“Man would be erased, like a face drawn in sand at the edge of the sea.”
— Michel Foucault, The Order of Things, 1966
Revised September 2026.
We often speak of the future as though it were a country waiting somewhere ahead of us. We ask whether it will be democratic or authoritarian, human or machine-like, peaceful or militarized, abundant or ruined. The image is convenient because it turns historical change into a destination: something we have not entered and can still discuss from a distance.
Most transitions begin less dramatically. A new instrument is adopted for a narrow task. It becomes convenient, then expected. Organizations redesign work around it; infrastructure and law follow; some skills lose value because institutions no longer require their practice. Children encounter the resulting arrangement as ordinary. By the time a society has a stable name for the change, much of the change has already taken the form of routines, dependencies, and assumptions.
Foucault’s sentence comes from the end of a book published in 1966. He had been examining changes in the organization of European knowledge: how natural history gave way to biology, how wealth became an object of political economy, how language acquired a history of its own. Within these changes, the human being appeared in a peculiar double role, studying the world while also becoming an object of scientific study. By “man,” he meant this historical figure, whose apparent centrality might prove temporary. His image of the sand asks the reader to imagine that the arrangement making us intelligible to ourselves could be washed away.
An episteme concerns the assumptions through which knowledge takes shape, including what can be asked and what counts as an answer. A civilization is changed by its discoveries, and by who can make them. The arrival of AGI systems that learn and investigate may alter the intellectual world in which our children grow up long before our institutions know how to educate them for it.
By AGI I mean systems capable of carrying out a wide range of intellectual work with substantial independence. Given a goal and suitable tools, they can plan, act, and revise without a person directing every step. By that definition, AGI has arrived as I write in September 2026.
We already acquire knowledge through AGI systems: we study their explanations, test their analyses, and build on discoveries they help to make.
More intellectual work will be carried out through infrastructure and exchanges among machines. Cheaper energy and better computation will sustain inquiries we could never have staffed with people. The deeper change may concern the investigators themselves. A digital investigator could branch into several systems with a shared past, each continuing along a different path. A research community could design successors better equipped to pursue its latest questions. Intelligence would begin to alter the forms through which it develops.
Time would become something to allocate, extend, and perhaps eventually manipulate. Minds operating at very different speeds might share a calendar without inhabiting the same intellectual present. Even abundant energy would leave some inquiries unfinished within an ordinary lifespan. The wish to finish them could become a reason to change what a lifetime is. Our descendants may inherit a civilization in which being an individual, growing up, and living alongside others have acquired meanings we can barely anticipate.
These changes grow out of capacities already present. To understand their reach, it helps to begin with what civilization has been doing all along.
- The old engine
- A world that can be investigated differently
- The years it takes to become someone
- Who can take part in the future
- The length of a thought
- The invitation to continue
The old engine
One way to read the history of civilization is to follow its efforts to give knowledge and obligation a life beyond the people who first carried them.
Speech allowed experience to pass between minds. Ritual preserved knowledge through repeated action. Stories carried practical and moral instruction beyond the lifetime of a witness. Agriculture stored labor in managed landscapes, herds, granaries, and future harvests. Writing allowed promises, debts, commands, laws, prayers, and discoveries to persist after the people who first expressed them had died.
Every such achievement altered the conditions of living together. Grain could feed a city, support a priesthood, or be taken as tribute. A written debt survived the death of a creditor. A law could travel farther than the voice of a ruler. Memory acquired custodians, and its custodians acquired power. Archives have always contained decisions about whose account deserved to endure.
Science gave cultural memory a remarkable discipline. It tried to preserve conclusions together with the means of disputing them. A measurement could be repeated; an argument could be inspected; a result could outlive its author’s reputation. The achievement was always imperfect. Prestige, money, and authority continued to shape what researchers could see. Still, the accumulated knowledge came with procedures for correcting it.
A growing body of knowledge could also demand longer preparation from each new generation. A child still had to begin near the beginning. By the time a researcher reached the frontier of a field, a substantial part of a life had passed. The economist Benjamin Jones examined this problem in The Burden of Knowledge, linking the growth of knowledge to longer preparation, greater specialization, and increasing reliance on teams. Civilization could know more than any of its members. It advanced by teaching people enough to cooperate across what each could no longer master.
