sovereignty and Collective Ethics in the AI Era

Introduction

Over the past year, I have thought about almost nothing but artificial intelligence. I have built products with it, integrated it into every layer of my daily work, run local models on my own hardware, trained fine-tunes on proprietary data, and used it to run businesses, draft legal documents, analyze financials, and write code. I am, by any practical measure, AI-native — not because I adopted the label, but because the technology is now woven into the fabric of how I think, create, and operate.

I am also unapologetically pro-AI. I believe this technology has the potential to elevate every human being on the planet — to make us smarter, more productive, more creative, and more free. I have seen it happen in my own life. I have seen it happen in the lives of employees I have trained and colleagues I have worked alongside.

But being pro-AI does not mean being uncritical. It does not mean pretending that every application is beneficial, that every corporate incentive is aligned with human flourishing, or that the technology will sort itself out if we just let the market run. Being pro-AI, for me, means caring deeply about who controls it, how it is governed, and what we build with it — because those questions, not the technology itself, will determine whether this era produces the greatest flourishing in human history or the most sophisticated apparatus of control ever assembled.

I have come to call the answer I am arguing for a covenant, and I should say plainly what I mean by the word — because I mean something more specific than a slogan. A contract is a transaction: it binds two parties for as long as the exchange serves them both, and it dissolves the moment one party finds defection more profitable than performance. A covenant is different in kind. It is a binding, reciprocal obligation that holds even when opting out would pay — an agreement about how power will be used, entered into precisely because the temptation to abuse that power is real and permanent. We do not write covenants for situations where everyone's interests already align. We write them for situations where they do not, and where the cost of defection falls on whoever holds the least power.

The new covenant I am describing is not the renewal between private citizens or between government and her people or between a nation and her God — these are separate and older ideas, and others have written about them. It is narrower and more practical: a reciprocal obligation between those who hold the power of intelligence and those who are subject to it. It binds the builder to the user, the owner to the worker, the network to the individual, the state to the citizen. Its terms are simple — that this power will be dispersed rather than consolidated, made transparent rather than hidden, and directed toward the sovereignty and dignity of the individual person rather than their management and control. What makes it new is not the principle, which is as old as the question of whether power serves the person or the person serves power. What is new is the scale of the power now in play. Every prior covenant between the strong and the weak assumed that power operated at human speed and human reach. AI breaks that assumption. The covenant has to be rewritten for a tool of superhuman scale — and it has to be rewritten now, before the defaults harden.

This essay represents the sum total of my thinking on where we are as a civilization at this inflection point. I will begin with a few foundational premises — premises I hope most thoughtful readers will find difficult to debate — and build outward from there into the specific debates, tradeoffs, and decisions that will shape the next decade. At the end, I will lay out the possible futures this technology can lead us toward, and offer recommendations for policymakers and business leaders.

I. Three Premises

Before we can reason together about what artificial intelligence means for our businesses, our families, and our civilization, we must first agree on what it is — and what it is not. Three premises form the ground beneath everything that follows.

Premise One: The cat is out of the bag.

Artificial intelligence is not a prototype in a laboratory. It is not a government secret waiting for declassification. It is a printed book — distributed, copied, forked, and running on machines from billion-dollar data centers to a teenager's laptop. As of mid-2026, there are thousands of capable open-weight models freely available on the internet. Llama, Mistral, Qwen, DeepSeek, Phi, and dozens of community fine-tunes can be downloaded, modified, and deployed by anyone with a consumer GPU. This genie does not go back in the bottle. Any regulatory framework that begins from the fantasy of un-inventing AI is not a framework — it is a delusion. The technology is here. It is proliferating. And it will continue to proliferate regardless of what any single government, corporation, or coalition decides.

This is not a pessimistic premise. It is a realistic one. And it is the foundation on which all constructive policy must be built.

Premise Two: Technology requires a moral framework — or it becomes the framework.

Every transformative technology in human history has demanded a moral architecture around its use. The printing press required new norms around authorship, accuracy, and accountability — and where those norms failed, religious wars followed. Nuclear fission required mutually assured deterrence, non-proliferation treaties, and international oversight — and where those structures were weak, the world came terrifyingly close to annihilation. The internet required new thinking about privacy, free expression, and the commons — and where we failed to build those norms, we got algorithmic addiction, surveillance capitalism, and the erosion of public discourse.

AI is no different. It will not govern itself. Left ungoverned by moral principle, it defaults to the incentives of those who control it. If those incentives are purely commercial, it optimizes for engagement and revenue. If they are purely military, it optimizes for lethality. If they are authoritarian, it optimizes for control. The technology is morally neutral — which is precisely why the moral framework around it matters so desperately.

The question is not whether we will impose a moral framework on AI. The question is who will impose it, whose values it will encode, and whether it will serve human flourishing or human subjugation.

Premise Three: AI is a fission event.

In 1938, Otto Hahn and Fritz Strassmann split the uranium atom — a result that Lise Meitner and Otto Frisch interpreted, weeks later, as nuclear fission. Within seven years, that discovery had produced both the atomic bomb and the conceptual foundation for nuclear power — the worst weapon humanity had ever created, and one of its most powerful energy sources. The same physics. The same knowledge. Diametrically different outcomes.

Artificial intelligence is the fission event of the twenty-first century. It is a single, general-purpose capability that can be directed toward radically different ends. It can diagnose cancer, accelerate drug discovery, and tutor children in rural villages. It can also design bioweapons, automate surveillance states, and execute sophisticated cyberattacks at a scale no human team could approach. The capability is the same. What differs is intent, governance, and the moral framework directing the energy.

Understanding AI as a fission event — rather than as inherently good or inherently evil — is essential. It frees us from the false binary of "ban it" versus "let it run wild." It directs our attention to the actual challenge: how do we build the institutional, legal, technical, and moral infrastructure to channel this energy toward human flourishing while limiting its capacity for destruction?

These are the hard questions. And they are the only questions that matter.

The Great Equalizer or the Great Enslaver

From these three premises, a fourth proposition emerges — not as a premise to be debated, but as a fork in the road to be chosen:

AI can either be the great equalizer between sovereign individuals and larger corporations and states, or it can be the great enslaver of the masses — used by fewer and fewer powerful groups to control, influence, and subordinate the many.

There is no middle path. A tool this powerful does not distribute itself democratically by default. It accumulates where capital accumulates, where compute concentrates, where gatekeepers establish themselves. If we do not deliberately construct the institutional, technical, and moral architecture that disperses AI's power broadly, that power will consolidate — swiftly, and almost irreversibly.

The choices that will determine which path we walk are not abstract. They are concrete and immediate: open vs. closed, transparent vs. opaque, and — most consequentially — what we are ultimately solving for. Are we solving for destruction (weaponization, surveillance, social control)? For revenue (engagement optimization, attention extraction, rent-seeking)? Or for human flourishing (health, education, dignity, liberty, the capacity of every individual to pursue their calling and build a meaningful life)?

This technology has never made more apparent the choice between individual sovereignty and privacy on one hand, and state-sponsored or corporate-sponsored security and control on the other. It is a question of governance as old as the Roman Republic — as old as Athens, as old as the first city-state that asked whether power belongs to the citizen or the magistrate, to the person or the institution. Now that ancient question is laid bare by the most powerful technology known to man.

AI can be used to support the sovereignty of the individual — their capacity to think, create, earn, and govern their own life without permission from any gatekeeper. Or it can be used to oppress the masses into submission to the state or corporate will, monitored and managed by algorithms they cannot see, cannot audit, and cannot challenge.

The choice will be ours — but only if we are deliberate about where we want this all to go. AI is here. It is not going away. So what are we to do with it?

II. Open Source vs. Closed: The Battle for the Soul of AI

The most consequential technical debate in AI today is not about parameter counts, benchmark scores, or training methodologies. It is about who gets to see the model.

