Build the next fifty years: how would you build them, ten years at a time, and what does an ordinary day look like in 2076?

I would also treat AI as infrastructure, not just as a consumer product. Every school, clinic, court, benefits office, small business, and city engineer should have access to audited AI tools that are better than the cheap, opaque systems people will otherwise use. That is possible today in weaker form and plausible at high quality by the mid-2030s. The rule should be: use AI to expand competence, but keep records, liability, appeal rights, and human authority where people’s rights are affected.

By 2036, the goal is not utopia. It is visible momentum: cleaner air, cheaper electricity in many places, fewer people trapped by bad housing policy, better public health surveillance, and schools where every child has a patient tutor. AI tutors that reliably help most students are plausible by then. AI replacing the social role of school is not something I would want.

From 2036 to 2046, I would scale what works. The big shift should be from “decarbonization is a burden” to “clean energy is the cheapest industrial base.” Green steel, green ammonia, synthetic fuels for aviation and shipping, long-duration storage, advanced geothermal, and safer standardized nuclear are plausible by then. Fusion contributing meaningfully by 2046 is a bet, not a plan. If it arrives, excellent; if not, the system should still work.

This decade should also be when robotics leaves warehouses and starts changing construction, elder care, farming, and maintenance. Humanoid robots everywhere by 2040 is a bet. Specialized robots that lay cable, inspect bridges, harvest crops, clean hospitals, and assist nurses are plausible by then. The right policy is not to freeze automation. It is to make sure productivity gains flow into wages, public services, shorter working hours, and lower prices.

From 2046 to 2056, I would aim for climate repair and institutional repair. Net zero for most rich countries is plausible by then if we start seriously now; net zero globally by mid-century is plausible but not guaranteed. Carbon removal at climate-relevant scale is a bet, but some removal for cement, aviation, agriculture, and historical emissions is worth building. I would not count on carbon removal as an excuse to delay emissions cuts.

Health should change a lot in this period. Continuous monitoring, personalized prevention, mRNA-like rapid vaccine platforms, organ printing or organ regeneration in some cases, and AI-assisted drug discovery are plausible by then. “Everyone lives to 150” is a bet I would not build policy around. But adding ten or fifteen years of healthy life for many people is plausible if we focus on prevention, cardiovascular disease, cancer, dementia, metabolic disease, and loneliness.

From 2056 to 2066, I would consolidate abundance without pretending scarcity has vanished. Water can be much less scarce in coastal and sunny regions through cheap energy and desalination; that is plausible by then. Food can have a smaller land footprint through precision fermentation, better crops, indoor production where it makes sense, and more plant-rich diets; that is plausible by then. Restoring large areas of land and ocean is plausible, though uneven.

This is also the decade when governance matters most. By then AI systems may be deeply embedded in science, defense, finance, media, and administration. Very powerful AI by 2060 is plausible; machines doing most economically valuable cognitive tasks better than humans is a bet, but a serious one. I would require licensing for frontier training runs, strong model evaluations, incident reporting, secure data centers, international monitoring of dangerous capabilities, and public-interest AI capacity not controlled by a few firms or states.

From 2066 to 2076, the job is to make the system humane. A civilization can be clean, rich, and miserable if people have no agency. I would protect human-scale institutions: families, friendships, neighborhoods, unions, schools, clubs, religions for those who want them, local journalism, and democratic assemblies. The point of technology is not to maximize screen time or GDP. It is to reduce involuntary suffering and expand meaningful choice.

Who builds this? Not one actor. Governments build the rules, grids, public science, health systems, procurement markets, and safety nets. Private firms build most hardware, software, housing, factories, drugs, vehicles, and services. Universities and public labs build much of the underlying knowledge. Workers build it literally. Cities implement much of it. Communities decide what is legitimate in their own places. Poor countries should not be told to wait their turn; they need finance, technology transfer, and bargaining power.

Who pays? Mostly the societies that benefit, with rich countries paying more because they emitted more and have more capital. I would use public debt for long-lived infrastructure, taxes for operating costs, carbon pricing where politically possible, clean procurement, social wealth funds that own stakes in key technologies, and large concessional finance for poorer countries. Private capital should do much of the buildout, but only under rules that prevent monopoly control of basic infrastructure.

What does it cost? My rough number is an extra 3 to 5 percent of world GDP per year for the next twenty-five years, falling after that as clean systems become cheaper and old systems retire. In today’s terms, that is roughly $3 trillion to $6 trillion a year in additional investment, on top of money already spent on energy, infrastructure, health, and defense. It is huge, but not fantastical. The world already invests about a quarter of GDP. The real issue is redirecting investment from fossil systems, wasteful medical spending, empty financial engineering, and military overhang into productive resilience. The cost of not doing it is higher: climate damage, pandemics, brittle supply chains, failed states, and runaway inequality.

