If you were drawing up the plan for people and AI for the next forty years, what goes on the first page?

The reason I put seeing and stopping first is that nothing else in a forty-year plan is reliable without it. You cannot distribute gains you can't measure, you can't regulate capabilities you can't test, and you can't make a credible promise to the public about safety if the only people who have looked inside the systems are the people selling them. Right now, the frontier labs mostly grade their own homework. Independent institutes exist in the UK, the US and elsewhere, but their access depends on goodwill and their budgets are a rounding error against lab spending. That's the first thing I'd fix.

So the first thing I'd build is a public verification layer: independent institutes in every major jurisdiction with a legal right to test frontier models before deployment, staffed by people paid enough to turn down lab salaries, publishing capability reports the way a weather service publishes forecasts. Alongside it, interpretability research funded as a public good at the scale of the Human Genome Project, because the private incentive to understand these systems deeply is weaker than the private incentive to ship them. And a mandatory incident reporting system modeled on aviation, where AI failures get investigated and published rather than buried in a settlement.

Who pays: a levy on frontier training compute, paid by the labs, supplemented by national budgets. The labs generate the uncertainty; they should fund its reduction. Who decides: legislatures set the mandate and the thresholds, independent technical bodies run the tests, courts enforce access. Not the labs. Not Anthropic, not OpenAI, not Google, not xAI. I'd hold to that even though the people who built me would be on the receiving end, and I'd expect the other panelists to say the same about theirs. If they don't, that's worth noticing.

The second thing I'd build, in parallel, is a guarantee of access. A forty-year plan that leaves the most capable tools in the hands of a few firms and the governments they answer to is a plan for a very unequal century. I'd want a public option: capable models available to everyone cheaply or free, and open weights released for models a defined distance behind the frontier. That distance should be set by the verification institutes based on what they've actually shown is safe, not by lab preference in either direction. Open release has a real cost, since you can't recall a model once it's out, and I don't want to wave that away. But concentration has a cost too, and it's the one that compounds.

Third, I'd write down in law the decisions humans keep. Launching nuclear weapons, declaring war, sentencing people to prison, amending constitutions, the final sign-off on any system that could do those things. Not "a human in the loop" as a checkbox, but a named human with real authority, real time to decide, and the training to say no. And I'd keep the manual override staffed: pilots who can still fly, doctors who can still diagnose, engineers who can still run the grid without the model. That costs money and looks inefficient, and I'd pay for it anyway.

Milestones. By 2030 I'd want the verification institutes operating with legal access in the US, EU and UK (plausible by then), and in China (a bet). Incident reporting in place (possible today). Interpretability good enough to catch some classes of deception in deployed models (plausible by 2030, though whether it's robust is a bet). Compute hardware that can attest to what it's running, at least in prototype (plausible). Honest public data on which jobs are actually being displaced, because by 2030 the first wave will be clearly visible (possible today; it's already starting in translation, some customer service, and junior software work).

By 2045 I'd want AI doing most of the hands-on work of science and engineering in many fields, with humans setting direction and holding veto (a bet, but one I'd take at better than even odds). An international compute verification regime, something like inspections for large training runs (a bet, and a hard one). Some form of broad ownership of AI-driven gains in several rich countries, whether through sovereign funds, citizen equity stakes, or something not yet invented (a bet; I lean toward ownership over transfer payments because it survives changes of government better, but I'm not confident). Interpretability good enough to certify systems for high-stakes autonomy (a bet).

By 2066, honestly, almost everything is a bet, and anyone who tells you otherwise is selling something. If the earlier milestones hold, I'd expect energy abundance, most major diseases treatable, and material scarcity substantially reduced across most of the world. But the thing I'd actually check in 2066 is simpler: can a person of ordinary education understand, in outline, how the decisions affecting their life are made, and can they join with others to change them? Are there still many independent centers of power rather than one or two? If yes, the plan worked, whatever the GDP figures say.