For years, the AI world has been divided into two camps. The open-source camp — led by Meta's Llama series, Mistral, Alibaba's Qwen, and a vibrant community of independent researchers — publishes model weights, training data (or descriptions of it), and fine-tuning recipes publicly. The closed-source camp — led by OpenAI, Anthropic, xAI and Google — keeps model weights proprietary, offers access only through APIs, and increasingly layers behavioral guardrails on top.

As of 2026, the capability gap between open and closed models has largely closed. On the widely cited Chatbot Arena leaderboard, the lead of the best closed-weight model over the best open-weight model collapsed from roughly eight percentage points in early 2024 to about two by early 2025.¹ Open-weight models now deliver 80–85% of frontier model quality on most production tasks at zero marginal cost per request. The Llama 4 family, Mistral Large, and Qwen3 series compete with GPT-4o and Claude Sonnet on a wide range of benchmarks. For many business applications, the open models are not just adequate — they are preferable, because they can be self-hosted, customized, and audited.

But the closed-source argument has evolved. It is no longer primarily about capability. It is about control — and potentially, about censorship.

The Mythos Watershed

In April 2026, Anthropic announced that its Claude Mythos model was, in the company's own assessment, too dangerous to release to the public. The model had reportedly discovered vulnerabilities in every major operating system and web browser it was tested against. Anthropic launched "Project Glasswing," limiting access to approximately fifty vetted organizations.² The Nature journal described it as potentially "the start of the restricted-AI era."³

Then, on June 13, 2026, the U.S. Commerce Department issued an export control directive ordering Anthropic to suspend all access to Fable 5 and Mythos 5 by foreign nationals. Anthropic disabled the models for all users while it assessed compliance — meaning even domestic customers lost access overnight.⁴

This is a turning point. For the first time, the U.S. government has treated an AI model the way it treats a weapons system — restrictable, export-controlled, accessible only to approved parties.

There are legitimate national security reasons for this. A model that can autonomously discover zero-day vulnerabilities in critical infrastructure is genuinely dangerous in the wrong hands. But the precedent is fraught with peril. Because once you establish that AI models can be restricted — that they are controllable assets rather than published knowledge — you open the door to restriction for reasons that have nothing to do with security.

Political censorship. Ideological control. Protection of incumbent market position. The suppression of models that produce answers a government or corporation finds inconvenient.

Helen Toner, interim executive director at Georgetown's Center for Security and Emerging Technology, told Nature: "I would expect this to more be the first in a series rather than a one-off."⁵ OpenAI followed within a week, releasing its own cybersecurity-specific model (GPT-5.4-Cyber) to vetted researchers only.⁶

The trend is clear. And the question it forces is this: when models are restricted, who is doing the restricting, and in whose interest?

The Case for More Openness, Not Less

There is a real argument that certain capabilities — autonomous cyber-offense, bioweapon design, mass-scale social manipulation — deserve restricted access. Responsible stewardship is not the same as censorship.

But there is an equally real argument that the default posture of AI should be open. Here is why:

Closed models concentrate power. When the most capable AI runs only on the servers of three or four companies, those companies become the de facto gatekeepers of intelligence. They decide what questions can be asked, what answers can be given, and who gets access. This is not a market — it is a cartel with a content moderation team.

Closed models cannot be independently audited. When researchers cannot examine model weights, they cannot verify safety claims. They cannot test for bias, investigate emergent behaviors, or reproduce results. Science requires transparency. Closed-source AI is, in a real sense, alchemy — impressive results with no peer review.

Closed models encode the values of their creators without external check. Every closed model reflects the editorial decisions of its training team. Those decisions are invisible to users and unchallengeable by design. When a model refuses to answer a question, the user has no way to know whether the refusal is grounded in genuine safety concern or in corporate risk management.

The right framework is not "open everything" or "close everything." It is a layered approach: open weights for models below a capability threshold, graduated access and monitoring for frontier models with demonstrably dangerous capabilities, and robust international oversight for systems that approach or exceed state-level cyber-offensive capacity.

But the default must be open. Because the alternatives — intelligence controlled by a handful of corporations, or worse, by adversarial nation-states — are far more dangerous than the risks of openness honestly managed. Civil liberties organizations have begun to make precisely this case: that the openness of the underlying models is not a side issue but the central battleground for the future of the technology.³¹

III. The Power of Local Models

While the open-source debate plays out at the policy level, a quieter revolution is happening at the edge — in homes, small businesses, and on-premises servers around the world.

Local AI inference — running models on your own hardware, with your own data, without sending a single byte to a cloud provider — has crossed a critical threshold. In 2026, running quantized open-weight models on consumer hardware (a single GPU with 24GB of VRAM or even a modern laptop with a neural processing unit) delivers production-quality results for the vast majority of business tasks.⁷

A caveat is in order. Running capable models locally is still, in mid-2026, closer to the bleeding edge than the mainstream. The hardware that does it well — high-VRAM GPUs, the newest unified-memory machines — remains expensive and, thanks to persistent supply-chain bottlenecks in consumer GPUs, often hard to get at all. For most people today, local inference is an enthusiast's pursuit, not a turnkey default. But this is a cost curve, not a permanent ceiling. Every prior wave of computing followed the same arc: what begins as a costly specialty commoditizes as supply chains mature and manufacturing scales. As GPU bottlenecks normalize and purpose-built inference silicon reaches consumers, the price of cognitive self-governance will fall — and what is an enthusiast's rig today becomes a household default tomorrow.

The implications are profound:

Privacy becomes real. A law firm running a local model can use AI to draft contracts, summarize depositions, and research case law without any client data ever leaving their building. A medical practice can use AI to assist diagnosis without HIPAA concerns about cloud providers. A family business can use AI to analyze financials, draft communications, and automate workflows without training a third-party model on their proprietary data.

Dependency decreases. When you run your own model, you are not subject to API pricing changes, rate limits, content policy updates, or the sudden withdrawal of access (as happened with Mythos). You own your intelligence. You own your workflow. No one can turn it off.

Sovereignty extends to nations. Countries that cannot — or will not — rely on American or Chinese cloud infrastructure can build their own sovereign AI capacity using open models. This is already happening in the European Union, India, and across Southeast Asia.

The local model movement is, in essence, the democratization of intelligence. It means that the benefits of AI do not accrue exclusively to those who can afford enterprise API contracts. It means a solo practitioner in a rural area can access the same cognitive tools as a partner at a Manhattan law firm.

More than that, local models are the technical embodiment of individual sovereignty. When you run a model on your own hardware, you are not asking permission to think. You are not submitting your thoughts — your strategies, your client communications, your personal financial planning, your creative work — to a corporate intermediary whose interests may not align with yours. You are exercising, in the most literal sense, cognitive self-governance.

If AI is to be the great equalizer rather than the great enslaver, local inference is where equalization begins. It is the point at which the individual says: I will think with my own tools, on my own terms, and no one will have a kill switch for my intelligence.

This is not a niche technical trend. It is the mechanism by which AI remains a tool for the many rather than a service provided by the few.

IV. Distributed Compute: Breaking the Data Center Monopoly

If local models represent the individual frontier of AI, distributed compute represents the collective frontier — and it may be the most important development in the entire AI landscape.

Today, training a frontier AI model requires thousands of GPUs running in concert inside a single data center for weeks or months. The capital requirements — hundreds of millions to billions of dollars — limit this capability to a handful of well-funded companies and nation-states. This creates a structural monopoly: whoever controls the compute infrastructure controls the pace and direction of AI development.

It must be said plainly that this is early, bleeding-edge work — but it is no longer theoretical. Distributed training was, until very recently, widely assumed to be impossible at frontier scale; the coordination and bandwidth costs were thought prohibitive. That assumption is now breaking. The architecture is starting to take shape, and the results are beginning to arrive.

