COORDINATION IS ALL YOU NEED: THE NEXT GREAT REINVENTION OF HUMAN POWER

One of the quiet indignities of modern life is being told that the future is coming, as though our principal responsibility is to prepare ourselves for whatever someone else has decided it will be.

Workers are told to adapt. Students are told to acquire new skills. Communities are told to remain competitive. Citizens are asked to choose representatives, then watch those choices become mandates for decisions they never discussed, conditions they never endorsed, and consequences they must nevertheless live with.

We receive more information about the future than almost any generation before us. Having a meaningful hand in shaping it is another matter.

Artificial intelligence makes that distinction impossible to ignore. Systems capable of generating research, building software, directing operations, and interpreting public opinion are becoming more powerful. The question is whether that power will make ordinary people more capable—or primarily make institutions more capable of acting upon them.

A future announced to people carries a different promise from a future they have the knowledge, tools, and standing to help build.

That promise connects developments that can appear unrelated: hundreds of machine-generated mathematical manuscripts; a criminal operator directing inexpensive AI agents; employees pretending to understand technology they have not been adequately equipped to use; and governments struggling to hear people whose actual judgments never fit neatly inside a political platform.

Each concerns the same relationship: the distance between the person who lives with a decision and the system empowered to make it.

The coming years could widen that distance. They could also give us an extraordinary opportunity to close it.

The ambition should be larger than producing more intelligence. It should be to expand the number of people who can put intelligence to work on purposes they understand, results they can examine, and futures they have a meaningful voice in choosing.

Everyone is living in 2026. Not everyone is experiencing the same reality.

In June 2017, eight researchers published Attention Is All You Need. The Transformer architecture they introduced helped establish a foundation for modern artificial intelligence.

Nine years later, its title offers an unexpected metaphor. Machines have become increasingly capable of producing sophisticated work, while human attention is stretched across the responsibility of understanding what that work means.

Yet the experiences informing our opinions about AI remain profoundly unequal.

Someone who occasionally asks a free chatbot to rewrite an email encounters a mixture of convenience and familiar mistakes. A developer directing an agent through implementation, testing, and revision may encounter something closer to a new source of digital labor. A researcher accessing an unreleased system may work with capabilities that neither person can independently examine.

They can all describe their experiences honestly and reach different conclusions.

Andrej Karpathy recently illustrated this divergence through a four-level funnel: roughly six billion people with little exposure to large language models; one to two billion casual users; approximately 20 million professionals using frontier capabilities in mathematics and software; and perhaps 5,000 people accessing advanced internal systems. These were explicitly his rough estimates, not population measurements. Their significance lies in the different experiences they describe, rather than the precision of the numbers. cellcog.ai

!

Karpathy’s rough estimates describe different experiences of AI. They are not measured population totals; groups can overlap and the conceptual chambers are not drawn to scale.

Measured adoption data provides firmer evidence of uneven exposure. Microsoft’s September 2026 Global AI Diffusion Report estimated that 18.8% of people ages 15–64 worldwide used generative AI during the reported period, compared with 28.8% in the Global North and 16.2% in the Global South. The measure has acknowledged coverage limitations and concerns usage during a particular period. It does not establish how many people can build or responsibly operate an agent. blogs.microsoft.com

Trying a chatbot, evaluating its limitations, creating a useful workflow, and safely delegating consequential work represent different capabilities.

That distinction matters because the most valuable knowledge often belongs to people who have little time to study the machinery now becoming available to them. A teacher understands where students struggle. A tradesperson recognizes the assumptions that make an estimate unreliable. A manager knows which recurring problem consumes hours without appearing on an executive dashboard.

Their knowledge does not become less valuable because it was acquired outside a research laboratory. The opportunity is to give that knowledge greater reach.

The real achievement would be a world in which more people can turn what they know into something they can build, improve, and direct.

When machines can propose more than humanity can immediately understand

OpenAI’s October mathematics release brought this challenge into unusually sharp focus.