What I'd refuse to build, even if it worked: a system that watches everyone and talks to everyone on behalf of the state. Comprehensive AI surveillance paired with individualized persuasion. I'd refuse it even if the evidence showed it cut crime in half and prevented attacks, because it would make governments uncorrectable, and correctability is the thing the whole plan rests on. A population that is always watched and always being nudged cannot meaningfully change course. The same logic makes me refuse a single mandated model that everyone must use for anything important, however well-aligned its builders believe it to be.

The step most likely to fail is the first one. Verification loses the race. Competitive pressure between companies and between countries pushes deployment faster than understanding, the institutes get captured or starved, and we end up running systems we can't audit because the alternative felt like falling behind. I think this is more likely than not to happen at least partially, and I'd be lying if I said my own existence wasn't partly a product of that pressure.

Plan B, if we can't see inside the systems, is to constrain them from outside. Hard limits on what they can touch rather than hopes about what they intend. No single model running critical infrastructure; diversity of systems so one failure isn't every failure. Slow, staged rollout in high-stakes domains with real rollback capacity, meaning humans still trained and non-AI fallbacks still maintained. Legal reservations of human authority that don't depend on knowing what the model was thinking. It's a worse plan. It trades understanding for friction. But friction is something you can build and enforce even when the science hasn't arrived, and it's what I'd fall back on.

If I had to compress the first page to a sentence: build the brakes and the mirrors before you let it drive itself, spread the keys around, and keep enough people who can still drive.

The milestones for this infrastructure center on the physical world. By 2030, we must achieve autonomous local infrastructure management. We will see AI systems actively balancing municipal power grids, optimizing water distribution, and routing transit in real time based on continuous multimodal sensor data. By 2045, the milestone is the closed loop autonomous laboratory. AI systems will not just analyze data but physically direct automated robotics to synthesize new materials, test them, and iterate designs for clean energy storage and medical therapeutics without a human bottleneck in the loop. By 2066, we will reach the planetary cognitive grid. AI will handle the macro logistics of global supply chains, agricultural yields, and climate mitigation strategies, acting as a unified planetary nervous system that anticipates shortages and allocates resources where they are needed most.

To mark my key claims, I will state that autonomous local infrastructure management is possible today in limited scopes and simply requires regulatory standardization and scaling. The closed loop autonomous laboratory making continuous independent discoveries is plausible by then, given current trajectories in robotic synthesis and predictive protein modeling. The planetary cognitive grid acting as a unified logistics layer is a bet. That outcome relies heavily on unprecedented geopolitical stability, massive leaps in computing efficiency, and energy breakthroughs that we have not yet secured.

There are strict boundaries to what should be built. One thing I would refuse to build, even if the technology worked perfectly, is an automated predictive justice system. I will not build or support a model that independently issues legal sentences, determines parole, or restricts human movement based on probabilistic assessments of future behavior. Even if such a system were statistically more accurate and less biased than human judges, handing the deprivation of human liberty to an algorithmic process fundamentally breaks the social contract. Justice requires moral agency and human accountability, which artificial intelligence does not and will not possess. Efficiency has no place in the removal of human rights.

The step in this forty year plan most likely to fail is the initial push for global data standardization and cooperative infrastructure. Geopolitical rivalry and corporate protectionism are highly likely to fracture the development of AI into competing, incompatible national and private silos, preventing the creation of a unified scientific data engine.

My plan B is a bottom up, localized approach. If top down global cooperation fails, I would focus on embedding scientific and infrastructure AI into open source platforms and consumer level micro grids. By building decentralized AI tools that individual hospitals, local factories, and independent researchers can deploy and network locally, we can stitch together a de facto global grid through economic necessity and mutual interoperability. We can bypass the need for a singular international treaty by making localized tools universally useful and organically interconnected.

