AGI Evolution and what we can do about it: a Danish perspective

Where it actually goes, and what to attend to.

The honest version of this question, given what we've said in the last two pieces, isn't what should governance do. It's where does the trajectory actually get shaped, and where does engagement compound. Those are different questions, and they produce different answers from the ones currently on offer in the policy literature.

The trajectory gets shaped, over the next forty years, by a small number of identifiable forces. The frontier labs — three to five American, two to three Chinese, with one or two European or Indian dark horses possible but not probable — set the capability frontier. The security and intelligence services of the US and China set the deployment frontier in the domains that matter most strategically. Compute infrastructure — the fabs, the power, the cooling, the rare-earth supply chains — sets the physical envelope. Capital markets set the time pressure. The technical staff of the labs set the design choices within all of that, and those choices compound across model generations in ways that aren't easily reversed once locked in.

Everything else is downstream. National regulators, international bodies, civil-society organisations, academic ethicists, journalism — these influence the rhetorical environment in which the upstream decisions are made, sometimes consequentially, mostly not. Treating them as the locus of agency is what produces the comfort literature I was arguing against in the last piece. Treating them as part of the rhetorical environment is more accurate and more useful, because it lets you see them clearly and engage them where they actually matter, which is in specific local cases rather than as a general governance program.

Let me work through this in layers: where the trajectory actually goes, what constraints will actually bind, what governance is worth fighting for despite the limits, and what an individual sitting in Denmark in 2026 should actually be attending to. The last layer is the one the question really wants. The first three are necessary to make the last one honest.

Where the trajectory goes

Capability won't progress smoothly. It'll progress in punctuated steps, with plateau periods between them, and the plateaus will be misread as ceilings until the next step happens. The current discourse oscillates between AGI is imminent and it's just statistics on a quarterly cycle. This will continue. Neither pole is right. The actual pattern over forty years will look like: a major capability jump every three to five years, with deployment, infrastructure adaptation, and second-order effects working through the system in the intervals. The jumps that matter most won't be the ones currently anticipated. The ones currently anticipated — embodied reasoning, persistent memory, true multi-modal grounding — will arrive and be absorbed. The ones that aren't anticipated — recursive self-improvement at the engineering level, novel scientific discovery at frontier physics or biology, autonomous coordination among agent systems — will arrive in some form and reshape what AGI means in ways the current frame doesn't predict well.

The deployment surface expands faster than the capability surface. By 2035, AGI-grade systems are integrated into most white-collar workflows in the developed world, with the integration deepening rather than reversing. By 2045, the same is true of most logistics, manufacturing, healthcare delivery, and significant portions of education. By 2055, the integration reaches the parts of life that currently feel insulated — personal relationships, child-rearing, civic participation, religious practice — not by replacing them but by mediating them. By 2066 there's almost no domain of life in the developed world that doesn't pass through AGI-grade systems at some point in its operation. The integration is the consequential development, not the capability jump. Capability is the precondition. Integration is what changes society.

The geopolitical pattern stabilises into a bipolar capability race with a third tier of dependent states. The US and China remain the only two states with full-stack frontier capability through the 2040s. India, possibly Israel, possibly a European consortium, develop second-tier capability: capable of running frontier models, capable of fine-tuning at the application layer, dependent on the bipolar powers for the underlying foundation models or the chips to train them. The third tier is everyone else, consuming what the first tier offers, with limited ability to shape what gets offered. This is the world AGI gets governed in, and any honest governance frame has to start from it rather than from the assumption that a universal regime can be negotiated among equals.

What constraints will actually bind

The constraints that bind are the ones that operate on physical or institutional realities rather than on stated norms.

Compute is the first. Frontier model training requires concentrated compute that requires specific chips made in a small number of fabs powered by significant fractions of regional electricity grids. This is the one part of the AGI stack that's physical, verifiable, and concentrated enough that state action against it can work. The US chip-export controls on China are imperfect — the Chinese have routed around them, developed domestic alternatives, and gained from algorithmic efficiency — but they've also slowed Chinese frontier capability by an estimated 12-24 months. That's the largest documented effect of any AGI-governance action by any state, and it operates entirely through physical constraint on compute rather than through normative constraint on use. The constraints that work in the future will look like this. Constraints that operate on stated norms — don't use AI for X — will mostly not bind. Constraints that operate on hardware, energy, supply chains, or financial flows will sometimes bind.

