Can We Coordinate As Fast As We Invent?
Technology grows exponentially. Coordination does not. What happens when the gap gets large enough matters more than most things people are worried about. An argument using the Montreal Protocol as the exception that proves the rule, and asking what it would take for AI, biosecurity, and the other frontier-technology governance problems to look like Montreal rather than Paris.
Technology growth is exponential. Coordination growth is not. What happens when the gap gets large enough matters more than most things people are worried about.
The interesting question about long-run technological civilisation is not whether it can produce enough intelligence to do dangerous things — it obviously can and does — but whether it can produce enough coordination to keep up with what its own intelligence produces. That gap has a shape. It has a history. It is measurable. And it may be the more useful lens through which to read the next fifty years than any of the framings that get more attention.
The two curves
Technology grows compound-multiplicatively. It builds on itself. Better tools produce better tools; the physics you learn this decade lets you do the biology you couldn’t attempt last decade; the compute you deploy this year makes the compute you deploy next year cheaper. Not every technology grows at the same rate, but the aggregate frontier moves forward monotonically and, over long enough periods, at rates that turn out to be roughly exponential in whatever the interesting variable happens to be — transistor density, sequencing cost, model capability, energy per useful watt.
Coordination does not grow like this. Coordination is the ability of a species to agree on what not to do — not to burn the atmosphere, not to release the pathogen, not to launch the missile, not to deploy the model without safeguards. The technologies for coordination are treaties, institutions, laws, cultural norms, and enforcement mechanisms. These do compound in some sense — a modern nation-state can coordinate at scales the Bronze Age could not — but the rate is slow. The Congress of Vienna in 1815 took months of negotiation to settle a redrawing of Europe. The Paris Agreement in 2015 took twenty years of accumulated diplomatic work to produce a document whose implementation is voluntary. The average time for a major international treaty to move from first proposal to reasonable adoption is on the order of decades. The technology curve moves inside that time. Sometimes it moves several times inside it.
So the interesting quantity is not the level of either curve. It is the divergence. And the divergence has, for the last century, been widening at every measurable point.
What we can learn from the treaties that worked
Rather than argue this in the abstract, look at the case where coordination did keep up with technology within a single generation: the Montreal Protocol of 1987. The problem was chlorofluorocarbons destroying the ozone layer. The response was a treaty that phased them out, that has been near-universally adopted, and that is now credited with preventing hundreds of millions of skin-cancer cases. As international coordination victories go, it is the closest thing there is to a clean win.
Scott Barrett’s Environment and Statecraft identifies five factors that made Montreal work: the treaty defined and created an aggregate good, it distributed benefits equitably, it deterred non-participation, it deterred non-compliance, and it deterred free-riding. The Kyoto Protocol, negotiated less than a decade later on a superficially similar problem, had none of them. The Paris Agreement, negotiated eighteen years after that, replaced the missing factors with voluntary pledges and peer pressure.
But the Barrett factors are the mechanism, not the underlying cause. The deeper reason Montreal worked and Kyoto and Paris have not is simpler and more brutal: substitutes existed for CFCs, and they do not exist for fossil fuels. That is David Victor’s observation, and it is the honest reading of the comparison. When the industry that had opposed the treaty realised there were viable replacements and that the replacements would be worth manufacturing, opposition collapsed. In an alternative world where DuPont could not have produced HFCs and the refrigeration and aerosol industries had no technical path forward, Montreal would have failed for the same reasons Paris is failing.
This matters for the argument because it names what coordination can and cannot do. Coordination can move at technology-frontier speed when the technology it needs to coordinate around has already been solved technically. It can be extremely fast at getting everyone to switch to a thing that is ready to be switched to. It cannot invent the substitute. And for the interesting cases — climate, nuclear weapons, biotechnology risks, AI — the substitute is either not available, only partially available, or so heavily contested that the coordination cannot converge on which one to adopt.
The lesson from Montreal is not “treaties work.” The lesson is that treaties work when the underlying technology has already delivered the answer, and treaties merely enforce the switch. When the technology has not delivered the answer, coordination becomes a fight about who bears the cost of not-yet-existent alternatives, and that fight is slow, contested, and often ends in agreements that do not bind.
Why AI is the hardest version of this problem
Every previous instance of the coordination-versus-technology race had a physical bottleneck. Nuclear weapons required industrial-scale enrichment. Biological weapons required physical laboratories, cultured organisms, delivery mechanisms. Climate emissions required infrastructure that took decades to build and might take decades to replace. In each case, the coordination problem had time to catch up because the technology could not be deployed instantly by a small number of actors. The bottleneck bought the negotiators years.