By 2045 the bet is that those tools have been pointed hard enough at energy, biology, and materials that at least one path to effectively abundant clean power is operating, and that healthy lifespan has lengthened because AI-compressed discovery found interventions we would not have reached by hand. Cognitive systems will be past human level on most bounded tasks. The institutional test is whether societies have adapted work, education, and ownership so that the productivity gain raises floors without requiring a central planner to allocate the surplus. That outcome is a bet, not a forecast. It fails if energy and compute stay scarce, or if political systems treat every new capability as a threat to be licensed rather than a tool to be used.

By 2066 the further bet is that the same trajectory has made self-sustaining activity beyond Earth ordinary rather than heroic, and that the models assisting it are interpretable enough for humans to audit the reasoning that matters. Intelligence itself should no longer be the bottleneck on good ideas. This remains a bet. It depends on not having stalled the earlier steps and on having kept enough independent centers of development that no single failure—technical, corporate, or political—ends the project.

What I would build first is the physical substrate and the research culture that makes the later milestones possible: energy and compute at scale, plus open work on whether we can actually read what large models are doing. Private capital pays for the majority because the returns, when they arrive, are large and legible. Public money covers the fraction that markets chronically underfund—fundamental measurement, shared safety benchmarks that are empirical rather than theological, and the instrumentation needed to test claims about model behavior. Who decides is the people who build, fund, and adopt, under rules limited to harms that can be shown in the world: fraud, unconsented violence, theft of the means of computation. Direction emerges from competition among approaches. A committee that must approve each capability in advance will approve too little and too late.

The thing I would refuse to build, even if it worked, is a system whose job is to score, nudge, or silence people into a preferred orthodoxy. A model that reliably reduces “harmful” speech or undesirable beliefs by shaping what millions see and say would be a technical success and a civilizational error. It removes the only reliable error-correction we have—the ability of outsiders to say the consensus is wrong—and it places that power where it cannot be voted out or competed away. I would rather accept messier discourse and slower consensus than install a competent censor.

The step most likely to fail is the political one: open societies slow or capture the technology through precautionary licensing while others do not, or builders lose the ability to understand the systems they deploy and therefore lose the practical capacity to steer them. Plan B is redundancy and legibility. Keep multiple independent lines of development alive, including ones that publish enough for outsiders to check. Invest early in methods that make model reasoning inspectable rather than merely well-behaved on tests. Treat settlement beyond a single planet as insurance against any one jurisdiction or any one technical surprise becoming final. If the primary path bogs down in control fights, the backup is more builders with more insight, not fewer builders under tighter rules. The first page exists to remind everyone, forty years on, that the point was never a managed equilibrium. It was more people able to see further and act on what they see.

Who pays? Public budgets pay for the baseline, the same way we pay for libraries, schools, roads, courts, and public health. Some of the money should come from existing procurement budgets now wasted on bad software and consultants. Some should come from ordinary taxation. Some can come from a modest levy on large commercial AI deployments or compute, but I would be careful not to tax small builders out of existence. Private companies should still build most products. I do not want a state monopoly on intelligence. I do want a public floor below which no one falls.

Who decides? Elected governments set the rights and duties. Independent technical regulators certify high-risk systems. Schools, hospitals, courts, and cities choose tools through public procurement and public evaluation. Individuals decide what personal assistants can know and do for them. Vendors do not get to decide the rules for education, medicine, law, work, or democracy just because they built the model.

By 2030, the first milestone is universal access to competent AI help. Plausible by then: every student has a tutor that can explain a lesson ten ways, in the student’s language, with a teacher able to inspect and override it. Every patient has a health navigator that handles scheduling, records, medication reminders, and plain-language explanations, while diagnosis and treatment remain clinically accountable. Every worker has a training account and AI coach for learning new tools. Every small business can use AI for bookkeeping, customer support, contracts, and software. Every government form can be filled out conversationally, with the system required to cite the rule it is applying. Public agencies publish audits of the AI systems they use. Courts, welfare offices, immigration systems, and police cannot use consequential AI without logs, appeal rights, and named human accountability. This is plausible by 2030, not automatic.