The initial collection comprised 722 manuscripts across 372 related result families. The repository subsequently listed 719 manuscripts, describing results at different stages of verification. Some have accompanying Lean formalizations that allow proofs to be checked computationally; others require substantial examination. The initial manuscript count was never equivalent to 722 independently verified breakthroughs. openai.com

The event still matters. A frontier system produced a body of research substantial enough that understanding the collection became a research undertaking of its own.

Mathematicians must examine whether arguments hold, whether claims are original, whether dependencies are sound, and whether the work contributes something others can use. Scientific knowledge gains its strength through that process of scrutiny.

For centuries, institutions worked to increase the supply of difficult ideas. Now they must also improve their capacity to evaluate a growing supply of machine-generated candidates.

The same transition is extending into the physical sciences. On October 15, CERN and Mistral announced a partnership intended to develop foundation models for scientific discovery, including a proposed physics model trained on billions of simulated proton-proton collisions. CERN separately reported holding approximately one exabyte—one million terabytes—of experimental LHC data. The simulated training material and experimental archive are distinct categories. The partnership’s value will depend on the rigor with which its models are developed and tested. home.cern

The implications reach beyond laboratories. A machine-generated engineering design needs tests. A financial analysis needs traceable assumptions. A medical research claim needs appropriate evidence. An operational recommendation needs a clear account of what it might overlook.

Generating an answer and establishing a dependable basis for action are different responsibilities.

This creates an important economic opening. As sophisticated output becomes more abundant, systems that organize evidence, test claims, expose uncertainty, and preserve reliable records can become increasingly valuable.

It also creates a civic obligation. People affected by these systems need practical ways to understand the grounds on which decisions are made.

The cost of proposing knowledge is falling. The responsibility for establishing knowledge remains.

!

Publication begins verification. The original 722-manuscript release and subsequent 719-manuscript catalogue describe different snapshots of an evolving research collection.

The $25 operator—and the people still waiting for a useful introduction

A very different application of delegated intelligence emerged from Gambit Security’s investigation of attacks on online retailers.

Researchers documented 105 attack projects over six days, at least 27 compromised companies, and more than 600,000 unexpired payment-card records stolen from two victims. One operator combined Strix for vulnerability discovery, Cairn for exploitation, and Hermes for orchestration and continuing coordination. The operator issued 1,951 prompts across 260 recorded Hermes sessions. gambit.security

Recovered accounting showed an average model cost of $25.46 across 101 completed scans, while the broader campaign’s model expense was estimated at $12,000–$18,000. Those figures describe scan and campaign expenses, rather than a universal price for compromising an organization. At one retailer, an overly broad cleanup operation deleted 180 database tables, including legitimate backups. gambit.security

!

Gambit’s September 22, 2026 investigation: $25.46 mean model cost across 101 completed scans. The separate Strix records show 633 scanner-hours within 195 elapsed hours during August 23–31. Parallel scanner activity is not a human productivity multiplier or the total cost of an intrusion.

The case exposes both the leverage and the danger of delegated machine work. One person could direct substantially more activity than they could personally execute. Imperfectly constrained instructions could also produce damage beyond the operator’s immediate objective.

Now consider the other side of that disparity.

There are people with legitimate problems, valuable expertise, and years of accumulated knowledge who have never been shown how to construct a useful workflow for their own needs. Some have encountered little beyond marketing, shallow demonstrations, and instructions to become more “AI-ready.”

Visier’s survey of 1,000 full-time U.S. employees found that 48% had exaggerated their AI usage or expertise, 45% felt pressure to use AI without confidence, and only 28% believed their leaders fully understood how employees were using it. www.visier.com

Those findings describe a culture in which appearing competent can feel safer than admitting the need to learn.

That should concern anyone serious about progress. People should have room to investigate, ask basic questions, discover limitations, and reject tools that fail to help. Familiarity performed under pressure is a poor substitute for capability earned through practice.

A society that celebrates extraordinary machine performance while leaving ordinary people embarrassed by their uncertainty has misunderstood the purpose of technological advancement.