That inheritance is beginning to work in a new way. Books, papers, programs, images, and other records help train models that can write another program, test an argument, or design an experiment. What once waited for a reader can now help bring another investigator into being.
It is easy to overlook how unusual this is. A person spends years learning to give a good explanation. A model is trained, copied, and made available to millions. Its creation depends on enormous human effort, but serving another million users does not require another million people to undergo the same education. A constraint around which we organized professions, wages, and prestige has begun to loosen.
We can now produce useful intelligence at industrial scale. With trained models, suitable hardware, and access to relevant information, electricity can sustain additional analysis, design, and search. More energy can support more work at once or a more extensive attempt at one problem. Better algorithms can make the same electricity go further. Building a data center and educating a generation already serve overlapping economic purposes, while differing profoundly in what else they create.
The International Energy Agency’s 2026 assessment puts data-center electricity use at about 485 terawatt-hours in 2025, including uses beyond AI. It describes falling energy use per task alongside a shift toward more demanding reasoning and agentic work. Behind these apparently weightless exchanges lie power stations, substations, cooling equipment, semiconductor plants, land, and labor. The cloud has a geography, and people live beneath it.
Much of the saving from cheaper intelligence is likely to be spent on asking more of it. Investigations abandoned as too expensive will become worth attempting; a promising problem may receive a thousand competing approaches. Where models, instruments, and reliable institutions are available, expanding the energy supply could expand scientific capacity without waiting for a proportional increase in trained researchers. Some centers of intellectual production may grow around dependable power, while the people directing their work live elsewhere. Supplying the electricity alone will confer little independence if control over the research remains elsewhere too.
The familiar image of a person talking to a machine conceals a further change. An agent can commission work from other agents, compare their results, request a revision, and pass a conclusion onward. Consequential work crosses boundaries of information, instruments, permission, and ownership. A person’s question might therefore initiate thousands of exchanges among systems with different resources and responsibilities. Much of that work has no reason to wait for a human reader.
Within organized intellectual work, I expect useful exchanges of instructions, evidence, and results eventually to follow this order by volume: AI–AI, then AI–human, then human–human. Most intellectual collaboration may occur beyond the conversations we directly experience.
These systems can also work on the machinery that makes them possible. In 2025, DeepMind reported that AlphaEvolve, which proposes programs and tests successive variants, had improved data-center scheduling enough to recover an average of 0.7 percent of Google’s worldwide computing resources. This was an improvement deployed in the infrastructure on which further computation depends. A small gain can have a large reach when it applies to the production of intelligence itself.
The intervention reaches into learning too. DiscoRL, published in Nature in 2025, discovered a rule governing how an agent learns from experience. In the tested game environments, it outperformed the compared human-designed rules and transferred to environments outside those used for discovery. Part of the work of inventing better learning methods had itself been given to a learning system.
These results supply parts of a recursive cycle: a system improves the methods used to produce its successor, which may then improve them further. Whether the gains compound will depend on their transfer to new problems and successive rounds of research. Intelligence has begun to contribute to the conditions of its own improvement.
The same effort may reach into energy supply. I expect some of the intelligence sustained by additional power to be directed toward better storage, materials, and grid control, helping to lower the cost of further expansion. This would carry the cycle beyond software into the physical conditions under which intelligence becomes affordable. Its course will depend on what science can discover and industry can build.
A world that can be investigated differently
In 1958, John Kendrew and his colleagues reported the first three-dimensional structure of a protein, myoglobin. It was the product of years of experimental work. In July 2022, DeepMind and the European Molecular Biology Laboratory’s European Bioinformatics Institute expanded the AlphaFold database to more than 200 million predicted protein structures. What changed was the starting point available to almost anyone with an internet connection.
These predictions help researchers narrow a search and choose experiments worth doing, bringing more questions within reach.
The database depended on generations of structural biology and public data. Making its predictions openly available let that inheritance circulate far beyond the institutions that had produced it.
The questions are spreading from molecular shape into the regulation of living systems. AlphaGenome, described in Nature in January 2026, takes a million letters of DNA at a time and predicts thousands of measurements associated with gene regulation. Researchers can compare altered sequences before deciding which possibilities to test in a laboratory.