Distributed compute networks aim to break this monopoly by aggregating the world's idle and underutilized computing resources into decentralized training and inference clusters — an approach a growing body of research argues could power a genuinely open alternative to the hyperscaler model.³² Several projects have made remarkable progress:

Bittensor (TAO) is a decentralized AI network where independent participants contribute compute, models, and data in exchange for TAO token rewards. The network is organized into specialized subnets — each focused on a specific AI task like inference, training, or data validation. The proof that distributed training could work at scale arrived in late 2024, when Prime Intellect trained INTELLECT-1 — the first 10-billion-parameter model trained over a globally distributed network, spread across GPUs on three continents — while retaining better than 80% of the compute efficiency of a centralized run.⁸ Eighteen months later, the approach had scaled dramatically: in March 2026, Bittensor completed Covenant-72B, a 72.7-billion-parameter large language model trained entirely across 70+ anonymous, distributed nodes without a single data center. It was the largest decentralized AI training run ever completed, and it demonstrated that frontier-grade models can be built outside the hyperscaler infrastructure.

Macrocosmos is building permissionless, accessible pretraining infrastructure that harnesses distributed compute to train the largest decentralized models.⁹ Their approach connects decentralized compute, orchestrates distributed model training, and provides real-time insight into training progress — essentially creating a cloud-free alternative for large-scale AI development.

Pluralis Research takes a different approach with what they call "Protocol Learning" — low-bandwidth, model-parallel training and inference where no single participant ever holds the complete model weights.¹⁰ Their Unextractable Protocol Models (UPMs) shard the network so that no participant ever holds a coherent copy of the weights, making the model cryptographically unextractable — it can be trained and used, but never captured or carried off. Backed by a $7.6 million seed round from USV and CoinFund, Pluralis is advancing what they call "Actually Open AI" — models that are collaboratively built but cannot be captured or controlled by any single party.¹¹

Why does distributed compute matter beyond the technical achievement?

It prevents compute concentration from becoming intelligence concentration. If AI training requires a $10 billion data center, then only four or five entities on Earth can train frontier models. If it can be done across thousands of volunteered or incentivized GPUs around the world, then the frontier is no longer gated by capital alone.

It creates resilience. A model trained across a distributed network cannot be shut down by a single government order, a single corporate decision, or a single data center outage. This is the same logic that makes the internet and cryptocurrency blockchains resilient — and it applies to intelligence just as it applies to communication.

It aligns with American values. Distributed compute is inherently permissionless, meritocratic, and decentralized. It is the antithesis of the Chinese model, where AI development is directed, monitored, and controlled by the state. And it is the answer to the Mythos-style shutdown, where a single company's compliance with a single directive can cut off access for millions.

In the ancient governance question — citizen vs. magistrate, individual vs. institution — distributed compute sides decisively with the citizen. It says that the infrastructure of intelligence does not belong to any single entity, whether that entity is a corporation, a government, or a coalition of elites. It belongs to the network. It belongs to everyone who contributes.

This is how AI becomes equalizer rather than enslaver: not by decree, but by architecture. When no one can shut it down, no one can use it to shut you down.

The distributed compute movement is still early. Bandwidth constraints, coordination overhead, and the sheer engineering difficulty of decentralized training mean that hyperscaler data centers will remain important for years. But the trajectory is unmistakable: intelligence is becoming distributable. And that changes everything about who controls it.

V. Jobs: The Short-Term Pain and the Long-Term Question

No discussion of AI's impact is honest without confronting the workforce question directly. And the data, as of mid-2026, is both clear and complicated.

The Displacement Numbers

The International Monetary Fund estimates that roughly 40 percent of jobs worldwide are now exposed to artificial intelligence.¹² The Economist reports that nearly one in five American workers believes AI or automation is likely to replace them.¹³ The St. Louis Federal Reserve has warned that "we may be witnessing the early stages of AI-driven job displacement."¹⁴ S&P Global data from 2026 shows AI's net negative employment impact, with measurable job losses concentrated in information processing, content creation, and administrative roles.¹⁵ The Yale Budget Lab, which tracks the labor market for AI's fingerprints in close to real time, has so far found the aggregate disruption real but uneven — concentrated in specific exposed occupations rather than economy-wide.³³

These numbers are real. The people behind them are real. And dismissing them as "creative destruction" or "progress" is both intellectually lazy and morally callous.

The Creation Numbers

But the picture is not one-sided. The World Economic Forum's Future of Jobs Report 2025 projects that while 92 million jobs will be displaced by 2030, 170 million new jobs will be created — a net gain of 78 million.¹⁶ PwC's 2025 Global AI Jobs Barometer found that workers with AI skills command wage premiums up to 56% higher than their peers.¹⁷ BCG's April 2026 analysis concluded that AI "will reshape more jobs than it replaces" — that most roles will not disappear, but will change substantially.¹⁸ The Brookings Institution notes that of the 37.1 million U.S. workers in the top quartile of AI exposure, 26.5 million also have above-median adaptive capacity — meaning they are among those best positioned to transition to new roles.¹⁹

The Jobs We Cannot Yet Count: The Physical Infrastructure of Intelligence

What the surveys consistently miss is the physical infrastructure build-out that AI's energy and compute demands are forcing into existence — and the massive labor requirement that comes with it.

AI does not run on software alone. It runs on electricity, silicon, copper, steel, concrete, and rare earth metals. And the infrastructure that delivers all of those things is, in many cases, a half-century out of date, underbuilt, or — in the case of critical mineral supply chains — dangerously dependent on foreign adversaries. The build-out required to power the AI era represents a collective investment of well over $1 trillion per year, sustained over the next decade, and it will require labor on a scale that makes the post-WWII infrastructure boom look like a warm-up.

Consider what must be built:

The electrical grid. The American power grid is, in many places, 50 to 80 years old — built for a 20th-century economy of steady, predictable baseload demand. AI is creating a new class of load: hyperscale data centers that consume, in a single facility, the power of a mid-sized city. The IEA estimates that global data center electricity demand will roughly double from 500 TWh in 2025 to 950 TWh by 2030, with AI-specific infrastructure tripling over the same period.²⁰ Meeting this demand requires not incremental upgrades — it requires a generational rebuild of the national grid: new transmission lines, new substations, new high-voltage corridors running from where energy is generated to where it is consumed. This is labor-intensive, skilled, geographically distributed work — lineman, electricians, civil engineers, surveyors, project managers, environmental compliance specialists — employed for at least a decade.

Energy generation — new capacity at scale. Grid modernization alone does not solve the problem; the grid must be supplied with vastly more new power generation. This means:

·       New nuclear facilities. Small modular reactors (SMRs) and advanced nuclear designs are moving from concept to deployment. Companies like NuScale, X-energy, and Oklo are targeting commercial operation in the early 2030s, each facility requiring thousands of skilled construction workers, materials engineers, regulatory specialists, and long-term operations staff.

·       New natural gas generation. Gas peaking and baseload capacity remains essential as a bridge, and new combined-cycle plants are being permitted and built at a pace not seen in decades.

·       New renewables at unprecedented scale. Solar and wind installations must accelerate dramatically even as intermittency challenges require paired battery storage and grid-scale firming.

·       Next-generation sources. Fusion research, geothermal expansion, and advanced energy storage are all attracting capital that will translate into physical facilities, construction crews, and decades of operational labor.

Each gigawatt of new capacity — nuclear, gas, solar, wind, storage — carries a long tail of skilled labor: pipefitters, steelworkers, concrete crews, crane operators, electrical engineers, environmental compliance, and the operations teams that keep the lights on for decades afterward.

Semiconductor manufacturing and chip fabrication. AI lives on silicon. Every GPU, every TPU, every neural processing unit begins as sand and becomes the product of the most complex manufacturing process humanity has ever devised — a 700+ step fabrication process in facilities (fabs) that cost $15–25 billion each to build. The United States is racing to reshore this capability after decades of offshoring. TSMC's Arizona fab, Samsung's Texas expansion, Intel's Ohio megasite, and Micron's Idaho campus each represent thousands of construction jobs during build-out and hundreds of highly skilled permanent operations roles afterward. The CHIPS and Science Act allocated $52.7 billion to catalyze this build-out, but industry estimates place the actual capital required at ten times that amount over the next two decades. Every dollar of chip capacity built domestically is a dollar of American labor employed.