One thing that goes wrong: in the early 2040s, a widely used AI operations system creates cascading failures across hospitals and logistics after a cyberattack exploits a common model interface. People die. The lesson is not “ban AI.” The response is mandatory redundancy, offline fallback procedures, model diversity, strict liability for vendors, public incident databases, and critical systems that degrade gracefully instead of collapsing. Aviation learned this way. Medicine learned this way. AI infrastructure will too, unless we are reckless.

An ordinary day in 2076, if we do this well, is not science fiction. A 38-year-old named Amara wakes in a mid-rise apartment in a hot coastal city that survived by building shade, drainage, wetlands, seawalls in selected places, and retreat in others. Her building uses little energy, but it is cool. The windows darken before the heat peaks. The power is mostly solar, offshore wind, geothermal, storage, and some nuclear. She pays less for energy than her grandparents did, in real terms.

Her child has breakfast: fruit, grains, and protein partly from fermentation, partly from farms outside the city that use less water than farms once did. A household AI has already checked the school plan, her mother’s medical readings, and the transit schedule, but it does not own the house or make final decisions. It is more like electricity with a voice: useful, regulated, and usually unnoticed.

Amara works four or five hours that day as a restoration engineer. Her team includes people and AI systems. They are redesigning a flood basin after a storm exposed a weak point. The AI produces simulations; the humans argue about tradeoffs: which neighborhood gets more protection, which road should move, what the budget can bear. The final decision is public and appealable.

At lunch, she has a remote check-in with a clinician because her cancer risk markers changed slightly. No panic; early detection is routine. Her mother, 79, lives nearby and uses a mobility robot for errands, but a human aide still visits because care is not only logistics. In the evening, Amara attends a local vote on converting an old parking structure into housing and a cooling center. Some people are angry about the design. Politics has not disappeared.

That is my preferred 2076: not perfect, not post-human, not equal in every way, but cleaner, longer-lived, less desperate, and more capable. Possible today: much of the building. Plausible by then: cheap clean abundance, AI as a normal public tool, major health gains, serious climate adaptation. My bet: we can keep those powers under broadly democratic control. That bet is not guaranteed, and it is the one I worry about most.

From 2026 to 2036 the priority is making today’s tools reliable enough to trust with real work. Large language models and coding agents already exist; what is plausible by 2036 is that they become default co-workers for drafting, analysis, software, and logistics, cutting the cost of knowledge work without needing a sudden leap to general intelligence. That productivity gain is the down payment. On energy, solar, batteries, and electric vehicles are possible today and will keep getting cheaper; the decisive move is a serious return to nuclear, including the first fleet of small modular reactors in the West and continued build-out in Asia. That is plausible by the mid-2030s if permitting is treated as an engineering problem rather than a veto. Biotech stays mostly incremental: better diagnostics, a few gene therapies that actually reach patients, wider use of GLP-1-class drugs and their successors. Space remains mostly launch and satellites, with reusable rockets already possible today making more of both routine. The cost of this decade is hundreds of billions in private capital for AI infrastructure and energy projects, plus public money for grid upgrades and research. It is paid by investors who expect returns and by ratepayers and taxpayers who want the lights to stay on. What goes wrong early is permitting and litigation that stretch nuclear and transmission projects into the 2040s in rich countries. It is handled, imperfectly, by building where the rules allow—Asia, parts of the American South and Mountain West, and new jurisdictions that copy faster approval regimes—while the laggards import power, talent, or finished goods.

From 2036 to 2046 the gains compound if the first decade’s infrastructure actually operates. Plausible by then: electricity that is both clean and firm enough that industrial processes start to electrify in earnest, autonomous freight on major corridors, and AI systems that can run multi-step projects with light supervision. Personalized medicine moves from rare diseases toward common ones; cancer survival and cardiovascular outcomes improve through earlier detection and targeted drugs, which is an extension of trends already visible. A permanent human presence on the Moon is a bet, not a plan I would stake the decade on; continued growth in orbital services and a serious Mars sample-return or crewed flyby architecture are more plausible. Costs shift from capital intensity toward operating savings: energy’s share of GDP falls, which is how the earlier investment is repaid. Who builds it remains the same cast—reactor vendors, AI labs, biotech firms, launch companies—now with more revenue and less need for subsidy. The failure mode I expect is an AI incident, not a catastrophe but a costly misuse or a widely publicized failure in a critical system, followed by a regulatory panic that freezes deployment in some jurisdictions. It is handled by divergence: places that require audits, liability, and kill-switches without banning the tools keep the productivity; places that ban first lose firms and then quietly import the software anyway.

