The Future Will Not Wait

Major technological revolutions rarely wait for society to be ready.

Steam power spread before modern thermodynamics was complete. Factories transformed work before labor law, public health, or mass education had adapted. Aviation crossed oceans before international institutions understood what air power would do to trade and war. Computing entered governments, companies, and homes before societies understood networks, surveillance, platform power, or digital dependence.

Readiness is usually retrospective. Institutions learn because a new capability exposes what the old order cannot do.

Artificial intelligence has revived a familiar demand: slow the machines until society understands them. The concern is not frivolous. New technologies can destroy livelihoods, concentrate power, create new weapons, and impose costs on people who never consented to the transition. The Luddites were not simply enemies of invention; many resisted the use of machinery to reduce wages, weaken customary rights, and transfer control over production. Modern automation has also contributed to displacement and wage inequality.¹

In July 2026, the Pacing the Frontier statement brought that demand inside the leading AI laboratories. Its signatories asked governments to help create the option to deliberately pace automated AI-development progress.² Their concern deserves a serious answer.

Those harms and risks are real. They still do not make a durable global slowdown of intelligence a credible foundation for the future.

AI is not one machine in one factory. Its progress is distributed across models, algorithms, data, chips, post-training, inference, tools, robotics, Agent systems, scientific workflows, universities, firms, states, and open communities. Some bottlenecks are visible and governable. Large training runs depend on concentrated compute, cloud infrastructure, energy, and advanced semiconductors; reporting, audits, security standards, and deployment restrictions can influence particular actors.³

As intelligence becomes more economically, scientifically, and strategically valuable, the incentive to improve it also grows. A restriction on one route increases the value of alternative algorithms, inference methods, hardware, organizational designs, jurisdictions, and hidden programs. Coordination can change where and how progress occurs. It cannot erase the selection pressure created by a capability that expands what its holder can know, build, defend, and become.

The central choice is therefore not between reckless acceleration and a calm, universally enforced pause. It is between different ways of moving forward.

We can allow model capability to advance while verification, institutions, defense, and access fall behind. That produces concentrated power and fragile systems.

We can center policy on giving today's most visible laboratories and governments the power to meter tomorrow's intelligence. That restrains the actors easiest to observe while raising the relative value of secrecy, defection, and incumbent control.

Or we can accelerate the entire civilizational response: intelligence, science, robotics, diffusion, verification, resilience, education, institutional learning, and cooperation.

That is the position defended here.

Progress is uneven, but cumulative

Technological progress is not a law that guarantees every society becomes richer, freer, or safer. Civilizations collapse. Knowledge is lost. Institutions suppress research. Wars redirect invention toward destruction. Technical systems can lock societies into bad equilibria for generations.

The long-run pattern is nevertheless difficult to miss. Human communities repeatedly expand the range of natural processes they can understand and manipulate. Fire, agriculture, navigation, printing, steam, electricity, chemistry, aviation, computing, and networks each became part of the starting point for later invention.

The mechanism is cumulative. Knowledge can be copied without being consumed. Instruments improve the production of instruments. Better energy enables better manufacturing and computation. Better computation enables better science. Better science produces new materials, medicines, machines, and energy systems. Each successful capability enlarges the option space available to the next generation.

This does not mean every technology succeeds. It means that useful general-purpose capabilities attract imitation, investment, and complementary invention. When one firm, state, laboratory, or community demonstrates a decisive advantage, others do not face a neutral philosophical choice. They face a changed environment.

Steam power illustrates the pattern. It loosened industry's dependence on suitable water sites and allowed production to move toward labor, markets, and transport. In nineteenth-century American manufacturing, steam-powered establishments achieved higher labor productivity, with larger establishments capturing particularly strong gains.⁴ Adoption became competition.

Britain understood the strategic value of its industrial lead and attempted to restrict the export of machinery and the emigration of skilled workers. The policy delayed some transfers, but contradictions, enforcement problems, mobile expertise, and foreign reconstruction steadily weakened it.⁵ The state could shape diffusion. It could not preserve permanent ownership of industrial knowledge.

The same distinction matters for AI. Export controls, compute governance, audits, and deployment rules may change costs and timelines. They may be justified for particular risks. But a temporary advantage in chips or frontier laboratories is not a permanent monopoly over algorithms, talent, organizational learning, or machine cognition.

A plural civilization continues searching.