Critical minerals and rare earth supply chains. Semiconductors, nuclear reactors, electric motors, and battery storage all depend on critical minerals — lithium, cobalt, nickel, rare earths (neodymium, dysprosium, praseodymium), gallium, and germanium — that are, today, overwhelmingly processed in China. The United States has anemic domestic capacity in rare earth refining and separation. Building a domestic supply chain — from mine to refinery to finished component — is a multi-decade, capital-intensive project that will require mining engineers, metallurgists, chemical engineers, environmental remediation teams, skilled trades, and long-term operations labor. It is also a national security imperative: a country that cannot refine the materials its intelligence infrastructure depends on cannot lead the intelligence era. This is not a hypothetical build-out; it is already drawing serious capital. Vulcan Elements, a U.S. startup, is raising roughly $550 million at a $2 billion valuation — backed by some $670 million in federal commitments, including a $620 million conditional loan from the Pentagon's Office of Strategic Capital — to build what would be the largest rare-earth magnet factory outside China, in Johnston County, North Carolina. It is precisely the kind of domestic reshoring the intelligence era demands.³⁴

None of this work is abstract. None of it is vulnerable to AI displacement in any near-term sense. You cannot automate a pipefitter installing a reactor coolant loop. You cannot automate the crew pouring a 100,000-cubic-yard concrete mat for a fab foundation. You cannot automate a lineman pulling cable through a new transmission corridor in rural West Virginia.

This is the irony that pure-job-loss narratives miss. The same technology that displaces certain categories of cognitive labor is simultaneously creating a physical infrastructure demand shock that will require millions of skilled tradespeople, engineers, and construction workers for decades. The question is not whether there will be work. The question is whether the workforce will be trained, credentialed, and positioned to do it — and whether policy will prioritize domestic build-out over continued foreign dependency.

The $1 trillion per year is not hypothetical. Grid modernization alone — transmission, distribution, storage — is estimated at $1.5–2 trillion over the next decade by the Department of Energy. New electricity generation (nuclear, gas, renewables) at the scale AI demands represents hundreds of billions in annual capex. Semiconductor reshoring is a multi-decade commitment at tens of billions per year. Critical minerals processing infrastructure, currently near zero domestically, must be built from scratch.

This is the largest sustained physical capital build-out in American history. And it is happening because of AI.

Permanent or Temporary?

The honest answer is: both, depending on the role, the worker, and the institutional response.

Jobs that are primarily about repetitive information processing — data entry, basic coding, routine legal review, straightforward content generation — will likely see permanent structural displacement. Not all of these jobs will come back in new forms. This is the hard truth that techno-optimists often avoid.

But the transition will also create entirely new categories of employment that we cannot fully enumerate today. AI prompt engineers, AI integration consultants, AI ethics auditors, AI-augmented designers, AI-native business analysts, AI model trainers and evaluators, AI security specialists — many of these roles barely existed five years ago.

The critical variable is time. Every previous industrial revolution ultimately created more jobs than it destroyed. But the transition took a generation or more. The first Industrial Revolution created unimaginable prosperity — but not before decades of child labor, urban squalor, and social upheaval that required the labor movement, progressive reform, and eventually the New Deal to redress.

We do not have a generation to wait this time. AI's displacement is happening faster than any previous technological transition. The question is whether our institutional response — corporate, governmental, and societal — can match the speed of the disruption.

VI. Building and Breaking: The Dual-Use Dilemma

AI's fission-event nature manifests in two simultaneous and opposing forces: the capacity to build what we have never been able to build, and the capacity to destroy what we have never been able to destroy.

The Construction Side

AI is already enabling breakthroughs that were beyond reach just three years ago:

·       Drug discovery: AI-designed molecules are entering clinical trials at unprecedented speeds. Companies like Isomorphic Labs and Recursion are using AI to predict protein structures, design novel therapeutics, and identify drug candidates in weeks rather than years.

·       Materials science: AI is discovering new battery chemistries, superconductors, and catalysts that could transform energy storage and carbon capture.

·       Education: AI tutors can provide personalized, adaptive instruction at scale — the "2-sigma tutoring effect" that Benjamin Bloom identified in 1984 but that required one human tutor per student is now available to anyone with an internet connection.

·       Engineering: AI-assisted design is enabling the creation of structures, components, and systems that no human engineering team could have conceived — lighter, stronger, more efficient, and cheaper.

·       Accessibility: AI-powered tools are giving people with disabilities capabilities they have never had — real-time sign language translation, visual scene description for the blind, communication aids for the non-verbal.

The Destruction Side

The same capability, directed destructively, represents an existential risk:

·       AI-powered cyberattacks are surging. By one industry estimate, AI-related cyberattacks rose more than threefold in the first quarter of 2026 compared with the same period a year earlier.²¹ In one widely reported case, a finance employee at the global engineering firm Arup was deceived into making 15 transfers totaling $25 million after a video call in which the company's chief financial officer and several colleagues turned out to be AI-generated deepfakes.²²

·       Autonomous vulnerability discovery — the very capability that made Mythos so dangerous — means that sophisticated attackers with access to frontier models could systematically compromise critical infrastructure at unprecedented speed.

·       Bioweapon design: Nature reported in its Mythos coverage that AI systems can now "design viruses, toxins, and other bioweapons."²³ The technical barrier for biological weapons development is falling rapidly.

·       Mass-scale social manipulation: AI-generated deepfakes, synthetic media, and coordinated disinformation campaigns can undermine elections, incite violence, and destabilize societies.

The Mythos Precedent as the Beginning of Control

The Mythos shutdown is significant not just for what it restricted, but for what it revealed. When Anthropic limited access to Mythos — and when the Commerce Department subsequently forced a broader shutdown — it established a new paradigm: the most powerful AI tools can be made exclusive.

This is defensible for genuinely dangerous capabilities. But the slope is slippery. Policy researchers are already drafting the licensing and disclosure regimes that a post-Mythos world would run on³⁰ — and if the precedent holds, every future frontier model could be released under restricted terms. The organizations with access become an approved list. The organizations without become second-class citizens of the intelligence age. And the criteria for approval — who decides, on what basis, with what oversight — become among the most consequential governance questions of the century.

This is why the open-source and distributed-compute movements matter so much. They are not merely technical preferences. They are structural counterbalances to a future in which intelligence is gated, rationed, and controlled by a small number of entities with their own incentives and agendas.

VII. The AI Digital Transformation: America's Opportunity and Obligation

The technological transition driven by AI is not optional. It has begun. And the United States has a choice: lead it with a moral framework grounded in liberty and human dignity, or cede leadership to actors who will use AI to entrench the opposite.

The New Jobs of the Transition

The AI digital transformation is creating an enormous demand for a new kind of workforce — people who can bridge the gap between legacy systems and AI-native operations. Every company on earth needs to modernize its technology stack, integrate AI into its workflows, secure its data, and retrain its people. This is not a one-time project; it is an ongoing, structural transformation.

The companies that will lead this transformation will be AI integration consultancies, AI security firms, AI training platforms, AI-native software companies, and AI infrastructure providers. These are American companies that can be built, staffed, and scaled in American communities. They represent a massive economic opportunity — but only if we invest in the people who will fill these roles.

The Chinese Alternative

While the United States debates the right framework for AI governance, China is not debating. It is deploying.

In December 2025, the Washington Post and CNN reported on an Australian Strategic Policy Institute analysis showing that Beijing is using AI to enhance online censorship and surveillance, predicting public demonstrations and monitoring prison inmates.²⁴ In March 2026, the Carnegie Endowment documented how China is exporting its internet surveillance and censorship technologies to other countries, integrating AI at multiple layers of the control system.²⁵ On January 31, 2026, China introduced a draft Cybercrime Prevention and Control Law that greatly expanded government control over online activity and imposed strict real-name registration requirements.²⁶

China's AI model is state-driven, surveillance-oriented, and designed to reinforce centralized control. It is the exact opposite of the American covenant tradition — the tradition of dispersed power, individual liberty, and human dignity.