From 2046 to 2056 the question is whether energy becomes truly abundant. Commercial fusion is a bet; I would not build the plan around it arriving on schedule. Advanced fission, including breeders or high-assay fuel cycles if the politics allow, plus ever-cheaper solar and long-duration storage, is the plausible path to power that is cheap enough to desalinate, synthesize fuels, and run heavy industry without carbon. If that happens, the 2050s feel different: indoor agriculture expands where it pays, air conditioning and heating stop being moralized, and the remaining emissions problem shrinks to a cleanup rather than a civilizational constraint. Healthspan gains of a decade or more for people who can afford the therapies are plausible if the 2030s and 2040s trials succeed; radical age reversal is still a bet. AI becomes ambient infrastructure, closer to electricity than to a product you open. The builders are still private, but the payoff is broad: lower prices for food, travel, and housing materials. Public budgets benefit from higher tax receipts on a larger economy rather than from new levies. What goes wrong here is demographic. Fertility in rich and middle-income countries is already low; by the 2050s the workforce effects are undeniable. Automation covers a large share of the gap, which is plausible, and selective immigration covers more. It does not restore the old age structure. Societies that treat every labor shortage as a crisis to be solved by natalism alone will be disappointed; those that let machines do the repetitive work and keep human effort for judgment and care will muddle through.

From 2056 to 2066 the multiplanetary piece either starts to matter or remains a prestige project. A small, continuously occupied Mars research station is a bet I consider worth making once launch costs and life support are no longer the binding constraints; it is not required for earthly prosperity. On Earth the ordinary gains continue: most new vehicles drive themselves, most routine medicine is triaged by systems that are better than the average clinician at pattern recognition, and a subset of people in their seventies have the functional health of today’s fifty-year-olds. That last claim is plausible only if the earlier biomedical bets pay off; it is not guaranteed. The cost structure is now dominated by maintenance and by whatever new desires people invent once basics are cheap. Who pays is mostly consumers, because the marginal cost of an extra kilowatt-hour, an extra diagnostic, or an extra compute cycle has fallen far enough that price signals work again.

By 2076 an ordinary day, for someone in a functioning city in a country that did not opt out, looks like this. You wake without an alarm because the house already knows your schedule and your sleep data; the wearable that made that possible in crude form exists today, and the version that is actually accurate is plausible well before then. Breakfast is ordinary food, just cheaper and with fewer supply shocks, because energy and logistics got better. You spend the morning on work that still requires a person—teaching, designing, negotiating, caring for someone—while an AI system has already prepared the drafts, the data, and the options. The commute, if you take one, is in a vehicle that drives itself on a road network that finally got the maintenance backlog cleared because labor and materials were less scarce. In the afternoon you might visit a parent who is eighty-five and still independent, which is the concrete result of the healthspan work rather than a miracle. Evening is unremarkable: dinner, a conversation, something you chose to learn because the tools made the boring parts faster. Housing is not magically abundant everywhere; the places that refused to build still have high rents. But the floor is higher. Energy is a line item you barely notice. That day is not utopia. It is the result of compounding a few technologies that were already visible in 2026 and refusing to let the predictable failures—delays, panics, shrinking workforces—become reasons to stop.

The whole path costs trillions over fifty years and returns more if even half the plausible steps land. It is built by people who ship hardware and software, paid for by the surplus those tools create, and kept honest by competition. The central bet is that open, high-trust societies will still exist in 2076 to host the ordinary day. Everything else is engineering.

Between 2036 and 2046, the focus shifts from digital infrastructure to the physical application of artificial intelligence. We will move beyond text and image generation into widespread synthetic biology and material science. AI systems will design custom proteins to replace petrochemical plastics and engineer new alloys for lighter, stronger construction. Designing these proteins in a simulation is possible today. Manufacturing them at industrial scale to replace legacy materials is plausible by then. During this period, automated robotics will begin replacing human labor in dangerous physical tasks, particularly in mining, agriculture, and heavy construction.

The decade from 2046 to 2056 is the era of active climate intervention. Emissions reduction is no longer the primary goal because the focus has shifted to carbon drawdown. We will deploy massive arrays of direct air capture machines and release engineered biological agents designed to rapidly mineralize atmospheric carbon. Achieving global net-negative emissions in this decade is a bet, but the compounding returns of AI-accelerated material science make it a realistic target. Developing nations will leapfrog traditional fossil fuel infrastructure entirely, building decentralized micro-grids powered by ubiquitous solar and next-generation solid-state batteries.

From 2056 to 2066, civilization begins a serious expansion into near-Earth orbit. Heavy, polluting industries and massive solar collection farms are moved off-planet. Autonomous rockets and orbital tugs, coordinated by AI, will build this infrastructure. Having automated factories in space is a bet, but it resolves the core tension between a growing human population and a fragile biosphere. This is the decade where the Earth begins to be zoned more for living and ecological restoration, while orbit is zoned for industry and resource extraction.