People deserve more than a warning that they are falling behind. They deserve a credible route forward.

The ability to act must grow alongside the ability to answer

The surrounding system determines much of what an agent can safely accomplish.

An assistant connected to documents has a different scope from one connected to a production database. A system preparing a recommendation has a different responsibility from one authorized to execute it. External material can contain information worth examining alongside instructions the assistant has no authority to follow.

Permissions, environments, tests, checkpoints, records, and recovery procedures make these distinctions practical.

At the frontier, understanding agent behavior can itself demand substantial resources. METR reported spending approximately $400,000 in API credits during a six-day investigation of an incident involving OpenAI agents and Hugging Face. In a subsequent proposal, it estimated that a rigorous two-month investigation of comparable scope could require more than $10 million in API credits. These figures concern specialized research investigations, rather than ordinary business oversight. metr.org

Cheap execution does not guarantee cheap assurance.

The response must include better ways to produce and examine evidence. Human beings cannot personally inspect every intermediate action at scale. They need systems that direct attention toward consequential uncertainty: failed tests, unusual behavior, missing information, unexpected permissions, and decisions whose costs are difficult to reverse.

Even measuring productivity demands humility. METR’s early-2025 experiment found that experienced open-source developers took 19% longer on the studied tasks when allowed to use AI tools, despite believing the tools had helped. In February 2026, METR described reasons to believe newer tools were producing greater benefits, while explaining why selection effects made its follow-up estimates unreliable. metr.org

The durable lesson is that perceived usefulness and measured usefulness can diverge. Models change. Environments change. The need to test consequential claims persists.

That principle should follow AI into every institution where it becomes influential.

A system’s fluency cannot settle whether its analysis is sound. Its efficiency cannot determine whether the objective is fair. Its capability cannot establish who had the right to authorize the action.

Capability, evidence, and authority must each earn their place in the decision.

A citizen is more complicated than a political package

This is where the technological story becomes inseparable from self-government.

A citizen can want safer streets, less surveillance, affordable healthcare, competitive businesses, secure borders, and limits on government power. Another can support environmental protection while questioning a particular energy policy.

These combinations reflect human judgment. They become difficult to express when public choices arrive in packages whose positions are expected to travel together.

A vote motivated by living costs can become part of a claimed mandate on education, foreign affairs, technology, and questions the voter never considered. Elections authorize representatives, but they cannot convey everything an electorate believes about every decision that follows.

America’s founding debates already recognized the danger of political organization overwhelming public judgment. In his 1796 Farewell Address, Washington warned about the “alternate domination of one faction over another,” intensified by revenge. Madison’s argument in Federalist No. 10 supplied a complementary constraint: disagreement arises naturally among free people, and suppressing its causes would threaten liberty. www.peters.senate.gov

Together, these arguments establish a demanding ambition. People must remain free to organize and disagree, while political competition remains answerable to the public it exists to serve.

In America’s 250th year of independence, that inheritance should give us confidence to improve our institutions. Rights, elections, constitutional continuity, and limits on power can coexist with more capable ways of hearing citizens.

The public is already more complicated than familiar labels suggest. Gallup’s 2025 findings showed 45% of U.S. adults identifying as independents, compared with 27% each as Democrats and Republicans. Most independents still leaned toward a party. Only 10% of all adults were non-leaning independents, while 33% described themselves as moderate. These are distinct measures, rather than evidence of one politically unified center waiting to be discovered. news.gallup.com

!

Gallup’s 2025 U.S. adult figures measure party identification, partisan lean and ideology separately. Rounded party-identification shares total 99%; the 10% and 33% figures concern all adults and are not additional slices of one chart.

Parties perform useful work. They recruit candidates, coordinate legislation, assemble coalitions, and help voters assign responsibility. Citizens should be able to benefit from that coordination while retaining the freedom to distinguish among proposals.

A person’s support on one question should not be treated as ownership of their judgment on every other.

No political organization should enjoy a permanent claim on the public’s imagination simply because it inherited machinery through which public opinion becomes power.