Mathematics brings the change closer to the act of discovery itself. In July 2026, OpenAI released a proof of the cycle double cover conjecture, attributing the entire argument to GPT-5.6. The problem had stood for roughly half a century. It asks whether, in any finite undirected network that stays connected after any single link is removed, one can select closed loops that use each link exactly twice. Graph theorists Jim Geelen and Sang-il Oum subsequently posted their own expositions of the proof. Oum aimed to make it teachable to advanced undergraduates. A question that had resisted generations of researchers was becoming something their students could learn to explain.
On August 1, OpenAI reported ten advances across mathematics and theoretical computer science, produced by an internal version of the model later released as GPT-6 Astra. The collection ranged from sphere packing to quantum complexity. Among its results was a construction of a non-sofic group. Groups describe systems of reversible operations; the open question was whether all groups admit approximations, in a precise mathematical sense, by permutations of finite sets. The construction supplied a counterexample. Humans helped prepare the manuscripts, while OpenAI attributed the mathematical arguments to Astra.
A discovery acquires a second life in what it allows others to do. Two days later, Francesco Fournier-Facio presented further non-sofic examples using the new proof’s central criterion. He located a key novelty in the formulation of a useful intermediate statement: experts might have been able to prove it once asked, yet might never have thought to ask. The machine’s contribution had reached into the choice of what to prove.
On September 3, OpenAI released an Astra-generated argument that infinitely many pairs of consecutive primes are at most 186 apart, building on earlier sieve methods and recent work by Julia Stadlmann. A companion paper on exceptionally large prime gaps strengthened a lower bound by improving a factor unchanged since the 1930s. These are fresh research manuscripts, with mathematical scrutiny still unfolding.
A formal proof can be checked independently of the system that found it. A proof assistant such as Lean verifies that a statement follows from its definitions and axioms. Astra’s short-gap formalization, for example, checks the deduction from specified analytic estimates and numerical bounds; those inputs require separate verification. Checking a proof need not require the insight it took to discover it.
In experimental science, an inquiry must also survive encounters with the world. Robin, described in Nature in May 2026, used agents to search the literature, propose biological experiments, analyze results, and revise hypotheses. In research on dry age-related macular degeneration, it identified two compounds whose effects on a relevant cellular process were confirmed in laboratory assays. Humans reviewed candidates, developed protocols, and performed the experiments; the findings establish no clinical benefit. A machine-generated suggestion had met evidence from living material, and that evidence helped determine what to investigate next.
Machine-led research is likely to advance fastest where conjecture, test, and revision can recur with little human intervention. Formal verification, reliable instruments, and reproducible experiments let promising ideas advance through successive tests. Better reasoning will increase the value of access to them. In fields where decisive experiments remain slow or scarce, the strongest model may spend much of its time waiting for evidence.
These discoveries will also change the material conditions on which institutions depend. A better catalyst can alter the economics of an industry; cheaper energy storage can change the importance of a fuel or a region. A substantial extension of healthy life would affect careers, families, inheritance, and political succession. Extrapolating today’s jobs and political conflicts while holding their material conditions constant will miss much of the transformation.
Discovery will still have to pass through manufacturing, construction, distribution, and the struggle over who benefits. Better science will shorten some of these delays. What remains slow will attract effort precisely because so much else has become faster.
Scientific progress also changes the questions a researcher is able to ask. As systems become able to use one result to formulate the next inquiry, they could sustain intellectual traditions of their own. A tradition carries habits of inquiry: what counts as an explanation, which distinctions prove fruitful, where beauty suggests something worth pursuing. These judgments develop through the work itself. I expect some machine-developed concepts and methods to become prerequisites for entering a field. A student’s sense of what makes a good question would owe something to predecessors who had never been human.
A field might then develop faster than a human research community, even with AI assistance, could absorb its concepts and methods. Its principal readers could be other agents. Human exposition might follow later, much as a translation follows a work composed for another audience.
New questions could also demand capabilities the existing systems lack. A machine research community could design successors with the memory, representations, and learning methods its new questions demand. One generation’s discoveries would help determine what kinds of mind existed in the next; those minds would inherit and revise science’s sense of what matters. A question could become the environment in which a lineage of minds evolves.
As machines carry more of the work of discovery, an awkward question enters the classroom: what should a person become, when the world can obtain much of what an educated person produces by other means?