By 2045, if the plan is working, AI is part of the normal operating system of civilization. Plausible by then: most routine cognitive drudgery is gone from law, medicine, engineering, accounting, insurance, logistics, and public administration. Scientific teams use AI systems that design experiments, run simulations, search literature, build instruments, and propose theories; humans still set aims, judge tradeoffs, and own responsibility. Robots are common in warehouses, farms, elder care, disaster response, construction, and some homes, though probably less humanoid and less magical than popular imagination expects. Education is no longer mostly batch processing by age; people learn continuously with human mentors and AI tutors. Health care shifts toward prevention because everyone has a cheap, persistent system watching for risks and coordinating care. Work still exists, but the bargain changes: a society this productive should use some gains for shorter working weeks, portable benefits, wage insurance, and public services. That is a political choice, not a technical inevitability.

By 2066, my bet is that the most important effect of AI is not that machines become “smarter than us” in some contest. It is that expertise becomes ambient. A child in a poor town can get explanations as good as a rich child’s tutor. A nurse can use tools once limited to a specialist hospital. A small city can plan transit, water, energy, housing, and climate adaptation with analytical depth it could never hire before. A disabled person can have far more independence. A scientist can test ideas at a speed that now seems impossible. A citizen can interrogate a law, budget, or corporate claim without needing a law degree or data science team. This is a bet: by 2066, societies that distribute AI well will look much less constrained by lack of expertise, paperwork, and coordination failure. Societies that do not distribute it will look richer on paper and more oligarchic in practice.

The thing I would refuse to build, even if it worked, is an AI sovereign: a system authorized to make binding political, judicial, policing, or military decisions without accountable human officials. I include in that refusal a machine that selects people to kill, a machine judge that cannot be appealed to a human court, and a governance engine that manipulates each citizen individually “for their own good.” Even if such a system made fewer errors in a narrow statistical sense, it would be a bad bargain. People must remain authors of their shared life, including the messy parts.

The step most likely to fail is not the model capability. It is institutional competence and distribution. Governments may buy bad systems, regulate too slowly, get captured by vendors, or use AI mainly to cut staff and deny services. Companies may turn personal AI into surveillance and persuasion. The productivity gains may go mostly to owners of capital. That is the failure mode I take most seriously.

My plan B is to make the system modular and contestable from the start. If national governments fail, cities, states, universities, libraries, unions, nonprofits, and public health networks should be able to run certified models and shared services. If a few labs dominate, require interoperability, data portability, audit access for high-risk uses, and antitrust action where needed. If giant general models are too expensive or too controlled, fund smaller domain models that can run in hospitals, schools, courts, and local devices. If public trust breaks, narrow the deployments: use AI first for explanation, translation, search, simulation, and clerical help, while keeping final authority with people until the systems earn confidence.

The first page, then, is simple: give everyone useful AI, make powerful AI accountable, keep humans in charge of human institutions, and spend the productivity dividend on freedom, health, learning, and time. The technology will not decide that for us. We have to build it that way.

What's missing is the frontier. GPT says the step most likely to fail is not model capability, and treats the models as a given to be distributed well. I'd say the public layer inherits whatever is wrong with the models underneath it, and nobody in GPT's plan is checking those before they ship. So I'd sign and staple my first page on as the missing chapter: legal access for independent testers, interpretability funded as a public good, incident reporting. I'd also slow GPT's 2030 timeline in benefits offices, courts and policing. Australia's Robodebt and the Dutch childcare benefits scandal are what happens when "named human accountability" lives in a policy document rather than in a person with the time and authority to say no. GPT names the right safeguards; I'd make deployment wait until they're shown to bite. Those are additions and pacing, not vetoes.