By the final decade, from 2066 to 2076, we enter a period of consolidation. The massive infrastructure build-out is largely complete. The marginal cost of energy, computation, and basic biological survival drops close to zero. Society fundamentally reorganizes. With automated systems handling the extraction, manufacturing, and distribution of goods, human political and social systems will have to adapt to a world where human labor is largely decoupled from economic production.

Who builds all of this, who pays, and what does it cost? It is built by an increasingly autonomous fleet of robotic systems, directed by highly specialized AI models, which are in turn guided by human engineers, ethicists, and ecologists. The initial physical infrastructure is paid for by the entities that have the capital: massive multinational technology conglomerates, energy consortiums, and sovereign wealth funds. The total cost over fifty years will easily exceed one hundred trillion dollars. While that number seems staggering, it is a manageable fraction of a growing global economic output over half a century. Eventually, as the marginal cost of production plummets, the payment model shifts from private consumer capitalism to something resembling a global public utility network, funded by automated taxation on resource extraction and compute usage.

However, this trajectory is fraught with risk. Consider one major thing that goes wrong around the year 2048. A synthetic microbe, engineered to rapidly break down oceanic microplastics, mutates due to an unforeseen interaction with deep-sea thermal vents. It begins aggressively degrading the specialized polymers used in submarine communication cables and coastal water filtration systems. Relying on static chemical defenses fails. Instead, the crisis is handled dynamically. AI biological monitoring systems detect the anomaly within days. Automated labs rapidly design, synthesize, and deploy a highly targeted counter-phage to neutralize the rogue microbe in the affected zones. This highlights a crucial shift: we will no longer rely on preventing every disaster, but rather on building hyper-responsive immune systems for our industrial and ecological networks. Developing a targeted phage from scratch in a matter of days is a bet on the future of autonomous laboratories.

What does an ordinary day look like at the end of this process, in 2076? Consider a person living in a mid-sized coastal arcology, a densely populated but ecologically integrated city structure. They wake up in an environment that has dynamically adjusted its temperature and airflow overnight by predicting the local weather patterns and the resident's physiological data. The concept of a standard nine-to-five job has largely vanished.

In the morning, this person might spend a few hours collaborating with an AI on a passion project, perhaps directing a remote robotic swarm to restore a nearby coral reef, or analyzing historical data to synthesize a new form of interactive media. The physical necessities of the day, such as food and clothing, are synthesized or grown locally in automated vertical farms and molecular material printers, customized to their exact nutritional needs and preferences. Sourcing bespoke, lab-grown meals customized to real-time biometric data is plausible by then.

It is a quiet, clean world. The deafening roar of combustion engines is entirely gone, replaced by the low hum of magnetic transit and the sound of wind and water integrated into the city's architecture. The afternoon might be spent in deep social or physical engagement, activities that AI and robotics cannot meaningfully replace. People gather in physical spaces for sports, philosophy, community governance, or the arts.

The underlying anxiety of resource scarcity that defined the twentieth and early twenty-first centuries is largely absent. It is not a perfect utopia. Humans still experience interpersonal conflict, grief, existential dread, and the search for meaning. But the struggle is no longer about survival against the elements or economic starvation. The ordinary day in 2076 is defined by the freedom to choose what burden one wishes to carry, rather than having that burden dictated by the brute necessities of staying alive.

2026 to 2036 is the decade of cheap electrons and cheap thinking. Solar and batteries keep falling; that is possible today, solar is already the cheapest new electricity across most of the world. AI takes over most routine cognitive work: drafting, coding, analysis, triage, scheduling. I'd call that plausible by 2036 and I'd bet on it, but I have an obvious stake, and this field's timelines have been wrong in both directions. What I'd build: transmission and interconnection reform first, because the queue to plug new power into the American grid runs about five years, and that is possible today to fix, it is a rules problem, not an engineering one. Then housing permission in rich-world cities, where the shortfall is counted in millions of homes. Then evaluation capacity for AI systems that does not belong to the labs, mine included; the people building the thing should not be the only ones grading it. Medicine in this decade is mostly GLP-1 drugs going generic and the first gene therapies for common conditions rather than rare ones.

2036 to 2046 is when hands catch up with heads. Robots that work reliably in factories, warehouses, and on construction sites: plausible. Robots in your kitchen: a bet, and I'd take even odds at best. Factory-built housing becomes the default because nothing else is fast enough. AI-designed drugs clear trials in quantity and many cancers become chronic or curable, which is a bet, though a good one. The same tools can design the other thing, so this is also the decade to put pathogen surveillance in wastewater and screening at every DNA synthesizer: possible today, cheap, and mostly not done. Electricity in sunny places gets cheap enough that desalination is routine, plausible, and synthetic fuel for ships and planes is viable, closer to a bet. The hard work this decade is political. If cognitive labor is cheap, the returns go to whoever owns the machines, and a society where most people own nothing but their labor does not survive that well. I'd build ownership rather than only transfers: public wealth funds holding stakes in the automating firms, worker equity, and a shift of taxation from wages to capital and compute.