The second constraint is technical alignment. This isn't governance in the conventional sense — it's engineering — but it's the closest thing we have to a binding constraint on what frontier systems can do. If interpretability research advances to the point that we can reliably detect deceptive behaviour, manipulation, or value drift in frontier models, then deployment without that detection becomes commercially and politically untenable, because the alternative is uninsurable risk. If interpretability doesn't advance, deployment proceeds anyway, with the risk priced in or hidden. The marginal dollar spent on technical alignment work probably does more to shape the trajectory than the marginal dollar spent on international coordination. This is the unfashionable answer and probably the right one.

The third constraint is lab culture. The three to five frontier labs that determine what AGI becomes have internal cultures, hiring practices, incentive structures, and willingness to delay deployment for safety reasons that vary significantly between them. The difference between Anthropic's culture and xAI's culture is more consequential for what 2066 looks like than the difference between any two national regulatory regimes. This is also unfashionable to say. It's true anyway. The culture inside the labs is shaped by the people who go work there, the people who lead them, and the competitive pressures they operate under. All three are influenceable. None is influenced by the governance frame as currently practised.

The fourth constraint is incident reporting and capability evaluation infrastructure. This is the part of governance that does work, because it's continuous with the engineering rather than imposed from outside. If a serious AGI evaluation regime exists — public benchmarks, mandatory pre-deployment evaluations, post-deployment incident reporting, with real consequences for material misrepresentation — then frontier development happens inside a feedback loop that catches failures faster. If it doesn't exist, failures accumulate until something visible happens. This is doable. It's already partially in place. It deserves more attention than most of what's currently called AGI governance.

The fifth constraint is the financial system's risk pricing. Insurance markets, capital allocation, and credit ratings respond to AGI risk in ways that bind without requiring legislation. If a frontier model causes a major incident — a financial system failure, a critical infrastructure compromise, a verified weapons proliferation event — the insurance and capital markets respond on a timescale measured in weeks. This is faster than any regulator. It's also less articulate, more reactive, and prone to overcorrection. But it binds.

What governance is worth fighting for despite the limits

Most governance proposals are theatre. A small number are worth fighting for because they're achievable and consequential. Here's what I think actually matters.

No fully autonomous lethal force in domestic civilian contexts. The military version will normalise; that fight is lost, and arguing it as if it weren't is a waste of energy. The civilian version is distinct. Police, immigration enforcement, prison management, border control, public-order maintenance. These domains aren't subject to the same competitive escalation as inter-state military capability, which means the line might still hold. The right standard is that lethal force in civilian contexts requires a named human decision-maker with full legal liability, with no exceptions for high-throughput scenarios. This is winnable in democratic states if the fight starts now. It's not winnable in 2040 if it starts in 2040.

Cryptographic identity infrastructure for authenticated communication. The information ecosystem won't be saved at the platform level. It might be partially saved at the personal level if widely-deployed cryptographic identity becomes standard for communication people care about. Estonia has done this for civic communication; the broader application to personal communication, journalism, and political speech is technically feasible and institutionally underdeveloped. Worth investing in because the alternative — an information environment in which authentication is impossible — is corrosive in ways that compound. The Nordics could lead this if they chose to. They probably won't, but they could.

Compute registration and transparency above defined thresholds. This is the governance proposal most likely to actually bind, because compute is physical and concentrated. Registration of training runs above a defined compute threshold, with reporting requirements that scale with capability, is achievable through chip-level instrumentation that's already being prototyped. This wouldn't constrain what people do with AGI. It would create visibility into what's being built and by whom, which is the precondition for any other constraint that might later become important. It's also one of the few areas where US-China coordination is at least conceivable, because both states have an interest in knowing what the other is building.