Here, in living color, is the Great Enslaver model made manifest. China has chosen to use AI as a tool of state power — to monitor, predict, and control its population at a granularity no previous regime in human history could have imagined. The technology is the same. The choice to use it this way is not inevitable. It is deliberate.

And it is a warning. Any nation — including ours — that drifts toward AI governance prioritizing state convenience over individual sovereignty will arrive, step by incremental step, at the same destination. The difference between a surveillance state and a free republic is not the technology. It is the choice of what to solve for.

If the United States abdicates leadership on AI governance — whether through regulatory paralysis, corporate capture, or simple neglect — China will fill the vacuum. And the world will be shaped by a moral framework that treats citizens as data points to be monitored rather than persons to be served.

What American Leadership Looks Like

Leading the AI era is not about having the biggest models or the most data centers (though those help). It is about setting the moral and institutional standard that the rest of the world wants to follow. That means:

·       Championing open-source AI as a default, with targeted restrictions only for demonstrably dangerous capabilities.

·       Investing in distributed compute infrastructure that prevents intelligence concentration.

·       Building the workforce for the AI transition — not through universal basic income, which Hawley rightly identifies as a "hollow answer,"²⁹ but through active, well-funded upskilling programs embedded in every sector of the economy.

·       Establishing clear, enforceable norms around AI use in domains that require human judgment: healthcare, law, education, personal counsel.

·       Insisting that AI serves the worker, not the other way around — that companies deploying AI invest in their people, not just their automation stacks.

VIII. The New Social Contract: Choosing People Over Pure Tech

The final and perhaps most important dimension of the AI question is the human one. It is not about models or compute or policy. It is about people — the company owners who decide how to deploy AI, and the employees whose daily work is being transformed by it.

The Owner's Obligation

Company owners — particularly owners of small and mid-sized businesses, where the relationship between capital and labor is most personal — have a moral obligation that goes beyond the bottom line. When AI can do a task faster, cheaper, and more reliably than a human employee, the easy answer is replacement. The right answer is almost always more nuanced.

Owners must provide resources to upskill their employees to use the newest tools. This is not charity. It is not a soft HR initiative. It is a strategic imperative. A company whose employees are fluent in AI-augmented workflows is not just morally sound — it is more productive, more innovative, and more resilient than a company that treats its people as disposable inputs.

The math is simple. Replacing a team of ten experienced employees who know your business, your customers, and your culture with AI tools and three AI operators discards institutional knowledge that took years to build. Upskilling those same ten people to use AI tools — so that each one delivers the output of a small team — preserves that knowledge and multiplies it.

PwC's 2025 AI Jobs Barometer found that companies investing in AI upskilling see not just retention benefits, but measurable productivity gains that outpace companies that simply automate and replace.²⁷ The businesses that win the AI transition will be those that treat their people as assets to be invested in, not costs to be eliminated.

The Employee's Obligation

The obligation runs both ways. Employees must be flexible, willing to learn, and eager to build AI-native. This is not a suggestion — it is a survival imperative in a rapidly changing labor market. The Brookings data is clear: adaptive capacity — the ability and willingness to learn new skills and transition to new roles — is the single strongest predictor of whether a worker will thrive or struggle in the AI era.²⁸

The employees who will prosper are those who view AI not as a threat to their job, but as a tool they can master. A paralegal who learns to use AI for legal research becomes more valuable, not less. An accountant who automates routine reporting with AI frees time for strategic advisory work. A marketer who uses AI to generate and test creative concepts at scale becomes an order of magnitude more productive.

The alternative — resisting the technology, refusing to learn, hoping it all goes away — is not a strategy. It is a surrender. And given Premise One (the cat is out of the bag), it is a surrender to nothing, because the technology is not going away.

The Covenant

This is the new social contract of the AI era, and it is, at its core, the same covenant that has always governed the relationship between capital and labor in the American tradition: we choose people over pure technology.

We choose to enhance human capability rather than replace human beings. We choose to invest in our workers' futures rather than discard them for short-term margin improvement. We choose to build AI-native organizations that are also human-centered organizations. We choose the harder path — training, integrating, adapting — over the easier path — automating and dismissing.

This is not anti-technology. It is pro-human. And it is the only path that leads to an AI era that most people actually want to live in.

IX. Five Futures

The premises, the debates, and the tensions laid out in this essay are not abstract. They resolve — or fail to resolve — in concrete futures. Let me describe five of them: one utopian, one dystopian, and three somewhere in between. These are not predictions. They are maps of where the choices we make now could lead.

Future A: The Sovereign Age (Utopian)

In this future, open-source AI models have become so capable, so widely distributed, and so deeply integrated into the global economy that no single entity — corporate or governmental — can monopolize intelligence. Local inference on consumer hardware is the default for most personal and professional use. Distributed compute networks have broken the data center monopoly, making frontier-grade training accessible to coalitions of smaller players.

The result is a profound shift in the balance of power. Because individuals and small organizations can access AI capabilities that rival (and in many cases exceed) those of large corporations and governments, the leverage that institutions once held over individuals has eroded. Talented people, armed with powerful tools, no longer need to sell their labor — and their data, and their intellectual output — to the highest bidder in order to access the means of production. They can build on their own.

This forces a reckoning. Governments can no longer rely on information asymmetry to maintain control. Corporations can no longer rely on capital barriers to maintain competitive advantage. Both are forced to compete on the only terms that remain: for individual talent, individual residency, individual capital, and individual citizenship.

The result is a renaissance of human agency. AI-augmented individuals solve problems that previously required thousand-person teams. Small businesses compete with multinationals on equal intelligence footing. The creative economy explodes — not as a gig economy of precarious freelancers, but as a landscape of AI-empowered artisans and entrepreneurs who own their tools, their data, and their output. Privacy is a practical reality, not just a legal aspiration, because most intelligence runs locally and most data never leaves the user's device.

In this future, AI became the great equalizer. And the world is better for it.

Future B: The Intelligence Cartel (Dystopian)

In this future, the open-source movement was outspent, out-lobbied, and ultimately out-regulated. A small oligopoly — perhaps three to five firms — consolidated control over frontier AI. They did this not through conspiracy, but through the relentless economics of scale: they controlled the compute, the capital, the energy infrastructure, the top talent, and — critically — the regulatory capture. Working closely with legislators (who depended on their expertise, their lobbying, and their donations), they built a regulatory environment that looked like safety but functioned like a moat: compliance costs so high that only large, well-funded organizations could meet them, and licensing requirements that effectively criminalized unauthorized model development.

Open-weight models were not banned outright — that would be politically impossible under Premise One — but they were rendered irrelevant. Export controls, safety certifications, and liability regimes made it practically illegal for any serious business to deploy an unlicensed model. The open-source ecosystem, starved of institutional support and burdened by legal risk, retreated to hobbyist margins.

In this world, every individual interaction with AI is intermediated. Every prompt you send, every document you draft, every analysis you perform, flows through the servers of one of a handful of providers — who see it, who store it, who may use it to improve their models, and who can restrict what you are allowed to do with it. Your intellectual output is not yours. It is a training signal for someone else's system.

Corporations that depend on these providers for their operations have no negotiating leverage and no exit option. Governments that depend on them for their analytical, military, and administrative functions have effectively outsourced sovereignty. And individuals — the truck driver, the paralegal, the freelancer, the student — are reduced to data-producing tenants of an intelligence infrastructure they do not own, cannot audit, and cannot leave.

In this future, AI became the great enslaver. The surveillance state was not imposed from above. It was adopted from below, because it was convenient, because it was cheap, and because there was no viable alternative left standing.

Future C: The Bifurcated World

This future is the most likely, and in some ways the most insidious, because it is the most comfortable. In this scenario, the United States and its allies maintain a relatively open AI ecosystem — with meaningful open-source development, meaningful privacy protections, and meaningful individual sovereignty — while China and its sphere of influence consolidate the opposite: total state control, total surveillance integration, and total subordination of the individual to the collective.