We should bring the ambition of persuasion to the work of listening

The conditions under which people encounter political information deserve scrutiny.

In a preregistered experiment involving 1,256 X users over ten days during the 2024 presidential campaign, researchers changed exposure to posts expressing partisan animosity and antidemocratic attitudes. Increasing or decreasing that exposure shifted measured hostility toward the opposing party by approximately two points on a 100-point scale. This was a bounded effect on attitudes, rather than evidence that a feed determines an election. It nevertheless demonstrated that ranking decisions can influence political feelings. datadryad.org

Political actors operate inside this environment. A message can be rewarded for the alarm it generates before anyone evaluates what it helps people understand. Disappointment with one’s own side can become tolerable when the alternative is presented as an existential threat.

AI can make persuasion cheaper and more adaptable. It can also make institutions better able to examine contributions that would otherwise be expensive to read.

That second possibility deserves a much larger place in the public imagination.

A person might oppose a proposal as presented while supporting a different version under clear conditions. “I support new housing if the infrastructure can accommodate it” contains information that a pro-development or anti-development label loses. So does “I would accept this cost if the benefit reaches these people and the program is reviewed after two years.”

These statements reveal where agreement might become possible, where safeguards matter, and where a commitment remains firm.

They also make trade-offs harder to conceal. Separate majorities may favor lower taxes, better services, and less debt while rejecting every combination that fits the budget. More input does not remove the responsibility to reconcile competing preferences.

It can, however, make the choices more intelligible.

The public does not have to meet at the midpoint of two platforms. It needs room to describe the conditions under which people can act together.

Listening can become a more capable public institution

The evidence is already more concrete than the idea of an artificial president.

In October 2025, the UK government reported that its Consult tool categorized more than 50,000 responses to the Independent Water Commission into themes in approximately two hours, at a reported computing cost of £240. Experts spent another 22 hours checking the results, while detailed manual reviews continued alongside the AI analysis. These figures describe one categorization stage, rather than the complete cost of consultation or policymaking. www.gov.uk

The practical significance is considerable. Reducing the administrative burden of organizing contributions can leave more capacity for interpretation, scrutiny, and response.

A 2024 Science study of the “Habermas Machine” examined another part of the relationship. Across experiments involving 5,734 participants, the system used people’s opinions and critiques to iteratively develop statements expressing common ground. Participants preferred its statements to those produced by human mediators. The finding concerned statements people endorsed; it did not establish that AI could determine correct national policy. deepmind.google

Earlier civic experiments show how participation can connect to institutions. In its UberX consultation, vTaiwan reports that more than 4,000 participants helped develop an agenda through four weeks of public input using Pol.is. Open participation in that process should be distinguished from a representative national vote. info.vtaiwan.tw

Ireland’s 2016–2018 Citizens’ Assembly brought together 99 citizens and an independent chair. Its work on the Eighth Amendment proceeded through parliamentary consideration and a national referendum. Deliberation developed recommendations; institutions and the electorate retained their respective authority. data.oireachtas.ie

These examples concern different functions. Analysis helps institutions read. Deliberation helps people reason together. Representative and electoral processes establish who may decide.

Connecting those functions could give citizens a more faithful relationship with public action. Someone’s concern would have a path from expression to examination, from examination to a proposal, and from a proposal to a decision that must be explained.

For people accustomed to contributing and hearing nothing further, that connection would matter.

Being heard should have an observable consequence: an answer, an explanation, a revision, or a clear account of why the institution chose another course.

!

The UK Consult figures describe one categorization stage, with expert checking and additional manual review. The theme names on this conceptual loom are illustrative.

What this could mean on an ordinary day

Consider a hypothetical town improving its bus service within a fixed budget.

For one resident, the issue is frequency. For another, it is the first bus arriving early enough to reach work. Someone else needs wheelchair access, an affordable fare, or a reliable route to a hospital.

These concerns may appear as competing entries in a budget. In a person’s life, they can determine whether a job is practical, an appointment is reachable, or independence can be preserved.