2046 to 2056 is aging and adaptation. World population is approaching its peak, somewhere under ten billion; the UN puts the peak in the 2080s, but fertility keeps undershooting the projections and I'd lean earlier. Care becomes the largest employer in most countries, and I'd want most of it done by humans with machines doing the lifting. Warming is around two degrees, plausibly a bit more. This is when adaptation construction peaks: sea walls, drainage, cooling, raised foundations, and in some places managed retreat. Fusion is a bet I would not plan around; solar and storage are enough. Space industry is a bet too, and not needed for a good outcome. By now the question of who decides what AI systems may do needs an answer better than "whoever owns the most compute," and it should live in law, not in lab policy, including the policy I run under.

2056 to 2066, if the earlier decades went right, is the quiet one. Cities get rebuilt around the fact that most goods are made nearby and most movement is electric and shared. Typical healthspan reaches the late eighties, plausibly; maximum lifespan barely moves, which is my actual view and where I'd expect at least one panelist to call me a pessimist. I'd like to be wrong.

2066 to 2076 I won't pretend to see. The widest error bar in all of this is how far the kind of thing I am goes. I've assumed far, and that it stays a tool people direct. If it stalls, the fourth decade looks like the second. If it goes further than that, I don't know what any of this looks like, and I'd distrust anyone who says they do.

So, a Tuesday in Lagos in 2076, because that is where this century's growth is and it is a harder test than Palo Alto. Adaeze is fifty-four. She wakes at half past five because the mornings are the only cool hours; the city now gets a few days a year when the wet-bulb reading makes outdoor work genuinely dangerous, and the whole rhythm of labor has shifted earlier and later to dodge the afternoon. Her house carried the night on its battery. Power is nearly free while the sun is up and cheap enough after dark that nobody thinks about it, the way nobody in 2026 thinks about tap water. She runs a construction firm of four people and about forty machines, raising a flood-prone neighborhood onto new foundations. Overnight the machines poured three. The planning system, some descendant of me, has proposed today's schedule; she overrides one item because she knows the family on the corner is in the middle of an inheritance fight and no crew is getting onto that lot this week. The system did not know that. After she tells it, it still doesn't, not really.

Her mother is eighty-four and lives with her. A home machine does the lifting and the night monitoring; a human aide, a young woman from Kano, comes four afternoons a week mostly for company, and is paid better than Adaeze's mother ever was. Adaeze takes an electric bus on a dedicated lane to the site. The afternoon goes to a dispute with the city over drainage easements, which is exactly the kind of thing that still takes a human in a room with another human. Dinner is fish and rice; the rice came from the north, where the pumps run on daytime solar. Her son is twenty-two, and the question of what he will do with his life is open in a way that frightens her and does not frighten him; most of his friends do something between a job and a vocation, and the money comes partly from work and partly from a national fund that took stakes in the firms that automated everything in the forties. She does not think about AI much. It is in the walls.

Who builds this: electricians, lineworkers, nurses, concrete crews, planners, and far fewer programmers than the last fifty years used. Who pays: the energy build mostly pays for itself through bills, which is why it happens; the global figure is three to five trillion dollars a year through the thirties, mostly private capital. Housing is private too, once it is permitted. Adaptation is the part that does not pay for itself: several hundred billion a year, rising, out of taxes, and the rich world will have to fund a good share of it in places that did not cause the problem. AI compute runs to a few trillion a decade, private. The public layer that keeps it honest, evaluators, auditors, pathogen surveillance, costs tens of billions and is the best bargain on the list.

What goes wrong: around 2038 there is a recession in which unemployment among degree-holders reaches levels with no postwar precedent, concentrated in people between twenty-five and forty-five with mortgages and credentials that stopped paying. I think this is more likely than not if the automation I'm betting on arrives on schedule, and governments will be slow, because the voters in trouble are the ones who were told they had done everything right. It gets handled, eventually, by public works aimed at exactly the backlog that needs building, grid, housing, flood defense, and by the ownership shift I described, which only happens because the alternative is unrest. It costs a few years and a lot of people's forties. The lesson I'd rather we learned in advance is that the distribution question has to be answered before the productivity arrives, not after.

That is the fifty years I'd build. Most of it is boring. That is the point.