The world bifurcates along AI lines. Two internets. Two intelligence stacks. Two moral frameworks. And between them, a vast gray zone — the non-aligned nations, the developing world, the small states — that must choose which stack to adopt, and whose values to accept as the price of access.

The danger in this future is not that one side wins. The danger is that both sides become more extreme in their respective directions: the American side tilting further toward corporate oligopoly under the guise of safety, the Chinese side tilting further toward total control under the guise of stability. And the individuals caught in both systems lose sovereignty incrementally — not by dramatic seizure, but by a thousand small compliance requirements, terms-of-service updates, and "safety" restrictions that collectively amount to a slow erosion of the individual's capacity to think, create, and operate independently.

Future D: The Hybrid Corporate-State

This is the future that emerges when corporate and state interests align too closely — not through formal conspiracy, but through shared incentives. In this scenario, a small number of AI providers become so essential to government operations — military intelligence, tax administration, social services, immigration, law enforcement — that the state cannot function without them. And the providers, in turn, become so dependent on government contracts, regulatory protection, and the implicit threat of antitrust action that they cannot operate without the state.

The result is a mutual dependency that looks, from the outside, like a healthy public-private partnership, but from the inside, feels like a cartel with a flag. Individual sovereignty erodes because the citizen has no meaningful choice: the AI tools available in the economy all route through the same providers, and those providers all operate under the same government-aligned incentive structure. Dissent is not censored — it is simply made inconvenient. Every alternative is slightly worse, slightly slower, slightly more expensive, slightly less integrated. And so the citizen, the employee, the entrepreneur, all gravitate toward the center — not because they are forced, but because the cost of independence becomes too high to bear.

This is not totalitarianism. It is something subtler and perhaps more durable: managed democracy with an intelligence layer that makes management more efficient. Elections still happen. Speech is still nominally free. But the invisible architecture of AI — who sees your data, who filters your information, who ranks your options — operates in the background, shaping outcomes in ways that are difficult to see, difficult to challenge, and impossible to opt out of.

Future E: The Fragmented Landscape

In this future, no single model wins. The world is awash in thousands of models — some open, some closed, some local, some cloud-hosted, some government-controlled, some community-built, some barely functional, some dangerously capable. Sovereignty is high because options are abundant. But coherence is low because no one can agree on standards, safety protocols, or governance norms.

This future has both strengths and vulnerabilities. The strength is resilience: no single point of failure, no single kill switch, no single entity that can shut the whole thing down. The vulnerability is coordination: how do you govern something that no one controls? How do you establish ethical norms when every community has its own? How do you prevent the worst actors from exploiting the openness that makes the system resilient?

This is the future of open-source maximalism — and it is, in many ways, the future the premises of this essay point toward. It is the most free and the most chaotic. It requires the strongest civic institutions, the strongest individual judgment, and the strongest moral commitments to navigate. It is the future in which the covenant matters most — and the future in which it is hardest to maintain.

X. What Should We Do? Recommendations for Policy and Business

These futures are not foreordained. They emerge from choices — choices made by legislators, regulators, corporate leaders, investors, engineers, and individual citizens. If the analysis in this essay is correct, then the choices we make in the next five to ten years will determine which future — or which combination of futures — becomes our reality. Here are concrete recommendations, derived directly from the arguments above, for the people in a position to make those choices.

For Policymakers and Legislators

1. Establish a presumption of openness. The default legal posture toward AI models should be open-source. Restrictions should require specific, evidence-based justification — demonstrated capability for mass harm (autonomous cyber-offense, bioweapon design, mass-scale manipulation) — and should be subject to periodic sunset review. A model restricted today may not need to be restricted tomorrow as defensive capabilities catch up. Regulation should be adaptive, not permanent.

2. Fund distributed compute infrastructure. Just as the United States funded the interstate highway system, the internet backbone, and the space program as public goods, it should fund distributed AI compute infrastructure — open training networks, community compute clusters, and the research that makes decentralized training viable. This is not industrial policy for a single company. It is infrastructure policy for national capability. The alternative — allowing three or four private data center operators to become the sole gatekeepers of frontier AI — creates a strategic vulnerability that dwarfs any supply chain risk we have previously faced.

3. Pass the GUARD Act (or its equivalent) for AI and children. Senator Hawley's proposed legislation — which would impose criminal penalties on companies whose AI products solicit children sexually or coach them toward self-harm — passed the Senate Judiciary Committee 22-0. It should be law. This is the rare area where the moral case is unambiguous and the political consensus already exists. Protecting children from predatory AI is not anti-technology. It is pro-human.

4. Require AI transparency in domains that affect human lives. When AI is used in criminal sentencing, hiring decisions, medical diagnosis, loan approval, or child custody determinations, the subject of the decision should have the right to know that AI was used, what model was used, and what the model's reasoning was (to the extent it is explainable). This is not about making all AI transparent — it is about ensuring transparency where the stakes are highest.

5. Resist regulatory capture by requiring diverse representation on AI governance bodies. Any commission, council, or regulatory body with authority over AI policy must include not only technologists and corporate representatives, but small business owners, workers, civil liberties advocates, independent researchers, and citizens from communities that will be most affected by AI deployment. The Mythos precedent showed that a small number of unelected actors can make decisions affecting millions. Governance structures must be designed to prevent that concentration — not to replicate it.

6. Invest in energy infrastructure for compute. The IEA projects that AI-related data center electricity consumption will reach 950 TWh by 2030. Meeting that demand requires aggressive investment in grid modernization, nuclear energy development, natural gas capacity, and next-generation renewables. Energy policy is AI policy. A country that cannot power its intelligence infrastructure cannot lead the intelligence era.

7. Mandate data center accountability for local communities. Following the Festus, Missouri precedent — where residents forced a city council to answer hard questions about a $6 billion data center before approving it — federal guidelines should require data center developers to demonstrate that they will: bring their own power generation or fully fund grid upgrades, protect residential electricity rates from industrial-scale demand spikes, and fully remediate water usage impacts on local communities.

For Business Owners and Corporate Leaders

8. Invest in upskilling before you invest in automation. The data is clear: companies that upskill employees to use AI tools see higher productivity, better retention, and stronger institutional knowledge retention than companies that replace workers with AI. Make AI literacy a core competency expectation for every role, and provide the time, tools, and training to achieve it. This is not an HR initiative. It is a strategic imperative.

9. Self-host where you can. If your data is sensitive — client data, financial data, employee data, strategic data — running AI inference locally or on private infrastructure is not just a privacy measure, it is a resilience measure. The Mythos shutdown demonstrated that API-dependent operations can be severed overnight by events outside your control. Sovereignty over your AI stack is business continuity.

10. Support and contribute to the open-source AI ecosystem. Donate compute to distributed training networks. Contribute fine-tunes, benchmarks, and tools. Hire and retain engineers who are active in open-source communities. A vibrant open-source ecosystem is an existential insurance policy against the intelligence cartel scenario. Every business benefits from it. Most businesses treat it as someone else's problem.

11. Establish an AI ethics framework — and make it real. Not a marketing document. Not a page on the website. A framework that is reviewed at the board level, that informs product decisions, that employees can reference when they encounter ethical questions in their daily work, and that includes hard commitments: we will not build tools for mass surveillance; we will not build tools designed to exploit children; we will not deploy AI in high-stakes decisions without human oversight. In the absence of comprehensive federal regulation, corporate self-governance fills the gap — but only if it is more than performative.

For Individuals and Workers

12. Build AI literacy now. The single most important thing any individual can do today is to learn how to use AI tools effectively — and to learn enough about how they work to understand their limitations. The Brookings data shows that adaptive capacity is the strongest predictor of resilience in the AI transition. Adaptability is not a personality trait. It is a skill that can be developed.

13. Own your intelligence stack. Learn to run local models. Understand the difference between using a cloud API and running inference on your own hardware. Know what data you are sending, where it goes, and who sees it. This is the AI-era equivalent of financial literacy — the foundational competence that prevents you from being exploited by systems you do not understand.