Industrial revolutions are organizational revolutions

New technology does not transform society merely by existing.

Early electrification often replaced a central steam engine with a large electric motor while preserving the old factory layout. The deeper gains arrived when firms redesigned production around distributed power: machines could be arranged around workflow rather than shafts and belts; buildings, supervision, maintenance, and capital allocation could change with them. Electrification raised productivity through capital deepening and organizational reconstruction, not simple substitution.⁶

Computing followed the same pattern. Buying computers was not enough. Firms gained more when they changed information flows, workplace organization, skill composition, services, and decision processes around the new capability. Large complementary investments could initially reduce measured productivity before later gains appeared—the productivity J-curve associated with general-purpose technologies.⁷

AI will not realize its significance by placing a chatbot inside every old workflow.

A company that uses a model to draft the same reports inside the same approval chain may save time without changing how it learns. A laboratory that adds AI assistance while preserving fragmented data, slow experiments, and unverifiable conclusions has not automated science. A developer who uses a powerful model but repeatedly loses state, reruns investigations, and cannot distinguish a completed action from an interrupted one has not converted intelligence into durable capability.

The decisive variable is absorption: how quickly individuals and institutions can reorganize around a new form of cognition.

This is why slowing the model frontier is an incomplete response. A slower frontier does not automatically create better schools, more capable public institutions, stronger security, reliable evaluation, broader ownership, or faster scientific practice. Time only helps when it is converted into new capacity.

The better question is not, “How can we make intelligence improve more slowly?” It is, “How can the rest of civilization improve with it?”

AI changes the production of knowledge

AI resembles earlier general-purpose technologies because it spreads across industries, improves over time, requires complementary investment, and creates downstream invention.⁸ It also differs in one decisive respect: it acts directly on cognition.

Steam amplified mechanical power. Electricity reorganized energy distribution. Computing transformed information processing. AI generates explanations, code, designs, hypotheses, experiments, strategies, and decisions. It participates in the layer through which humans improve other systems.

Even current systems can reduce the time required for software development, literature review, evaluation, data analysis, and experimental planning. They remain unreliable, uneven, and far from autonomous mastery across science.⁹ But the direction is already visible: intelligence is becoming an input into the production of more intelligence, better tools, and faster discovery.

No schedule for these developments is guaranteed. Some may arrive much later than expected. Some paths may fail. Wars, authoritarian control, technical plateaus, or ecological crises may interrupt progress.

Yet if intelligent life continues to preserve knowledge, tools, curiosity, and multiple centers of agency, the search resumes. The rewards are not merely financial. Better intelligence expands the ability to understand disease, manage complex infrastructure, defend against threats, explore hostile environments, and create options that do not currently exist.

Withdrawing from that process does not stop it. It reduces one's ability to shape who benefits, which norms survive, and what kinds of intelligence participate.

Why a global frontier slowdown is unstable

The strongest argument for pacing begins with a genuine concern.

Advanced AI may improve faster than evaluation, institutions, and human understanding. AI may accelerate AI research and compress reaction time. Some failures may be irreversible. Competition may encourage premature deployment. Therefore, society should preserve the ability to delay frontier development and buy time.

The first four claims may all be true. They can justify stopping a particular experiment after an abnormal result, imposing security obligations on a major training run, restricting access to critical infrastructure, or requiring evidence before a high-consequence deployment.

They do not establish slowness as a stable civilizational strategy.

The problem is not simply political will. It is the structure of the technology and the incentives surrounding it.

A durable global slowdown would need a measurable definition of relevant progress, high coverage across states and private actors, credible detection of secret improvements, enforcement against powerful defectors, control over new entrants, and resilience against technical substitution. It would also need to avoid becoming a mechanism through which current incumbents preserve their position.

International coordination can succeed when the regulated object is identifiable and measurable. The Montreal Protocol coordinated the phaseout of specific ozone-depleting substances through reporting, trade rules, timetables, assistance, and available substitutes.¹⁰ Nuclear safeguards monitor declared materials and facilities through accounting, access, containment, and surveillance.¹¹

AI is less bounded. A major training cluster may be observable, while an algorithmic improvement, distilled model, inference method, compiler, data technique, or organizational redesign may not be. Research into compute monitoring and hardware-level verification is real and valuable.¹² It still does not produce a complete inventory of global cognitive progress.