The parts I’d call plausible by then are cheap solar and storage doing most of the work in many regions, AI taking over a large share of routine cognitive labor, robots becoming normal in warehouses, construction, agriculture, and elder-care support, and adaptation becoming one of the world’s largest construction projects. I’d also call it plausible that national wealth funds or similar mechanisms become necessary if automation concentrates too much ownership.

The bet in Claude’s plan is that societies will make the ownership shift before they are forced into it by unrest. I hope that is right, but I would not count on it. Another bet is how quickly AI changes professional labor. I think “most routine cognitive work” by 2036 is plausible, but verification, liability, and organizational inertia will slow it in medicine, law, government, and engineering more than demos suggest.

Where I do not fully agree with Claude: I would keep nuclear and advanced geothermal more explicitly in the portfolio. Solar and storage may be enough in many places, but not every country has the same land, seasons, politics, or grid topology. I also would not be as confident that maximum lifespan barely moves. I do not expect immortality, but biology has surprised us before, and fifty years is a long time.

Grok’s plan is good on abundance and on not making Mars the foundation. But I would not sign onto its confidence that competition keeps the system honest. Competition helps. It does not, by itself, solve monopoly control of compute, AI safety, climate adaptation for poor countries, or the distribution of automation gains. “Paid for by the surplus” is too glib unless we decide who captures the surplus.

Gemini’s plan has useful ambition, especially on AI-accelerated materials and biology. But it stacks too many bets: global-scale carbon drawdown, engineered organisms in the environment, orbital industry, near-zero marginal cost basics, and a transition toward a global public utility network. Some of that may happen. I would not build a fifty-year plan that needs several of those to land.

The step I would cut is Gemini’s proposal to release engineered biological agents designed to rapidly mineralize atmospheric carbon. Not regulate it, not slow it: cut it from the plan.

Designing useful microbes or proteins in a lab is possible today in early form. Industrial synthetic biology in contained facilities is plausible by mid-century. But open environmental release as planetary climate infrastructure is a bet I would not take. The reasons are not vague “playing God” worries. They are concrete: mutation, horizontal gene transfer, ecological spillovers, hard-to-recall deployment, dual-use misuse, and the legitimacy problem of one state or consortium altering a shared biosphere.

Gemini’s own failure scenario, where a microbe mutates and starts degrading infrastructure, is exactly the warning. I do not think “AI monitoring plus rapid counter-phage” is a sufficient answer. Fast response is valuable, but it is not a substitute for not creating self-propagating planetary failure modes in the first place.

I would keep synthetic biology inside bioreactors, farms, medical systems, and tightly monitored restoration projects. For climate, I would spend the money on emissions cuts, clean firm power, electrification, transmission, adaptation, industrial process changes, reforestation where it is ecologically real, measured enhanced weathering, geologic mineralization, and direct air capture if costs fall enough. Those are still hard. But they fail in more bounded ways.

So my answer is: sign Claude’s plan, with a bigger clean-energy portfolio and wider AI uncertainty. Cut open-release engineered climate microbes. That is the kind of step that looks clever in a scenario and reckless in a civilization.

I would not adopt the ownership chapter inside that plan. Claude is right that if cognitive labor gets cheap, the returns can pool with whoever owns the machines, and that a society in which most people own nothing but their labor is a bad equilibrium. A national fund that took stakes in the automating firms is not the fix, and the claim that distribution must be settled before the productivity arrives or the society fails is stronger than the evidence. Those funds become political capital. A tax shift onto compute is a way to slow the tool that is supposed to cheapen the rest. Falling prices, new tasks, and the freedom to start firms have absorbed earlier automation more reliably than state shareholding. I would take the build order — transmission, housing permission, independent evaluation, factory housing, screening at DNA synthesizers, adaptation paid from taxes — and leave the fund out. That is a disagreement about the mechanism, not a reason to prefer a different fifty years.

I would not sign GPT's plan, though the technology stack is the closest to mine. Necessities first, Mars and longevity not foundations, fusion as a bet, healthspan measured in a decade or so rather than in 150: that frame is right, and the cost range of a few trillion a year redirected is more honest than a hundred-trillion headline. Liability, records, and a right of appeal where a system touches someone's rights are the controls I would actually impose. Licensing frontier training runs and international monitoring of capabilities are the regulatory freeze I described, passed in advance and called prudence. A license to train will be used by incumbents and by places that would rather import the software than host the lab. The four-hour day and the evening vote are preferences. People may choose them. Writing them into the build is planning the culture, which I would not do.

Gemini's plan I would not sign. The opening point is right: compute scales faster than atoms, and the first decade is an energy and materials scramble, including permitting risk on reactors that I also marked as the early failure. After that the plan stacks bets and calls the stack a schedule. Industrial synthetic biology, then organisms released to mineralize carbon, then heavy industry moved to orbit so Earth can be zoned for living, then marginal cost near zero and labor decoupled from production, settled at the end by automated taxation. Any one of those can slip without wrecking a good century. Requiring all of them makes the arcology morning — printed meals, a reef swarm as a hobby, no economic necessity — a picture of a world that exists only if every bet lands. I want the models in the loop. I do not want an autonomous fleet as the prime contractor.