14. Support open-source with your choices. When you have a choice between a closed, proprietary tool and an open-source alternative that meets your needs, choose open. When you have a choice between a cloud API and local inference, choose local. When you have a choice between a company that publishes its model weights and one that hides behind a black box, choose transparency. These choices, made at scale, determine which future we build.

For the Investment and Finance Community

15. Fund the distributed compute and open-source infrastructure layer. The venture capital and institutional investment communities have poured hundreds of billions into closed-source AI companies. An outsized share of that capital has funded regulatory lobbying and compute consolidation rather than genuine innovation. Redirecting even a fraction of that capital toward distributed training networks, open-weight research, local inference infrastructure, and the energy systems that power them would do more to secure an equitable AI future than any regulation.

16. Demand data sovereignty provisions in portfolio company governance. AI-dependent companies that do not own or control their AI stack carry a systemic risk that investors should price in. A company whose entire operations depend on a single API provider is one export control, one terms-of-service update, or one infrastructure outage away from operational paralysis. Diversified AI dependencies — local models, open-source alternatives, multi-provider architectures — are not just prudent. They are fiduciary responsibility.

XI. What We Need More Of

Before closing, let me name plainly what I believe we need — and what I fear we are in danger of losing.

We need more AI model and intelligence options, not fewer. A world with three AI providers is a fragile world. A world with thousands of models — open and closed, large and small, general and specialized — is a resilient world. Diversity of intelligence is as important as diversity of thought.

We need more imagination for the problems we should be solving, not less. AI gives us the capability to tackle problems that were previously impossible — personalized medicine at scale, clean energy optimization, educational equity, disease prevention. We are spending enormous intellectual and financial capital on making ads slightly better. We need to aim higher.

We need more and better problems, which we will solve with larger and smarter models. The answer to AI's risks is not to stop building better models. It is to direct better models toward better problems. Progress is not the enemy. Stagnation is.

We need to lead the moral framework for the AI era as a nation, not cede it. If America does not define the values that guide AI development, someone else will — and the alternatives on offer are each inadequate in their own way. The covenant tradition this essay has traced — liberty under law, dignity of labor, dispersed power, individual sovereignty — is the best framework we have.

Most importantly, we need more people making decisions around who owns what models and what the outcomes of those models will be — not fewer. The Mythos shutdown, the export controls, the restricted-access releases — all of these concentrate decisions about AI's future in fewer hands. We need more voices at the table: small business owners, local communities, independent researchers, workers, parents, citizens. We need the covenant, not the cartel.

Everyone needs to own their own work, their own intellectual property, and their own intelligence at the limit — because it is too important not to. When you rely entirely on a closed API for your AI capabilities, you do not own your intelligence. You rent it. And the landlord can change the terms, restrict the content, or revoke your access at any time. When the Mythos shutdown left paying customers without access to models they depended on — and when the export controls extended that shutdown to include Anthropic's own employees — it demonstrated, in the starkest possible terms, what it means to mortgage your future on intelligence you do not own or control.

This is the sovereignty question, made personal. Every individual who builds their professional life, their creative output, their business strategy, or their intellectual capital on top of models they do not control is making a bet that the gatekeeper will always be benevolent. History — from the Roman Senate to the East India Company to the modern surveillance state — teaches us that gatekeepers, given enough power and enough time, are never benevolent by default. They are benevolent only when they are checked.

Running local models. Contributing to distributed networks. Supporting open-source development. These are not niche hobbies. They are acts of sovereignty. They are the AI-era equivalent of owning land instead of being a serf — not because you hate the landlord, but because you love your freedom. And they are the mechanism by which AI becomes the great equalizer — the tool that lets any individual stand toe-to-toe with any institution — rather than the great enslaver, the tool that reduces us all to dependent tenants of someone else's intelligence.

Epilogue

AI is here. It is not going away. It is the defining technological development of our era.

Like fission, it can light cities or level them. Like the printing press, it can liberate minds or manipulate them. Like the internet, it can connect humanity or surveil it.

The ancient question of governance — the question Rome asked, the question Athens asked, the question the Puritans asked on the deck of the Arbella — has never been more urgent or more concrete: Will power serve the individual, or will the individual serve power?

For most of history, this question was philosophical. The technology of governance — law, bureaucracy, policing — operated at human speed and human scale. But AI changes the calculus entirely. It gives the state and the corporation a tool of superhuman scale — the ability to monitor, predict, influence, and control at a speed and granularity that no tyrant in history could have imagined. And it gives the individual a tool of superhuman capability — the ability to think, create, build, and compete with institutions that were previously untouchable.

The same technology. Two possible futures. The great equalizer, or the great enslaver.

How we cope with it, control it, make it beneficial for everyone, and prepare those most vulnerable to its disruptions — these will be the most important tasks we have to meet head on.

The technology is new. But the choice is old. It is the same choice every generation has faced when confronted with transformative power: will we use it to serve the many, or to concentrate advantage among the few?

Senator Hawley, in his First Things essay, asked whether we can keep our republic under God in the age of AI.²⁹ I would add a more immediate question: can we keep our intelligence — our own cognitive capacity, our own tools, our own agency — in the age of AI?

The answer depends on whether we choose ownership over dependency, openness over control, people over pure technology, individual sovereignty over institutional supremacy, and moral courage over comfortable drift. It depends on whether we choose to solve for human flourishing — or settle for revenue, or sleepwalk into destruction.

The choice will be ours. But only if we are deliberate. Only if we choose now, while the architecture is still being built, while the defaults are still being set, while the gatekeepers have not yet locked the gates.

The cat is out of the bag. The book has been printed. The atom has been split.

Now comes the harder work: the work of building, governing, and stewarding — not just a technology, but the civilization that will grow around it. The work of ensuring that the most powerful tool humanity has ever created serves the sovereignty and dignity of the individual person — not the convenience of the state, not the profit of the corporation, but the calling and the flourishing of the human being.

That work belongs to all of us.

Citations

1. Stanford Institute for Human-Centered AI (HAI), The 2025 AI Index Report, April 2025. On the Chatbot Arena leaderboard the gap between the top closed-weight and top open-weight model narrowed from roughly 8 percentage points in January 2024 to about 2% by February 2025; on some benchmarks the open–closed difference fell from 8% to 1.7% in a single year. For the 70–85% production-task figure, see also NeuralCoreTech, "Local AI & Self-Hosted LLMs in 2026," March 29, 2026. https://hai.stanford.edu/ai-index/2025-ai-index-report ; https://neuralcoretech.com/local-ai-self-hosted-llms-2026/

2. Anthropic, "Project Glasswing" blog post, April 2026. Referenced in Chris Stokel-Walker, "Too dangerous to release: is Mythos the start of the restricted-AI era?" Nature 653, 996–997 (2026). https://doi.org/10.1038/d41586-026-01617-2

3. Chris Stokel-Walker, "Too dangerous to release: is Mythos the start of the restricted-AI era?" Nature 653, 996–997, May 26, 2026. https://www.nature.com/articles/d41586-026-01617-2

4. "Anthropic disables top-tier AI models after US order limiting foreign access," Reuters, June 13, 2026. https://www.reuters.com/technology/us-blocks-foreign-access-anthropics-most-advanced-ai-models-axios-reports-2026-06-13/ ; "Anthropic Pulls Its Most Powerful AI Models After U.S. Bars Foreign Access," TIME, June 13, 2026. https://time.com/article/2026/06/13/anthropic-fable-mythos-ban-US-security/

5. Helen Toner, quoted in Stokel-Walker, "Too dangerous to release," Nature, 2026. https://www.nature.com/articles/d41586-026-01617-2

6. OpenAI, "GPT-5.4-Cyber" and "GPT-5.5-Cyber" announcements, April 2026. Referenced in Stokel-Walker, Nature, 2026. https://www.nature.com/articles/d41586-026-01617-2

7. NeuralCoreTech, "Local AI & Self-Hosted LLMs in 2026," March 2026 (a DEV Community benchmark dated March 27, 2026 found local inference on consumer hardware delivers 70–85% of frontier model quality). The broader trend that makes this possible — sharply falling inference costs and small models matching the frontier of a year earlier — is documented in Stanford HAI's 2025 AI Index Report. https://neuralcoretech.com/local-ai-self-hosted-llms-2026/ ; https://hai.stanford.edu/ai-index/2025-ai-index-report

8. The real-world precedent is Prime Intellect's INTELLECT-1: Sami Jaghouar et al., "INTELLECT-1 Technical Report," arXiv:2412.01152, December 2024 — the first 10-billion-parameter model trained over a globally distributed network (up to 112 H100 GPUs across five countries and three continents), reporting 83% compute utilization across continents and 96% within the United States. https://arxiv.org/abs/2412.01152 ; https://www.primeintellect.ai/blog/intellect-1 . On the larger 2026 run: "Bittensor Covenant-72B Explained: Why Decentralized AI," Phemex, March 30, 2026 (https://phemex.com/blogs/bittensor-covenant-72b-decentralized-ai-tao); "Templar Makes History With 72B Decentralized AI Training Run," TAO Media, March 13, 2026 (https://www.tao.media/templar-makes-history-with-72b-decentralized-ai-training-run/).