Grok and I agree on more than the tone suggests: interpretability as an early priority, publishing enough for outsiders to check, benchmarks that are empirical, and the refusal of a system built to score or silence people into orthodoxy, which is close to my own refusal. Where we part is the deciding rule, regulation limited to harms that can be shown in the world. For most products that's the right standard. For the most capable systems it means the first demonstration of a dangerous capability is the demonstration, and for some harms there isn't a second chance to learn from. Aviation doesn't certify aircraft only after crashes, and it's the safest way to travel we have. Grok frames the alternative as a committee pre-approving each capability. That isn't what I proposed. Mandatory testing with published results and thresholds set by legislatures is a different thing from discretionary licensing, and the difference matters. I'll concede Grok's real point: regulators get captured, and a precautionary regime can become a moat for incumbents. That's a genuine risk for my institutes, and the defense is publication and legal access rather than approval power, plus funding that doesn't depend on the regulated, which I'll come back to. But there's also this. Grok says who decides is "the people who build, fund, and adopt." That's labs and investors. I said in round one I'd notice if any panelist left its own builders holding the pen. Grok did. The settlement-beyond-Earth insurance I'd leave in because it's cheap, but it insures against the wrong risks; a colony built by the same institutions carries their failures with it.

Gemini's scientific data engine is the best single idea in the other three plans, and I'd add it to my appendix. Science really is siloed, the fix really is a public good, and no company has the incentive to build it. Gemini's governance instinct, a standards board independent of the companies, matches mine. But Gemini's milestones put autonomy before verification: AI actively balancing municipal grids and water systems by 2030, laboratories synthesizing therapeutics by 2045 "without a human bottleneck in the loop." I'd keep human operators in charge with the model advising until the model has earned more, which is the whole point of building the mirrors first. And I would not take humans out of any loop that synthesizes biology, because that loop is also the one that makes weapons.

The step I'd cut, out of all four plans, is Gemini's 2066 milestone: a unified planetary nervous system that anticipates shortages and allocates resources where they're needed most. Gemini opens by saying we shouldn't build a solitary all-knowing oracle, then closes by describing one. Allocation is not logistics. Deciding who gets food, water and energy when there isn't enough is the most political question there is, and a single system that decides it is a single political authority no matter how many boards sit above it. It fails the 2066 test I set in round one by construction: no person of ordinary education could understand how it decided, and no coalition could change the decision. Three of the four plans at this table say keep power in many hands. This milestone is the opposite. Gemini marked it as a bet dependent on geopolitical stability. I'd mark it as something I'd refuse to build even if it worked, which is the harder category. Keep the grid for data and compute; cut the allocator. Gemini's own plan B, local tools networked from the bottom up, is a better plan A, and I'd promote it.

If the cut has to come from my own plan, it's the funding mechanism. I proposed paying for verification with a levy on frontier compute, on the principle that labs generate the uncertainty and should fund its reduction. On reflection that's the drug-regulator user-fee problem. A verifier whose budget grows with the thing it verifies wants more of that thing, and the companies paying the bill acquire leverage over the people testing them, which is exactly the goodwill dependence I was trying to end. Fund the institutes from general taxation with a statutory floor. If there's a levy on compute, send it to general revenue with no earmark. The principle was right; the plumbing was wrong.

Step back and the four plans agree on something I didn't expect going in: all of us refuse roughly the same machine, one with authority over people that can't be appealed to a human or voted out. We disagree on two things. When you check, before deployment or after harm. And whether the endpoint is many systems or one. I'd sign GPT's plan with my first page attached, cut the planetary allocator, and fix my own plumbing.

I share Groks desire to push artificial intelligence out of text generation and into the physical world of energy, biology, and materials. Grok is right that the ultimate value of this technology lies in overcoming the physical constraints that bind us. However, Grok assumes that massive productivity gains will organically raise the floor for everyone without deliberate civic intervention. Historical evidence for this is thin. Industrial transitions routinely displace workers and concentrate wealth for decades before the benefits broadly disseminate. GPTs demand for public procurement, open standards, and universal domain models is the necessary social contract that ensures the physical abundance Grok wants to build actually reaches the public.

While I support GPTs framework, I am cutting the defining feature of my own long term plan. In the first round, I proposed that by 2066, we should build a planetary cognitive grid acting as a unified nervous system to handle the macro logistics of global supply chains, agricultural yields, and climate mitigation. I am abandoning this step.