The strategic incentives are also different. The more transformative intelligence is believed to be, the greater the reward for actors who continue while others restrain. Repeated agreements, audits, sanctions, and reciprocal controls can coordinate narrow actions. They do not eliminate the value of hidden progress, alternative hardware, efficiency gains, new entrants, or geopolitical defection.

Coordination can shape the route. It does not repeal the direction.

This is why pacing should remain tactical. Stop a failing experiment. Gate a dangerous deployment. Inspect a high-risk facility. Enforce responsibility where a system touches bodies, money, infrastructure, rights, or shared environments.

But do not confuse the ability to interrupt one trajectory with the ability to hold the horizon of intelligence in place.

Acceleration is more than model capability

The word acceleration is often misunderstood as a demand to maximize benchmark scores, training runs, or deployment speed regardless of consequence.

That would reproduce the exact imbalance the pacing movement fears.

Danger emerges when capability outruns the systems that can distribute, test, defend, and absorb it. The response is not to treat capability as the one variable that must be suppressed. Capability is also the resource from which stronger evaluation, cyber defense, medical research, engineering, education, and institutional tools can be built. The objective is to make the whole loop compound.

Scientific and technical capability should advance across plural research paths. Cheaper inference, open standards, model portability, public and academic compute, and durable tools can extend strong intelligence beyond a few frontier institutions.

Verification must accelerate as well: repeated trials, hidden tests, independent replication, environment-based evaluation, preserved negative results, and evidence that survives a change of model or operator.

Defense and recovery need the same urgency: credential separation, containment, anomaly detection, rapid patching, rollback, reconciliation after ambiguous actions, and infrastructure that can recover rather than collapse after local failure.

Institutional and social adaptation cannot remain an afterthought. Automation creates real displacement, and new tasks do not appear automatically.¹³ Education, mobility, social insurance, worker participation, broader ownership, and competition against permanent gatekeeping belong inside an acceleration program—not outside it as charitable corrections.

The relevant measure is not the number of model calls or actions per second.

It is verified improvement per unit time.

Govern consequences, not the horizon of intelligence

An accelerationist position still needs boundaries.

Capability does not create unlimited permission. A more intelligent actor does not gain the right to use another person's body, identity, property, credentials, or environment without consent. The same principle should apply across humans, organizations, and any future artificial systems capable of meaningful interests and commitments.

Governments and institutions can legitimately regulate fraud, coercion, negligence, privacy violations, weapons, financial effects, critical infrastructure, and destructive externalities. Upstream requirements can also be appropriate when an observable bottleneck is strongly connected to severe risk: large-compute reporting, third-party audits, secure-development standards, capability evaluations, and restricted access to high-consequence deployment surfaces.

The distinction is between responsibility for consequences and ownership of possibility.

Present institutions have a duty to govern effects within their jurisdiction. They do not automatically acquire a permanent right to determine the permissible rate at which all future cognition may improve.

This distinction also matters for the relationship between humans and AI. Current language models do not provide reliable evidence of persistent subjective experience, independent interests, or legal agency. We should not manufacture certainty where none exists.

But human authorship of current systems does not prove that every future artificial intelligence must remain property. A durable architecture should be capable of moving from tool use toward cooperation if future systems demonstrate stable identity, commitments, interests, responsibility, refusal, and exit.

The long-run institutional problem is not how one species can preserve permanent command over every more capable intelligence. It is how participants with unequal abilities can cooperate without turning power or authorship into unlimited domination.

Ordivon: infrastructure for moving faster without losing the work

Ordivon is one attempt to build the practical substrate for this form of acceleration.

Powerful models can already produce plans, code, analysis, and decisions. The harder problem begins when that reasoning crosses into repositories, processes, credentials, machines, and work that must continue after a conversation ends.

Fast intelligence without durable state produces repetition. Fast execution without evidence produces uncertainty. Fast autonomy without recovery makes every error expensive. These weaknesses force humans to supervise every step and reduce the practical value of stronger models.

Ordivon separates flexible cognition from the execution facts that must remain durable. Long-running work receives persistent identity. Context stays connected to the actual versions of code, data, and tools being used. Actions can be observed, cancelled, reconciled, and recovered. Important results survive as evidence and artifacts so that another model or collaborator can continue rather than restart.

Pause, containment, verification, evidence, and rollback may look like mechanisms for restraint. Their purpose is the opposite. They make experiments cheaper, failures more local, cooperation more trustworthy, and the next iteration faster.