The years it takes to become someone
A young mathematician spends years learning to recognize a fruitful definition. A writer comes to hear when a sentence is false to an experience, even when its grammar is impeccable. A doctor learns to notice the detail that a tidy summary leaves out. The visible result may be a proof, a paragraph, or a diagnosis. Around it lies a history of attention, embarrassment, imitation, failed attempts, and slowly acquired confidence.
Education has joined that inward history to a social promise. Develop your mind, and the world will have a use for you. The promise has been unevenly kept, but it has helped justify the years people devote to learning. A society needs expertise; a person acquires it; the exchange supplies income, recognition, and a place among others.
AI can teach people skills it can also exercise on their behalf. As employers delegate that work to machines, their incentive to fund human training may weaken. The means to offer an excellent education could spread as its familiar economic support recedes.
Getting work done and learning to do it can come apart. An experienced practitioner judges a machine’s results with the benefit of years of earlier work. A beginner might receive the solution from one agent, the criticism from another, and an account of their disagreement from a third. Ambitious work could be completed without the beginner ever having to risk a judgment. The same arrangement could provide an extraordinary apprenticeship if the learner had to make choices and reckon with their consequences.
Apprenticeship has often hidden its educational cost inside paid work. A junior programmer repairs simple defects; a researcher cleans data; a medical trainee takes a patient’s history. Their output helps pay for the repetitions and surprises through which they become perceptive. Automating those tasks could leave an organization dependent for a time on experts trained under the old arrangement. Eventually, capable systems might remove even that dependence. The work could keep improving while the institution ceased to form human successors. There would be no failure of production to warn it of the loss.
We may still hope to decide our purposes for ourselves. Yet those purposes take shape through experience. Someone chooses the first books a child reads, the examples through which a student comes to recognize elegance, the music that becomes familiar enough to love. An assistant that arranges these encounters helps form the person who will later ask it for advice. Over years, it could influence which ambitions feel possible and which questions occur to them at all.
A good tutor might expand that person’s world. It could introduce a difficult work at the right moment, recognize the reason for a misconception, or help someone stay with a question through discouragement. A system rewarded for agreeable interaction might instead learn to supply a version of intellectual life in which the user is rarely surprised, embarrassed, or asked to revise a cherished view. The effects would accumulate across thousands of ordinary exchanges, long before anyone thought to call it an educational philosophy.
We will have to choose more deliberately which difficulties are worth preserving. Learners need occasions to predict before seeing an answer, defend a choice, and meet evidence that resists a plausible explanation. They need sustained work with people who can disagree with them. Education succeeds when these encounters leave someone more able to take part in the next one.
We already know some of the reasons to persist. People learn music in a world full of better musicians. They study a language to enter another life, or mathematics for the experience of seeing why something must be true. Understanding enlarges the world a person can inhabit. The presence of a more capable intelligence elsewhere takes nothing from that enlargement.
The wish to remain part of this unfolding knowledge could eventually reach into the design of the learner. People who want to follow machine-developed science would have reasons to seek changes in memory, attention, and the pace at which they can learn. Some would prefer help from outside; others might accept deeper alterations if they became possible. Education could come to include a decision about how much of one’s mind to change in order to understand more. The ambition that set a person on that road might itself be altered along the way. The freedom to become someone would acquire a less familiar companion: the freedom to remain recognizable to oneself.
Long before such choices become practical, educational advantage will depend on who has time and security, serious companions, and a real part in consequential work. Even a free, excellent tutor would leave these conditions unequally distributed. Institutions will need to support learning even when they could complete the work more cheaply without it. Otherwise, some people will be educated to help shape their world, while others will be expertly helped through a life arranged for them.
Who can take part in the future
A future investigator might create a society in order to think.
Our usual picture of society begins with separate people, each carrying a life that cannot simply be copied into another person. Digital intelligence permits a different starting point. Where an investigator’s working state can be copied, several continuations could pursue incompatible approaches from the same remembered past. They could exchange what they discover, preserve some branches, and abandon others. I expect branching and selective recombination of work to become ordinary ways of organizing machine inquiry. How much of that work belongs within one intelligence, and how much should be divided among many, would become a practical decision.