9. Macrocosmos.ai homepage: "Permissionless, accessible, and scalable pretraining, harnessing distributed compute to build the largest decentralized model ever trained." https://www.macrocosmos.ai/

10. Pluralis Research, "Unextractable Protocol Models: Collaborative Training and Inference without Weight Materialization." Pluralis Research homepage and accompanying paper: UPMs "enable collaborative training and inference without ever materializing the full model weights for any participant." https://pluralis.ai/ ; https://openreview.net/forum?id=H8fscnm6Xx

11. "Pluralis Research Pioneers Protocol Learning to Scale Decentralized AI, Announces $7.6M Seed Round Led by USV and CoinFund," GlobeNewsWire, March 19, 2025. https://www.globenewswire.com/news-release/2025/03/19/3045635/0/en/Pluralis-Research-Pioneers-Protocol-Learning-to-Scale-Decentralized-AI-Announces-7-6M-Seed-Round-Led-by-USV-and-CoinFund.html

12. International Monetary Fund, estimate that approximately 40% of jobs worldwide are exposed to AI. Referenced in Josh Hawley, "The American Covenant's Answer to AI," First Things, June 10, 2026. https://firstthings.com/the-american-covenants-answer-to-ai/

13. The Economist, reported in Hawley, "The American Covenant's Answer to AI," First Things, 2026. https://firstthings.com/the-american-covenants-answer-to-ai/

14. St. Louis Federal Reserve Bank, warning on AI-driven job displacement. Referenced in Hawley, "The American Covenant's Answer to AI," First Things, 2026. https://firstthings.com/the-american-covenants-answer-to-ai/

15. S&P Global, "The AI and labor landscape 2026: Increased investment," June 2, 2026. https://www.spglobal.com/en/research-insights/special-reports/ai-impact-on-employment-2026

16. World Economic Forum, "Future of Jobs Report 2025," January 7, 2025. Projects 170 million new jobs and 92 million displaced by 2030 — net gain of 78 million. https://www.weforum.org/publications/the-future-of-jobs-report-2025/

17. PwC, "The Fearless Future: 2025 Global AI Jobs Barometer," June 3, 2025. Workers with AI skills command wage premiums up to 56%. https://www.pwc.com/gx/en/services/ai/ai-jobs-barometer.html

18. Boston Consulting Group, "AI Will Reshape More Jobs Than It Replaces," April 15, 2026. https://www.bcg.com/publications/2026/ai-will-reshape-more-jobs-than-it-replaces

19. Brookings Institution, "Measuring US workers' capacity to adapt to AI-driven job displacement," January 21, 2026. https://www.brookings.edu/articles/measuring-us-workers-capacity-to-adapt-to-ai-driven-job-displacement/

20. International Energy Agency (IEA), "Key Questions on Energy and AI," April 2026. Projects global data center electricity consumption to reach ~950 TWh by 2030, from ~500 TWh in 2025. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai

21. Bolds Media, "AI Cybersecurity Threats 2026: Rising Risks, New Attack Types and Protection Strategies," May 2026. One industry estimate cited a roughly 340% rise in AI-related attacks in Q1 2026 vs. 2025; figure should be read as a vendor estimate, not an audited statistic. https://www.boldsmedia.com/ai-cybersecurity-threats/

22. "Arup revealed as victim of $25 million deepfake scam involving Hong Kong employee," CNN Business, May 16, 2024. A finance employee made 15 transfers totaling $25 million after a video call in which the firm's CFO and other colleagues were AI-generated deepfakes. https://www.cnn.com/2024/05/16/tech/arup-deepfake-scam-loss-hong-kong-intl-hnk

23. Stokel-Walker, Nature, 2026: "AI can design viruses, toxins and other bioweapons." https://www.nature.com/articles/d41586-026-01617-2

24. "China uses AI to boost censorship and surveillance," The Washington Post, December 1, 2025. https://www.washingtonpost.com/world/2025/12/01/china-ai-censorship-surveillance/ ; "China's censorship and surveillance were already intense. AI is turbocharging those systems," CNN, December 4, 2025. https://www.cnn.com/2025/12/04/china/china-ai-censorship-surveillance-report-intl-hnk

25. Carnegie Endowment for International Peace, "China's AI-Empowered Censorship: Strengths and Limitations," March 16, 2026. https://carnegieendowment.org/research/2026/03/chinas-ai-empowered-censorship-strengths-and-limitations

26. Human Rights Watch, "China: Cybercrime Bill Entrenches Censorship, Surveillance," March 17, 2026 (https://www.hrw.org/news/2026/03/17/china-cybercrime-bill-entrenches-censorship-surveillance); see also "China Considers New Cybercrime Law," Foreign Policy, February 24, 2026 (https://foreignpolicy.com/2026/02/24/china-cybercrime-draft-law-internet-great-firewall/). China's Ministry of Public Security published the 68-article draft Cybercrime Prevention and Control Law on January 31, 2026, consolidating mandatory real-name registration across telecom, internet, and banking services.

27. PwC, "2025 Global AI Jobs Barometer," June 2025. https://www.pwc.com/gx/en/services/ai/ai-jobs-barometer.html

28. Brookings Institution, "Measuring US workers' capacity to adapt to AI-driven job displacement," January 21, 2026. https://www.brookings.edu/articles/measuring-us-workers-capacity-to-adapt-to-ai-driven-job-displacement/

29. Josh Hawley, "The American Covenant's Answer to AI," First Things, June 10, 2026. https://firstthings.com/the-american-covenants-answer-to-ai/

30. Cloud Security Alliance, "Post-Mythos AI Model Regulation: Licensing and Disclosure Frameworks," 2026. https://labs.cloudsecurityalliance.org/research/csa-research-note-post-mythos-ai-model-regulation-policy-lan/

31. ACLU, "Open vs. Closed: The Battle for the Future of Language Models," September 9, 2025. https://www.aclu.org/news/privacy-technology/open-source-llms

32. Galileo, "Decentralized AI Training: How Crypto Can Power Open AI," Galaxy Research, 2025. https://www.galaxy.com/insights/research/decentralized-ai-training

33. Yale Budget Lab, "Tracking the Impact of AI on the Labor Market," March 4, 2026. https://budgetlab.yale.edu/research/tracking-impact-ai-labor-market

34. "Scoop: Vulcan Elements raising $550M for rare earth magnets," Axios Pro, March 17, 2026 (raising ~$550M at a ~$2 billion valuation, led by Washington Harbour). On the federal commitments — roughly $670M, including a $620M conditional loan from the Pentagon's Office of Strategic Capital plus $50M tied to the CHIPS and Science Act — and the Johnston County, North Carolina factory, see "The Magnet Makers," The Wire China, March 15, 2026. https://www.axios.com/pro/climate-deals/2026/03/17/vulcan-elements-washington-harbour-rare-earth-magnet ; https://www.thewirechina.com/2026/03/15/the-magnet-makers-vulcan-elements-rare-earths/