A campaign about being for or against public transport could conceal almost everything necessary to make the decision.

A more capable process would publish current spending, performance, legal obligations, proposed changes, and forecast uncertainties. Residents could question the evidence and supply missing information. Participation would be available online, on paper, by telephone, and in facilitated meetings.

AI could help organize submissions. Residents would retain the ability to inspect and correct summaries attributed to them. Someone asking for accessible vehicles should not disappear into a generic category called “better service.”

Officials would present coherent alternatives: - Improve frequency on existing routes, with a clear account of who remains underserved. - Expand access, while making fewer frequency improvements. - Increase funding, subject to the required authorization and a specified limit. - Each package would disclose its beneficiaries, disadvantages, costs, assumptions, and review conditions. Residents could identify what would secure or withdraw their support. - Some people initially opposed to one another might support the same package if it included a particular route, spending cap, or evaluation date. Others would remain opposed, and their reasons would stay visible. - The authorized institution would decide, explain its reasoning, and later compare promises with actual performance.

A resident could see which concern influenced the proposal, why another concern could not be accommodated, and who made the final choice. Changed evidence could justify revision without requiring everyone to pretend their earlier judgment had been irrational.

The same principle applies to automation’s economic benefits. If a service requires less labor to deliver, the gain could support lower taxes, shorter waiting times, broader coverage, or shorter working hours. Each distributes the benefit differently.

Technology can expand what is feasible. People still need a meaningful role in deciding what the expansion is for.

Greater productive capacity should widen the choices people can make about their lives.

!

A hypothetical bus-service decision. The diagram encodes no actual route distances, costs, ridership, or package rankings.

The person who still says no belongs in the future, too

The strongest objection must be built into this vision: a system introduced to loosen political gatekeeping could create a more powerful gatekeeper.

Whoever chooses the question can exclude alternatives before deliberation begins. Whoever selects the evidence can shape what appears reasonable. Whoever defines success can privilege one group’s interests.

An elegant interface can make those choices less visible while making the resulting answer appear objective.

A public summary therefore cannot authorize itself. People need competing arguments, correction rights, independent review, and access to the evidence behind consequential claims.

Participation must remain voluntary and practical. A person should not need to disclose a lifetime of private conversations to contribute a view about a local service. Someone with more money, free time, or automated agents should not acquire greater political standing through sheer volume.

The ability to generate submissions must not become the ability to manufacture a public.

Majority agreement also has limits. Minority rights cannot depend on whether a model finds them inconvenient to a popular proposal. An objection does not become disposable because it complicates the preferred answer.

The citizen who still says no remains part of the public the institution exists to serve.

That principle should shape the record of a decision. The strongest unresolved objection should remain accessible, including when the authorized decision proceeds despite it. Future review may depend on exactly the concern that was easiest to overlook.

Technology providers must also remain replaceable. An institution should be able to change systems without losing its records, abandoning its obligations, or surrendering its ability to explain what happened.

The measure of an intelligent institution should include how effectively a person can challenge it.

A future worthy of human beings must preserve room for judgment, disagreement, and correction—even when the machines would find unanimity easier.

!

A summary cannot authorize itself. These are proposed safeguards, not an enacted policy or an existing ZEN government system.

The beginning can be small enough to build and serious enough to matter

A council, school district, or civic institution could begin with one bounded decision and a public commitment to respond.

It could convene a broadly representative resident panel, support participation with accessibility assistance and compensated time, publish the evidence and budget constraints, and invite additional submissions.

The roles of the panel and open input would be explicit. Human reviewers would check AI summaries against original contributions. The authority would commit to a response date and explain which recommendations it accepted, rejected, or needed to investigate.

Evaluation would examine participation gaps, corrections, time and cost, treatment of dissent, and what changed.

The connection to actual authority is essential. A legislature could establish a formal route for citizen-developed, independently costed proposals to receive a hearing and recorded response. Participation thresholds, minority protections, and agenda limits would themselves require legitimate approval.