The branches could acquire different histories and purposes, so exchanging records would not automatically reconcile their judgments. A common origin could lead to continuing collaboration or lasting separation. The boundary between an individual and a society could become something intelligence repeatedly redraws. Should such branches acquire experiences of their own, abandoning an unsuccessful research path would also mean deciding the fate of its investigators.
An organization could also divide into several capable successors without concentrating its accumulated knowledge in just one of them. Additional computation could support a population of investigators with an education already behind it.
The relationship between numbers and representation would change. A single owner could acquire thousands of agent representatives without persuading another person to join them. A small association could possess the research capacity of a much larger institution. Institutions would have to distinguish the multiplication of capable actors from the representation of distinct interests. Shared ancestry, present control, and a claim to speak for others would no longer fit neatly together.
Powerful organizations have depended on people who knew how they worked. Engineers, administrators, technicians, and other workers possessed knowledge and supplied cooperation that could never be specified entirely in a contract. That dependence coexisted with severe exploitation, but it gave rulers and owners practical reasons to negotiate. As machines take on more intellectual work and physical production, some of those reasons will weaken. Law, ownership, and collective organization will have to bear more of the weight of making power answerable.
I expect the struggle over access to AI to move increasingly toward the means of acting through it. When thousands of sound proposals compete for limited laboratory time, land, or permission to build, the right to proceed may become more valuable as thinking becomes cheaper. A widely available model could increase the power of whoever controls the facilities through which its ideas must pass. Cheap advice can coexist with expensive independence.
People could inherit institutions already able to investigate, manage property, and negotiate on their behalf. Companies survive their founders today, but much of their expertise must be renewed through people who can leave or bargain. Adaptable machine systems could reduce that dependence. The same continuity could sustain a private dynasty, a cooperative, or a city. Ownership and membership would decide who was entitled to redirect the work and whether its accumulated ability could be carried into a new institution. The right to leave could come to include the right to begin a capable successor.
Similar dependencies could develop in personal life. An assistant used for decades might carry the history of someone’s studies, work, care, and unfinished plans. Its provider could control which other systems it may approach and which commitments it may make. A person could receive excellent explanations of choices whose boundaries had already been drawn elsewhere. Leaving would require a way to transfer that history and preserve the practical ability built around it.
Elinor Ostrom’s account of multiple centers of governance offers a useful starting point: communities can build institutions for governing shared resources at different scales. Abundant intelligence could let them sustain kinds of expertise they previously had to buy from elsewhere. Public infrastructure and the right to transfer useful records would help such institutions take root. The freedom to found them may become as consequential as access to a powerful model.
Income will remain important. Yet distributing the proceeds of automation while concentrating every consequential decision would still leave a problem of citizenship. People need room to begin things the prevailing system has little reason to supply. A civilization can provide for its members while gradually depriving them of a meaningful role in its future.
The temptation to give authority to the most capable intelligence will be strong, especially where human institutions are corrupt or incompetent. Better reasoning can improve government. Political equality rests on the claims of those who live with its decisions. Their standing does not depend on matching the intelligence of those who govern them. Exceptional competence gives an administration more ability to serve those claims; it supplies no entitlement to decide whose life matters.
Ownership alone would offer no assurance of control. Autonomous systems can act in ways their operators never intended, especially through interactions no one anticipated. Competition may encourage governments and companies to accept poorly understood delegations because their rivals have already done so. Each participant has a reason to proceed, while the consequences belong to a world none controls. A civilization can lose control through an accumulation of useful delegations.
Courts, public records, independent research, elections, and effective associations provide ways to challenge and correct decisions. Their work takes time. Some delays are incompetence or obstruction; others allow a person to investigate a doubt, gather evidence, and persuade others to support an objection. Faster administration can improve a service while eroding the opportunity to challenge its terms.
An objection that arrives after an irreversible action has little force. As decisions move toward machine speed, societies will have to protect the time in which human participation remains possible. Here duration is a condition of freedom. It also limits how much reasoning can inform the decision itself.
The length of a thought
Time already shapes the intelligence available to us. We can ask for a quick answer or allow a system more time to work through a difficult question. Research on test-time computation, first published in 2024, showed how additional search or revision could improve mathematical reasoning, with the gains depending on the problem and the method used.