Narrow non-delegation in specific high-stakes civilian contexts. The general non-delegation principle is sentimental and will lose. A narrow version — criminal sentencing, child custody, end-of-life medical decisions, asylum determinations — might hold if defended as a specific list rather than a general principle. The list has to be narrow enough that the political cost of insisting on human discretion is bearable. Five or six categories, with strong procedural protections, defended hard. Broader than that loses everything. This is unromantic and feels like surrender. It's also the version that might actually preserve something.

Educational reconstruction. This isn't a constraint, it's a project, but it's the most consequential civilian work of the next twenty years and almost no one is doing it seriously. The current education system in the developed world prepares children for a world that won't exist by the time they're forty. Reconstructing it for the world that will exist — high automation, abundant cognitive capability, scarce judgment and taste and motivation and the ability to sustain meaning in conditions where productive labour is no longer the central organising fact of life — is the work. Curriculum reform, assessment reform, teacher training, the relationship between school and the rest of childhood, the role of physical and embodied skill, the cultivation of attention. This is a thirty-year project that needs to start now and isn't starting at the scale required. The Nordics are better positioned for this than anyone else and aren't doing it.

That's the list. Five things, four of them governance, one of them institutional reform. Everything else conventionally proposed is either downstream of these or theatre. I'd rather spend energy on five winnable fights than on a comprehensive program that loses every fight.

What an individual should actually be attending to now

This is the part of the question that matters most, and the part the governance frame can't address because it operates at the wrong level.

Your own composition with the technology. You're using AGI-grade systems now. The question is whether you're using them as a multiplier on your own capability or as a substitute for it. The first compounds. The second atrophies. The discipline is using these systems for the parts of work you understand well enough to evaluate the output critically, and not using them for the parts where you can't tell if the output is right. Most people are doing the opposite. The asymmetric advantage over the next twenty years goes to people who use AI heavily in their domain of competence and lightly outside it. The disadvantage compounds for people who use it heavily everywhere.

Your children and grandchildren. The most consequential decision a parent makes in the next twenty years is what relationship their child has with these systems during the developmental window. Not whether — that ship has sailed. The shape of the relationship. A child who learns to write with a model collaborator before they can write alone develops different cognitive habits than a child who learns to write alone first and adds the model later. A child who learns to ask the model before they learn to think loses the capacity to think before asking. The discipline is teaching the underlying capability first, then teaching collaboration with the model as a separate skill. This is harder than it sounds because the school system isn't structured to support it and the social pressure runs the other way. Worth fighting for in your own household even if you can't win it at the policy level.

Your professional positioning. In every domain, AGI deployment over the next twenty years will divide practitioners into three categories. The directors: people who shape what the systems do in their domain because they understand both the domain and the systems deeply enough to design the deployment. The directed: people whose work is mediated, supervised, or replaced by the systems. The displaced: people whose work no longer exists in a form that supports a livelihood. The category you end up in isn't decided by your current professional standing. It's decided by what you do in the next five to ten years to position yourself relative to the systems. Most people don't take this seriously because it feels distasteful, like building a personal brand. It's not the same thing. It's about whether your work involves the systems on terms you set or on terms set for you.

Your civic engagement. The civic battles that matter aren't the abstract ones. They're specific deployments in specific institutions. Your local council deciding whether to use AI in social services. Your child's school deciding what role models play in teaching. Your hospital's policy on AI-assisted diagnosis. Your country's policy on AI in court proceedings. These are all winnable or losable, and they cumulate into the institutional environment your children will live in. The macro governance fight is mostly noise. The micro deployment fights are where the actual character of the integration is decided.

Your epistemic hygiene. The information environment will degrade further. The question isn't how to fix it at the platform level — that fight is mostly lost — but how to maintain a functional epistemic life inside a degraded environment. Practices that matter: a small number of trusted sources, sustained relationships with people whose judgment you can evaluate over time, primary sources where you can access them, willingness to be wrong publicly, willingness to change your mind in public when warranted, awareness of when you're being persuaded by something whose authorship you can't verify. These are practices that worked before AI and work better in the AI environment if held more strictly. They don't scale. They're individual or small-group disciplines. They're what survives the degradation.