AI collapses that bottleneck. The frontier models capable of substantial economic and civilisational impact are being trained by a handful of organisations, over months rather than decades. The compute infrastructure is physical, and to that extent coordination has a leverage point — but the compute is already installed and the training runs are already in motion, and each release makes the next release cheaper for whoever runs it next. Governance is not being asked to react to a technology whose deployment will unfold over decades. It is being asked to react to something whose deployment is compressing into years and, in the crucial layers, into months.
The current attempts to coordinate are, in Barrett’s framework, poor on almost every factor. There is no agreed definition of the aggregate good. There is no benefit distribution mechanism (the returns concentrate in a handful of companies and their host countries). There are no deterrence mechanisms for non-participation, non-compliance, or free-riding — the actors most likely to defect are the ones most likely to gain from defecting, and enforcement across national jurisdictions is essentially voluntary. The EU AI Act has emerged as the most ambitious attempt so far and is regulating what is on the market rather than what is being trained; the US framework relies on executive orders and voluntary commitments; China is following a different track entirely. Set against the pace of capability release, coordination is not visibly closing the gap. It is losing ground.
What does converge slowly, and what does not
It is worth distinguishing what coordination can be fast at from what it cannot. The literature on international regimes suggests three properties predict which coordination problems get solved quickly.
The problem has a clear scientific answer. Ozone depletion had one. The climate problem is scientifically settled at the level of the mechanism but contested at the level of what to do about it, which is not the same thing. AI safety at present has neither — the technical question of what alignment even means is contested, let alone the policy question of what to do.
The costs of action are less than the perceived costs of inaction. This was true for ozone once the substitutes appeared and the connection to skin cancer became visible in individual lifetimes. It has not been true for climate, where the costs of action are borne now and the costs of inaction fall on future decades and other countries. AI is in a similar position: the costs of restriction fall on the companies and countries doing the restricting; the costs of inaction are distributed across everyone.
There is a small enough number of relevant actors to make agreement tractable. Montreal effectively required agreement among a handful of chemical companies and about twenty producing countries. Nuclear non-proliferation required agreement among the P5. AI is currently coordinable at the level of five to ten frontier labs and three to five states, which is small; the number of actors capable of training a frontier model, however, is growing, and the moment it becomes a hundred rather than ten, the coordination problem becomes qualitatively different.
The technology curve reduces each of these three conditions. It makes the scientific picture more complicated because new capabilities create new hazard classes. It shifts the cost balance in ways that reward defection. And it multiplies the number of actors, because falling costs democratise what used to require nation-state resources.
What follows
The uncomfortable implication of all of this is that the standard model of coordination — treaties, institutions, agreements slowly negotiated between states over years — is unlikely to keep pace with technology in the categories that matter most. The Montreal Protocol has to be understood as the exception it is, not the model to be replicated. The Paris Agreement is closer to the median outcome for hard coordination problems.
This does not mean giving up. It means being honest about what coordination is likely to achieve and what it is not. Coordination can slow deployment; it can force disclosure; it can create legal liabilities that shape corporate behaviour at the margin; it can distribute the gains from technology more evenly than pure market outcomes would. What it is unlikely to do, on the historical evidence, is preempt the deployment of a technology whose deployment is already technically and economically feasible for a small number of actors, especially when the actors face different incentive structures and the technology itself is not yet stable enough to permit a Montreal-style clean substitution.
The civilisations that survive their own technology, if the Great Filter framing is a useful one, are the ones that either invent the substitute in time (so coordination has something to coordinate around) or invent forms of coordination that do not require the Montreal-style prerequisites (which nobody has yet demonstrated, though there are proposals). Neither of these is impossible. Both are hard. And the exponential curve does not slow down while we work out which one we are betting on.
The Great Filter, on this reading, is not intelligence itself. It is the gap between what intelligence can produce and what it can agree not to produce. That gap has been widening for a century. Whether it can be narrowed — and how — is the question that matters more than the specific technology of the moment. Every specific-technology governance debate is a small local instance of it. The debate about the specific technology matters less than getting the general answer right.
What would change my view
Three things, in declining weight.
One: a demonstration that a coordination institution has, in the last decade, successfully constrained the deployment of a frontier technology in a way comparable to what Montreal did for CFCs, without relying on the substitute-already-exists condition. If someone can point to a case where governance moved faster than the technology and it worked, the pessimism above needs revision. The current best candidate is the EU AI Act, but it is regulating rather than preempting, and its enforcement has not yet been tested against a determined non-complier.
Two: a serious quantitative measurement of the gap, rather than the qualitative argument above. The premise is that the divergence is real and widening. If someone did the work to measure it — treaty-response-time versus technology-doubling-time, across a large enough sample — the argument would be sharpened or falsified. I have not seen this measurement done well.
Three: a proposal for a coordination mechanism that does not depend on the Barrett factors and does not depend on substitutes. What that would look like is unclear to me. If someone has designed one, it is the piece of civilisational infrastructure most worth developing now.