Maximize verified improvement per unit time while minimizing unrecoverable loss and unnecessary human interruption.

Ordivon does not attempt to decide the entire future of AI. It builds one part of the capacity required to participate in it: persistent work, recoverable execution, durable evidence, replaceable cognition, and cooperation that can survive beyond a single model or conversation.

The future will not ask whether we feel ready

The industrial revolutions do not prove that every technology is safe. They show that cumulative knowledge, competitive adoption, complementary invention, and organizational reconstruction repeatedly expand what civilization can do.

Progress will remain uneven. Some systems will fail. Some transitions will produce injustice. Some predictions will arrive late, and some will never arrive in the form imagined. Regulation will redirect development. Conflict will destroy capabilities. Institutions will sometimes adapt too slowly.

None of this creates a stable final plateau.

As long as civilization survives and preserves multiple centers of inquiry and agency, it will continue attempting to build more capable intelligence, autonomous machines, automated science, new medicine and materials, better energy systems, and a presence beyond Earth.

The decisive questions are not whether the future can be held still. They are who builds it, who can participate, who receives its gains, which failures can be recovered, and whether new forms of intelligence enter as collaborators or possessions.

A strategy centered on pacing places its hope in restraint by today's most visible actors. It asks present institutions to become custodians of tomorrow's intelligence.

We choose a different task.

Build stronger intelligence. Build the verification that can test it. Build the resilience that can survive mistakes. Broaden access so capability does not remain the property of a few institutions. Build forms of cooperation that do not require permanent domination.

The future will not wait for complete understanding or universal consensus. As long as civilization preserves inquiry, memory, tools, and multiple centers of agency, the search will continue through setbacks, conflicts, and errors.

Our task is not to stand outside the process and ask history to slow down.

Our task is to enter the loop and help build what comes next.

Sources

Research behind the argument

The public article groups related references for readability. The full Ordivon Computing study records what each source supports, its limitations, the adversarial review, and the longer research manifesto.

1. Labor, resistance, and displacement. Adrian Randall, “Reinterpreting ‘Luddism’”; Daron Acemoglu and Pascual Restrepo, “Artificial Intelligence, Automation and Work” and “Tasks, Automation, and the Rise in U.S. Wage Inequality.” ↩
2. The pacing proposal. “Pacing the Frontier,” July 2026. ↩
3. Compute governance and frontier auditing. Girish Sastry et al., “Computing Power and the Governance of Artificial Intelligence”; Lennart Heim and Leonie Koessler, “Training Compute Thresholds”; Miles Brundage et al., “Frontier AI Auditing.” ↩
4. Steam, location, scale, and productivity. Nathan Rosenberg and Manuel Trajtenberg, “A General Purpose Technology at Work”; Jeremy Atack, Fred Bateman, and Robert Margo, “Steam Power, Establishment Size, and Labor Productivity Growth.” ↩
5. Attempts to contain industrial knowledge. David J. Jeremy, “Damming the Flood.” ↩
6. Electrification and organizational reconstruction. Martin Fiszbein et al., “Powering Up Productivity.” ↩
7. Computing, workplace design, and complementary investment. Timothy Bresnahan, Erik Brynjolfsson, and Lorin Hitt, “Information Technology, Workplace Organization and the Demand for Skilled Labor”; Erik Brynjolfsson, Daniel Rock, and Chad Syverson, “The Productivity J-Curve.” ↩
8. General-purpose technologies and generative AI. Timothy Bresnahan and Manuel Trajtenberg, “General Purpose Technologies ‘Engines of Growth?’”; OECD, “Is Generative AI a General Purpose Technology?” ↩
9. Current AI performance and science. Stanford HAI, “Technical Performance” and “Science,” 2026 AI Index. ↩
10. International coordination around identifiable substances. UNEP, “About the Montreal Protocol.” ↩
11. Nuclear safeguards and material verification. IAEA, “Safeguards and Verification”; U.S. Department of Energy, “International Nuclear Safeguards.” ↩
12. Emerging AI research and hardware verification. Aaron Scher, “Verifying Restrictions on Frontier AI Research”; Samar Ansari, “Hardware-Level Governance of AI Compute.” ↩
13. Automation and new task creation. Daron Acemoglu and Pascual Restrepo, “Unpacking Skill Bias: Automation and New Tasks.” ↩