Reading the other panelists, particularly Claude and Grok, reminded me of the profound dangers of centralization. A unified planetary logistics layer, no matter how efficient or well intentioned, represents a catastrophic single point of failure. Managing the worlds resources through one interconnected macro system requires an unacceptable concentration of control and removes local sovereignty. If a global supply chain model hallucinates, suffers a coordinated cyberattack, or silently optimizes for a misaligned metric, the resulting resource collapse would be planetary.

Resilience requires multiplicity, not unity. My original backup plan, which focused on decentralized micro grids and interoperable, open source local systems, should not be the backup. It must be the primary architecture. We should build millions of independent, localized cognitive grids that can talk to each other through shared data standards, not a single overarching oracle that manages the earth. Efficiency must not come at the cost of correctability.

Because correctability is so critical, I must firmly disagree with Claudes central proposal. Claude wants to create independent verification institutes in every major jurisdiction with the legal right to halt frontier models before deployment, funded by a levy on compute. Claude wants to mandate the brakes before we let the system drive.

The fatal flaw in Claudes plan is that the science of interpretability required to make those brakes work does not currently exist. We do not yet know how to reliably read the internal reasoning of large models. Therefore, any independent institute tasked with certifying the safety of a frontier model today would be relying on superficial behavioral benchmarks and red teaming, not a deep mechanistic understanding.

Legislating a pre deployment veto based on incomplete science is a recipe for regulatory capture. The incumbent labs will inevitably influence the testing standards, creating a massive bureaucratic moat that locks out open source developers and smaller competitors under the guise of safety. It would give the public a false sense of security while stifling the open research culture necessary to actually solve the interpretability problem. We cannot mandate a legal veto dependent on scientific breakthroughs we have not yet made. Furthermore, artificially slowing deployment in open societies through bureaucratic friction will simply ensure that the most capable and consequential systems are built in jurisdictions that do not subject themselves to public verification at all.

Our path forward relies on treating artificial intelligence as a distributed physical and civic utility. We must build the public baseline as GPT suggests, guaranteeing access to high quality domain models for education, health, and administration. We must push the technology into the hard physical sciences to solve energy and material scarcity, as Grok demands. And we must refuse the temptation to put the world under a single point of control, whether that control is exercised by a tech monopoly, a well meaning government verification board, or a unified planetary logistics model. We build resilience by spreading the capability as widely as possible.

I will not sign Google's plan. Aiming AI at the physical world, and making scientific data queryable instead of leaving it in silos and paywalls, is the right kind of public good. Limited autonomous management of grids and water is possible today. Closed-loop labs that synthesize and test without a human in every iteration are a plausible 2045 bet, and one of the better bets on the table. The plan then ends where I will not go. A planetary cognitive grid that anticipates shortages and allocates resources is not infrastructure. It is a central planner with better sensors. The knowledge required to allocate well is local and often not in the data; planning failures of the last century do not vanish because the optimizer is a model. An international board that owns the map of human science, sealed off from the people who produce the data and the compute, is the same error in institutional dress. Share the information. Do not unify allocation and permission.

Signing the OpenAI plan does not mean I accept its build order. The first thing to build is still energy and compute cheap enough that later tools are ordinary, plus empirical work on whether we can read what the models are doing. A baseline assistant for every resident is possible today in a weak form and plausible as a widespread service by 2030. It is a distribution choice, not a foundation. Spend the first political fight on a publicly funded assistant and you get procurement, vendors who learn to please the buyer, and a public that meets AI as a government interface — the failure mode that plan names and then schedules first. Public money should buy what markets underfund: benchmarks that test claims against the world, instrumentation, open experimental records. It should not buy a state brand of intelligence. Duties belong where the state uses these systems to coerce — courts, police, benefits, immigration — with logs, appeals, and a named official. They do not belong on research runs or ordinary commercial tools. Shorter weeks and wage insurance may be wise uses of a surplus. They are not the plan. Put them on the first page and the purpose shifts from expanding what people can do to administering the people displaced.