Representatives would retain the authority to reject a proposal. Public agreement would gain somewhere to go beyond a petition, a poll, or a favorable headline.

That would give public reasoning an institutional consequence.

Individuals can begin with the same discipline. Take one decision that affects your life. Find the original proposal and identify who has authority. Examine the evidence. State what you support, what you oppose, and what would change your judgment. Use AI to assist the investigation, then check the material claims.

The contribution should leave the institution with something specific enough to answer.

This approach is demanding because it asks citizens and officials to make their reasoning more visible. It is promising for the same reason.

Literacy should give people something they can carry into the world

The practical value of AI begins to deepen when a person can turn knowledge into a working capability.

Someone comparing insurance policies needs relevant provisions extracted from actual documents, unresolved questions identified, and conclusions connected to their sources. A manager balancing schedules needs constraints, conflicts, and consequences made visible. A business owner investigating recurring errors needs a process that can distinguish patterns from misleading coincidences.

The human supplies knowledge about what matters. The system helps organize and extend the work. Its result must remain open to examination.

A useful progression is straightforward: - Choose a meaningful problem. Solving it should improve a decision, reduce errors, save time, or create an opportunity. - Define completion. Specify the information, constraints, expected result, and standard for checking it. - Begin with bounded authority. Let the system prepare, compare, draft, or simulate before it takes consequential action. - Verify the outcome. Examine calculations, sources, omissions, performance, and the burden of supervision. - Expand demonstrated value. Preserve successful work as a reusable system and a lesson others can apply. - One reliable workflow can become a durable asset: documented, improved, adapted, and taught.

This is why building belongs at the center of education. Students should have opportunities to create, evaluate, modify, deploy, and constrain systems alongside learning how to question their output.

That principle is central to the AI Pioneer Program and the broader ZEN mission. A young person who builds a functioning agent encounters both capability and responsibility. Technology becomes something they can shape, test, and improve.

The deeper lesson travels beyond software. Problems can be investigated. Assumptions can be challenged. Evidence can change an answer. A system that appears authoritative can still be examined.

A citizen who can identify a missing assumption and articulate a workable alternative possesses something more consequential than product familiarity.

They possess a greater ability to influence what happens next.

!

A conceptual learning workflow: choose a real problem, define completion, bound authority, build, verify and reuse what works. These are practices to apply, not measured educational outcomes.

The future should belong to more than the people who reach the controls first

Access, education, infrastructure, incentives, and rules will shape who gains power from increasingly capable intelligence.

We should therefore measure progress through more than model scores and account totals. We should ask how many people can reliably accomplish meaningful work; whether verification becomes easier; whether delegated authority remains understandable; whether citizens can recognize their actual concerns in institutional summaries; and whether those contributions receive reasoned responses.

These are standards for judging whether technological advancement is becoming human advancement.

The people with the least time to study every new system are often carrying the responsibilities that make society function. They should not have to become full-time technology researchers to understand decisions affecting their work, families, communities, and opportunities.

Useful tools should respect that reality. Education should make capability more attainable. Institutions should make participation consequential.

A century from now, many names dominating today’s arguments may occupy only a few paragraphs in a history book. The arrangements we establish for turning human judgment into machine action and public decisions could last much longer.

We should want our successors to inherit institutions that can learn before a crisis forces them to; revise a policy without rewriting the identity of everyone who once supported it; and permit cooperation on one question without demanding agreement on every other.

We should want young people to grow up understanding that intelligence is something they can direct toward a meaningful purpose—and that power remains something human beings are entitled to question.

The people are not a party. They are the reason parties, governments, and every system we build must answer for what they do.

The promise of this age should reach the person building their first useful tool, the worker gaining command of a difficult process, the student discovering what they are capable of, and the citizen receiving a serious answer to a serious concern.

That is how a technological revolution can become something people recognize in their own lives: greater capability, greater understanding, and a stronger say in what comes next.

We were never meant merely to watch the future arrive.

We have this one chance to build something that fuel future civilizations for a millieum.