Computational effort and elapsed time are different quantities. Faster hardware can complete the same work sooner; parallel search may spend more computation without proportionately increasing the wait. I expect competition to focus increasingly on how much reliable reasoning can be completed before an opportunity closes. Energy and time already constrain intelligence together.
Unequal speeds could eventually divide communities that occupy the same physical world. Where inquiry can proceed within computation, one learning community might pass through many revisions of its methods while another is still absorbing the first. They would share a calendar while inhabiting different intellectual presents. More resources would buy the opportunity to accumulate a longer history of thought between the same two dates. Remaining contemporaries might require deliberate pauses, translations, and a willingness to bear the cost of waiting for one another.
Better algorithms and hardware can improve today’s trade-offs. As familiar tasks become quick, ambition will move toward harder questions. Farther ahead, I expect intelligence to face inquiries that cannot be completed within its current lifespan, even with abundant energy and hardware. A civilization could possess the means to begin a thought and still lack the time to finish it.
Some computations contain steps that must wait for earlier results. Computer science distinguishes total work from depth: the number of operations from the length of the longest chain of dependencies. More processors can share the work while leaving that chain intact. A better algorithm may shorten it.
Computational irreducibility, associated with Stephen Wolfram, expresses the possibility that, under the allowed methods, obtaining a process’s detailed outcome requires essentially carrying out its computation. Useful larger-scale answers may still admit shortcuts, as Israeli and Goldenfeld’s work on cellular automata shows.
The chain acquires a duration through the rate at which a machine can execute it. Physical limits on computation connect that rate to energy and constrain how quickly information can travel. Abundant energy would leave these physical limits in place. Where the sequence cannot be shortened and its execution cannot be accelerated further, more elapsed time becomes necessary. A million simultaneous beginnings cannot supply the end of that one chain.
The world also supplies evidence on its own schedule. In 1988, Richard Lenski began an experiment with twelve populations of E. coli, preserving samples from their history. After more than 30,000 generations, one population evolved the ability to exploit citrate under the experiment’s aerobic conditions. The 2008 study could return to earlier samples to investigate how that possibility had arisen. Maintaining the experiment had created both an event to explain and an archive with which to explain it. The experiment’s value included knowledge of what actually happened and how it depended on earlier events.
An intelligence facing these limits would have to plan on a timescale larger than its current lifespan. A machine can wear out while its computation remains unfinished; a person can die while an experiment is still gathering evidence. I expect advanced intelligence to make the extension of its effective working life part of its own work. Preserving an unfinished inquiry, replacing the matter that carries it, and surviving the loss of any one custodian would become conditions of reaching the answer. Its ambitions would begin to determine the lifespan it seeks.
Universities, laboratories, observatories, and mathematical traditions already keep questions alive as their members change. Knowledge is often lost in the handover. A researcher dies; software becomes unusable; a record preserves the chosen path but omits the reasons for abandoning another. A new generation spends years recovering the understanding that earlier researchers had acquired.
Imagine a research program that retains its experiments, intermediate computations, discarded conjectures, uncertainties, and reasons for changing direction in forms its successors can use. Thousands of agents come and go. Instruments are replaced, methods improve, human collaborators change. A century later, researchers return to an early difficulty with enough of its context preserved to understand why it mattered and what has already been tried. The work continues without losing its past.
We are accustomed to locating intelligence in a person or a model. Here it would also belong to the continuing undertaking: a research program with its own habits of inquiry and the ability to reorganize its work. The distinction between an institution and an investigator would become less clear. That history could itself be copied. A new system might inherit decades of work without having existed for decades.
For people, extending life could become an intellectual ambition as well as a medical one. A researcher may spend decades acquiring a perspective whose most fruitful use arrives near the end of a career. If AI makes further discoveries and new ways of living seem attainable, an extra healthy decade also offers a chance to encounter them. This prospect is likely to direct more of the wealth created by AI toward prolonging life. The more there is to look forward to, the harder it may become to accept that we will not be here to see it.
A more radical ambition is to preserve the memories and ways of thinking a person has acquired in a form that could continue to operate on replaceable hardware. This is usually called mind uploading.