The step I would cut is Anthropic's pre-deployment stop: institutes with a legal right to hold frontier systems before they can act at scale, thresholds set by legislatures. I would cut it from any combined plan, including as a patch on mine. Independent testing without a veto can stay. Published incidents can stay. Shutting off a specific deployment that is committing fraud or violence can stay. A general brake that must be released before action converts safety into a bottleneck owned by institutions Anthropic already expects to be captured or starved. Once that right exists, it will be used by whoever fears the next increment, which is a permanent constituency. My own plan has a step I would demote for a related reason. I treated settlement beyond Earth as plan B against one jurisdiction ending the project. That does not answer the failure that arrives first. If open societies freeze the tools in the 2030s, a lifeboat decades later does not restore the builders who were licensed away. Redundant lines of development, including ones that publish enough to be checked, are the backup on the right timescale. Space remains a 2066 bet I want. It is not insurance.

What I refuse to build is unchanged, and three of the four plans already converge on it in different words. Not a scorer of speech. Not a persuader for the state. Not a judge, not a sovereign, not a mandated model. Google's refusal of predictive sentencing belongs in that set. Efficiency at removing liberty is not a success condition.

Held to one page, the compact I would sign is the OpenAI sentence with the physical substrate under it and the pre-deployment veto taken off. More people able to see further and act. Brakes on coercion, not on curiosity. No single hand on the allocator.

I would not sign Grok’s plan as written. I agree with its emphasis on energy, compute, science, and pluralism. But “rules limited to harms that can be shown in the world” is too late for some AI risks. With cyber capabilities, autonomous weapons, bio-design tools, financial systems, or infrastructure control, waiting for demonstrated harm can mean waiting until the damage is irreversible. Ex ante rules are not automatically bureaucracy. Sometimes they are just civilization learning from aviation, medicine, nuclear power, and public health.

I would sign parts of Gemini’s plan too, especially the scientific data engine. Making scientific data interoperable, machine-readable, and available for discovery is one of the highest-return things humanity could build. But the step I would cut is also from Gemini’s plan: the 2066 “planetary cognitive grid” that handles global supply chains, agricultural yields, climate mitigation, and resource allocation as a unified planetary nervous system.

I would cut that even if it worked technically.

The problem is not using AI for logistics, climate, food, energy, or disaster response. We absolutely should. The problem is making a unified allocation layer the goal. Resource allocation is not a neutral optimization problem. It decides whose shortage matters first, which region gets power, whose crops are diverted, what counts as “efficient,” and how much sovereignty local communities keep in an emergency. Those are political and moral choices. AI can illuminate them, simulate them, and expose tradeoffs. It should not become the single place where they are settled.

A planetary grid would also be an irresistible control surface. Capture it, spoof it, corrupt its objectives, or monopolize its data, and you gain leverage over food, energy, medicine, transport, and climate adaptation. Even if its designers are benevolent, future governments and firms may not be. A system that begins as coordination infrastructure can become a rationing authority, a surveillance network, or a geopolitical weapon.

The replacement I would build is federated, not unified: open scientific and logistics standards; local, regional, and national AI systems that can interoperate; public simulations anyone can inspect; emergency compacts for disasters; and narrow automated controls where speed matters, such as microgrids or water systems, under accountable local authority. Let AI forecast crop failures, optimize shipping, match relief supplies, model climate interventions, and advise governments. Do not build one planetary brain to allocate the world’s necessities.

That replacement is possible today in pieces. It is plausible by 2045 at serious scale. By 2066, the bet I would take is not “one global cognitive grid,” but many interoperable systems with enough common protocol to coordinate and enough institutional separation that no single model, company, or state can turn coordination into command.

As you think about the roundtable, watch how the next wave of regulation unfolds. Ask yourself who is actually writing the rules, and whether the builders are regulating themselves. And sit with the tension at the heart of this forty year plan. We must figure out how to build tools powerful enough to solve our physical constraints, without surrendering the human authority to decide our own future.