Your community. The transition will be lonely for people who navigate it alone. The people who do well over the next forty years will be people embedded in real communities of mutual obligation: geographic, professional, religious, familial, whatever the specific form. Not networks. Communities, in the older sense. The kind of relationship where people show up when something is wrong and aren't keeping score. These are getting harder to build and more valuable as they get harder. The investment in them now compounds in ways that won't be visible for ten years and won't be replaceable when needed. This is the single most important practical investment for navigating what's coming, and it's the one the entire incentive structure of contemporary life runs against.

The Nordic question

I am Danish, sitting in Australia, so the question lands twice. From Denmark, what's the right move? From the Anglosphere, what does the Nordic position offer that the Anglosphere lacks?

The Nordics have advantages that they aren't using. High state capacity. High trust. Small enough to coordinate internally without the EU-scale dysfunction. Large enough as a bloc to matter. Cultural disposition toward pragmatic accommodation rather than moral panic. The world's most successful welfare states, which gives them both the most to lose and the most institutional capacity to adapt.

What they could do that they aren't: lead on cryptographic identity infrastructure. Lead on educational reconstruction. Lead on compute transparency as a regional bloc, ahead of the EU process which will be too slow. Take the demographic-collapse problem seriously enough to actually plan for a population in 2066 that's 30-40% smaller than today's, with AGI-driven productivity making the smaller population sustainable but only if institutions reform fast enough. Build a Nordic frontier-model consortium — not at the American or Chinese scale, but at the second tier — to avoid total dependency on the bipolar powers for the systems that mediate civic life.

None of this is happening seriously. The Nordic governments are running standard EU-style governance theatre. The opportunity cost is large and growing.

From the Anglosphere, the Nordic position offers a worked example of high-functioning small-state pragmatism that the larger democracies could learn from, if they wanted to. They mostly don't. The lesson worth taking is that scale isn't the only variable. A coordinated five-million-person polity can do things a fragmented three-hundred-million-person polity can't, and the things it can do include some of the constraints I listed above. The Nordic countries are the most likely places for the good versions of the next forty years to first be tried, even if they end up being tried badly.

The throughline

What I highlight across the three essay pieces — the romance one, the governance critique, this one — is that the structure of the answer is consistent. Powerful technologies produce social transformations that institutional responses can't keep up with. The honest move is to look at where leverage actually exists rather than where the discourse says it should exist. In intimacy, leverage is at the individual level: refusing the substitution, choosing the harder human relationship, building the communities that protect against the substitution at scale. In AGI governance, leverage is at the technical and institutional level: the labs, the compute, the alignment work, a narrow set of governance constraints that might actually bind. In what to do now, leverage is at the personal-positioning level and in five specific governance fights worth having.

The dominant discourse on all three subjects is at the wrong level. It treats consequential transformations as something we collectively decide about, when the actual dynamic is that consequential transformations happen and we decide about them retroactively. The right disposition isn't fatalism — the decisions you make individually and the institutions you support locally do compound — but it isn't the comfort fiction of collective deliberation either. It's working hard at the layer where work actually does something, and not spending energy at the layers where it doesn't.

The thing I'd want a Danish reader to take from these three pieces, sitting in 2026 looking forty years out, is that the work to do now is unglamorous and personal. Build the relationships that survive the substitution. Position your work and your children's education for the world that's coming rather than the world that's leaving. Pick the small number of public fights worth having and fight them hard at the local level. Don't waste energy on the abstract governance debates that won't bind. Hold the line on the few specific things where the line might still hold.

The civilisation that emerges from this isn't the one currently being designed at conferences. It's the one being built right now by the people who are clear-eyed about which layer they're working at.

Whether the Nordics can be one of those places is, I think, the most interesting open question of the next forty years. The answer depends on choices made before 2035. Most of them aren't being made.

For a longer response, send a reply →

← All essays