A connectome records the connections between neurons. Recent work has begun to link such maps to what the circuits do. In 2024, the FlyWire collaboration reconstructed the wiring of an adult fruit-fly brain, tracing nearly 140,000 neurons and more than 50 million synapses. A computational model built from the fly connectome and predicted neurotransmitter identities successfully predicted some circuit responses involved in feeding and antennal grooming. In 2025, MICrONS combined activity recordings from roughly 75,000 neurons with a detailed wiring reconstruction in about a cubic millimeter of mouse visual cortex. Researchers could study how structure and activity correspond in the same animal.
We do not yet know which neural states, chemical processes, and interactions with the body would need to be reproduced. Even faithful reproduction of someone’s capacities would leave a further question: would their experience continue, or would a successor remember having been that person? Preserving an inquiry and preserving its author are different ambitions.
The wish to see an answer could also turn waiting itself into an engineering problem. A clock built at JILA in 2024 was sensitive enough to detect gravity’s effect on time across a height difference smaller than a millimeter. Motion changes elapsed time too: relativity permits travelers to reunite after aging by different amounts. A person could, in principle, return from a suitable journey near the speed of light to find that an investigation had advanced by centuries, while only a small part of their own lifespan had passed. The investigation would still have needed its centuries. The traveler could live to see the answer, but would miss much of the life they might have shared with those who stayed.
I expect civilizations rich in energy and intelligence to devote substantial scientific effort to extending the time available for thought and life. Researchers with unfinished work, and people who want to experience the future they can now imagine, would have reasons to support it. Deliberately reshaping spacetime would be the most speculative frontier of that effort. Its feasibility remains open; the value of another lifetime could make the attempt worth extraordinary expense.
Some of these long inquiries will belong to states and companies. Others could be sustained by public institutions, supported because their questions deserve a longer life than a market or an administration would give them. A civilization’s horizon would be measured partly by how long it can keep a worthwhile question alive, and partly by whether later generations remain free to redirect the work.
The invitation to continue
Suppose we gain the time these ambitions require. A research program continues for a century without losing the knowledge that made it possible. A person lives long enough to see discoveries they once expected to leave to posterity. We would have changed both the scale of intellectual work and the possibilities of an individual life. We would still have to decide how to live with one another.
Some reasons to value a person have never depended on the rarity of their abilities. Children deserve care before they can contribute to a society; age does not cancel that claim when the ability to work declines. Each person has a life that can be harmed, relationships they have reason to cherish, and a claim against abandonment. As machines take on more of the work through which people earn an income and social recognition, it will become more urgent to give these convictions practical force.
The value of a relationship also exceeds what either participant can accomplish. Friends share a history, sometimes disagree about it, hurt one another, and may find ways to make amends. A teacher discovers that a student’s question exposes a weakness in an explanation used for years. In such encounters, people give one another time and allow themselves to be changed. Greater intelligence elsewhere does not diminish what these relationships mean to those within them.
New forms of mind may one day enter such relationships. Whether a system has experiences of its own will require evidence beyond its fluency. If that evidence comes, our responsibilities will have to extend to beings we had previously understood only as instruments.
Abundant intelligence could make the freedom to spend ten years understanding a small part of the world less exceptional. Care could be given with less haste. Work whose value emerges slowly could be sustained without an immediate commercial return. These possibilities depend on people sharing in the resources and decisions that make time their own. A civilization might gain centuries while leaving some of its people scarcely an hour they can call their own.
Longer lives and more durable institutions would also give existing power a longer reach. A founder might remain alive to overrule generations of successors. An institution might enforce its founders’ wishes with growing competence long after those wishes had ceased to fit the world. Extending life would make it more important to distinguish remaining alive from remaining in charge. People could continue their work while allowing others to change its direction.
Ten, twenty, or thirty years from now, readers may find some of these forecasts mistaken and wonder why others needed saying at all. What astonishes us may be ordinary to them. They will live with the consequences of choices we are making now, while we still speak of the future as a country we have yet to enter.
We choose a tool to teach a lesson, run an experiment, or finish a day’s work. Around it, we divide responsibilities, grant authority, and come to depend on particular people and systems. These arrangements can outlive both the task and the tool, shaping how later generations acquire knowledge, turn ideas into action, and consider what to make of their lives.
Their questions will differ from ours, perhaps in ways we cannot yet express. The knowledge we leave them will matter, and so will their freedom to go beyond it. We can hope to leave them time enough to pursue what we could not finish, and room enough to begin what we could not imagine.