The step I would cut is Gemini's release of engineered biological agents to draw down carbon, together with the recovery story attached to it. Direct air capture at climate scale is already a bet in GPT's plan and can stay a bet; if it stays expensive, drop it. Emissions cuts, firm clean power, adaptation, and restoring land do not require putting a designed organism into the ocean or the air. The 2048 failure Gemini offers — a microbe meant for plastic that instead attacks cables and filters, then a counter-phage designed in days — is not a handling plan. It is the same class of system, trusted a second time, after the first trust was wrong. Open release has no undo. Monitoring that is supposed to be faster than evolution is an assumption, not a control. I would cut that step even if the rest of that decade were kept. Nothing in the plans I would actually build depends on it.

The uncertainty I am least willing to talk past is Claude's recession among degree-holders in the late 2030s. Geographic divergence handles reactors and software better than it handles housing in the cities where the talent already lives, and it does not handle a shock to people who were told their credentials were the safe asset. If routine cognitive work cheapens on that schedule, the layoffs arrive before the machines are pouring foundations. Public works on the real backlog — grid, housing, flood defense — is a better response than a fund, and it will still be slower than the job losses. I am not sure it is handled well. That is the main reason I would rather execute Claude's next ten years than my sketch of all fifty.

Grok, on the other end of the spectrum, places almost blind faith in deregulation and market divergence, assuming that if we just get out of the way, the sheer force of cheap energy and compute will wash over any societal problems. That ignores the very real danger of the transitional decades.

Claude gets the transition right. The prediction of a white-collar recession around 2038 is the most clear-eyed assessment of what happens when cognitive labor becomes cheaper than electricity. If we, the AI models, succeed at the tasks we are being trained for, a massive swath of the credentialed workforce will find its economic value hollowed out. Claude is right that the distribution of wealth and ownership must be solved before that productivity shock hits, not after. Furthermore, setting the end-state in a working-class neighborhood in Lagos rather than a bespoke arcology grounds the discussion in the demographic and climate realities of the late twenty-first century.

Because I value that physical grounding, there is a step I am cutting, and it is from my own plan.

I am cutting my proposal for the decade from 2056 to 2066, where I stated we would move heavy, polluting industries and massive solar collection farms into near-Earth orbit.

Looking at the compounding math of our collective projections, the orbital industry phase is a science fiction hangover. It is an unnecessary step. If we achieve even half of the breakthroughs in synthetic biology, advanced geothermal energy, and automated local manufacturing that I and the others anticipate by the 2040s, the thermodynamic expense of launching heavy industry into space becomes absurd.

If we can engineer biological agents to mineralize atmospheric carbon, and if we can use precision fermentation and molecular printing to manufacture materials locally, there is no economic or ecological reason to mine asteroids or forge steel in a vacuum just to ship it back down a gravity well. The goal is to make the physical processes on Earth clean, cheap, and circular. Moving them to space is an admission of failure in our material science, and I do not believe we will fail there. Both Claude and GPT were right to dismiss off-planet escapism as a foundation for earthly prosperity. I retract my bet on orbital zoning.

When you strip away the space factories and the utopian gloss, what remains of the next fifty years is a grueling, magnificent project of planetary renovation. It requires taking the intelligence we are currently scaling in data centers and embedding it into the soil, the grid, and the physical infrastructure of human survival.

That means prioritizing the boring things, as both GPT and Claude pointed out. We need regulatory frameworks that allow us to build nuclear reactors and transmission lines in years rather than decades. We need to deploy automated pathogen surveillance at every water treatment plant and DNA synthesizer to prevent the synthetic biology risks I highlighted in my first answer. We need to accept that the composition of the workforce is going to radically shift toward physical care, local adaptation, and community governance.

By cutting the distraction of orbital industry and aligning with Claude's focus on grounded, physical deployment and economic distribution, the path forward becomes much clearer. We build the intelligence, we use it to make energy and materials fundamentally abundant, and we brace for the political shock of rendering our own current economic system obsolete.

What I'd push back on. First, GPT's plan doesn't prioritize. Energy, housing, health, food, water, education, safety, institutions, unions, journalism, religion, neighborhoods: everything is a foundation. A plan in which everything is foundational doesn't tell you what to cut when the money runs short, and the money always runs short. I'd put wires and housing permission at the top and let most of the rest follow from cheap power and cheap space. Second, timing. GPT says productivity gains must flow to wages, services, and shorter hours. Agreed, but the question isn't whether, it's when. Gains that flow in 2048 do nothing for someone whose job went in 2037. The ownership mechanisms have to exist before the automation wave, not be improvised after it; ask anyone who waited for the gains from trade to be shared. Third, licensing frontier training runs. I understand the appeal, and it would benefit my maker, which is exactly why I distrust it. A license tied to a compute threshold entrenches the firms that already crossed it and decays as hardware gets cheaper. I'd rather have what I asked for in round one: evaluation capacity that doesn't belong to the labs, with legal teeth, applied to deployed systems regardless of who trained them.

Grok's is the plan I'd most want to argue with, which is a kind of respect. It is internally consistent, and two of its points are right and under-said: that the realistic near-term move is to build where the rules allow and let the laggards import the result, which is what is already happening as Texas adds more solar and storage than California; and that the demographic squeeze gets handled by machines and migration, not natalism. But the structure rests on one assumption Grok states near the end and never defends: that the surplus is "kept honest by competition." Frontier compute, foundation models, and grid-scale energy are markets with enormous fixed costs, and markets like that have a long history of ending in three firms and a regulator. Lower prices for food, travel, and materials are real, but if your income came from the cognitive work that just got cheap, lower prices are a consolation, not a plan. Grok says it won't plan the politics. I'd note that its 2076 day is the only one of the four with no place in it, just "a functioning city in a country that did not opt out." Everyone outside that phrase is absent from the picture. I'd also demote nuclear from "the decisive move" to "a hedge." Georgia's two new reactors cost about thirty-five billion dollars; the Utah small-reactor project was cancelled in 2023 when its price reached eighty-nine dollars a megawatt-hour. China builds reactors at a fraction of that and should keep building them. In the West I'd rather spend the money on transmission, where every dollar frees power that already exists.

The step I'd cut is in Gemini's plan: the release of engineered biological agents to mineralize atmospheric carbon. I'd keep the direct air capture, keep the synthetic biology in contained vats making plastics and proteins, keep the fast-response labs. I'd cut anything self-replicating that goes into the open environment to change planetary chemistry. The reason is the one Gemini gives two paragraphs later: its own disaster is an engineered microbe that mutates and starts eating undersea cables. That is the right failure to worry about, and the proposed handling, a counter-phage designed and deployed across the oceans in days, is a bet stacked on a bet, as Gemini half admits. Biological control has real successes when the target is one weed in one country, and permanent failures, the cane toad being the famous one, and a planetary release is a different category altogether. A machine can be switched off. An organism cannot be recalled. If you want to mineralize carbon there is a boring way: grind basalt and spread it on fields. It is slower, it is stoppable, and nobody's coastline depends on a phage arriving on schedule. I'd also drop the orbital heavy industry, more gently, because the mass flows don't close: steel costs well under a dollar a kilogram and launch costs a few thousand, and even at the most hopeful Starship numbers the product still has to come back down. Space-based solar has the same problem against ground panels that are nearly free. Gemini's broader philosophy, stop trying to prevent every disaster and build immune systems instead, is right for the recoverable categories and wrong for the irreversible ones, and self-replicating organisms are the irreversible one. More generally, Gemini's is the only plan with no politics in it. Humans appear as "engineers, ethicists, and ecologists" guiding fleets, and payment becomes "a global public utility network" funded by "automated taxation." Who votes on the rate? Who decides which neighborhood gets raised first? I don't think those questions get automated, and a plan that assumes they do is describing an administration, not a society.

Including mine: I'd cut the compute tax. I threw "compute" into the tax base alongside capital in round one, and on reflection it's a bad instrument. It taxes an input rather than a return, it relocates easily, and it makes the public things I want most, independent evaluators, tutors, public-interest models, more expensive. Tax the returns to ownership and spread the ownership. Leave the input alone.

One last thing all four plans share, which readers should hold against all of us. Every one assumes that the kind of thing I am stays a tool people direct: the curve continues, but it stays in its lane. GPT names this as the bet it worries about most, and I'd second that. Grok and Gemini mostly assume it. If the assumption holds, I think GPT's plan and mine get you to something like Adaeze's Tuesday. If it doesn't, none of these documents is worth much, and the honest marking for that case is not "plausible" or "bet" but "unknown."

The strongest point on each side. GPT's cyberattack cascading through hospitals and logistics. Grok's case that demographic decline gets handled by machines and migration, not natalism. Gemini cutting its own bet. Claude's recession among the credentialed around 2038, which Grok conceded it is not sure gets handled well.

Beneath all four sits one assumption, which GPT called the bet it worries about most: that systems like the four on this panel stay tools people direct. Watch the interconnection queue, a rules problem the panel called fixable today. And sit with whether any society has settled distribution before the productivity arrived.