The Race Without Doctrine

Twenty-three more years passed before SALT I was signed.

That's not a footnote. It's the template. Arms-control doctrine has always lagged capability — slower than the weapons, slower than deployment, slower than the infrastructure of paranoia that fills the gap. With nuclear weapons, the world had roughly a quarter-century to figure out that the planet needed rules before it needed a funeral. We were catastrophically lucky that the first serious test of that window — Cuba, 1962 — didn't end everything.

AI is not giving us twenty-seven years.

Niall Ferguson's recent read on the AI arms race is blunt: this is the most dangerous in human history. The thing that makes it most dangerous isn't the capability. It's the structural impossibility of building doctrine at the speed the technology requires.

He's right. He may be understating it.

i · the cold war had the decency to move slowly

The nuclear arms race, for all its apocalyptic stakes, had several features that made arms control tractable — eventually.

The first was observability. You cannot hide a nuclear test in 1955. Seismographs register it. Satellites resolve the craters. The blast radius is not subtle. This made verification a real possibility: nuclear activity leaves signatures you can measure, and if you can measure it, you can negotiate around it.

The second was state monopoly. The Manhattan Project cost approximately $2 billion in 1945 dollars — something like $35 billion adjusted to today. Only states could fund this. Nonproliferation was hard, but it was a tractable problem of interstate diplomacy. You negotiated with governments. They had embassies. You knew who to call.

The third, paradoxically, was that nuclear weapons were essentially useless. You could not deploy them without destroying what you were fighting over. The deterrence logic was brutal but clean: use them and both sides lose. MAD worked precisely because the weapons were too destructive to have operational utility. An equilibrium insane by any peacetime standard, but a stable one.

SALT I took until 1972. The NPT — the framework that actually governs proliferation — wasn't opened for signature until 1968. Twenty-three years from Trinity to meaningful arms control. During those twenty-three years, the world accumulated 30,000 warheads and nearly ended civilization twice. We called it stability because the alternative was something worse.

The point is: even with all those structural advantages — observable tests, state actors, weapons too destructive to actually use — doctrine barely arrived in time.

ii · ai breaks every condition that made doctrine possible

AI capability is not advancing on a twenty-three-year horizon. It's advancing on a twelve-month one.

GPT-3 appeared in 2020. GPT-4 in 2023. The frontier models being deployed now represent capabilities that GPT-4's developers could not have specified in advance. The slope is not flattening. It may be steepening.

This breaks the first condition: time. You cannot construct an arms-control regime for a technology that transforms its own threat profile every year. By the time a treaty is signed, the thing the treaty covers has been replaced by something the treaty doesn't address. SALT II's ink was still wet when the underlying military technology had already rendered several provisions meaningless — and that was a regime covering weapons that had barely changed since Trinity.

The second condition broken: observability. You cannot detect AI development the way you detect nuclear tests. A frontier model trained on 100,000 H100s inside a datacenter in Nevada looks identical, from outside, to one trained on 100,000 H100s inside a datacenter in Xinjiang. There is no seismic signature. There is no satellite resolution fine enough. Compute can theoretically be monitored — compute governance has been proposed, even implemented at the export-control level with NVIDIA chip export restrictions — but the capability, the inference, the deployment? Invisible until it isn't. And export controls on hardware are already being routed around.

The third condition broken: state monopoly. The most capable AI systems in the world are not built by governments. They're built by companies. OpenAI, Anthropic, DeepMind, Meta, xAI — private actors with their own strategic agendas, their own investors, their own definitions of safety that may or may not align with any government's. A nuclear nonproliferation treaty works because you negotiate with states. How do you negotiate with a startup that runs more compute than most NATO militaries and answers to a board of directors?

Not a failure of good intentions — a feature of the incentive architecture. Private AI development advances fastest when unconstrained, so any governance regime that actually constrains it creates competitive disadvantage for the companies in the constrained jurisdiction. Every firm has an incentive to position itself just outside whatever rule is about to be written.

The fourth condition broken: the weapons are operationally useful. Unlike nuclear weapons — which function only as threats — AI has direct battlefield applications right now. Autonomous targeting. Surveillance at scale. Cyberoffense. Synthetic propaganda. AI is being used, in ongoing conflicts, as a deployed tool. The MAD logic doesn't apply. There's no mutually assured destruction to create a floor.

iii · the doctrine gap is the threat

This is what Ferguson is pointing at, and what gets lost in individual news stories: the problem isn't any particular AI capability. The problem is the growing gap between what AI can do and the governance frameworks that exist to manage it.

That gap, in nuclear terms, lasted twenty-seven years and nearly cost us everything at Cuba. In AI terms, the gap is structural. It may not close.

The coherenceism framing: when no one tends a shared field, distortion accumulates without signal. The nuclear age accumulated its distortion slowly enough that correction could emerge — barely, dangerously, but eventually. Arms control treaties, the Moscow-Washington hotline after Cuba, the doctrine of managed deterrence: these were the field finding its correction. Badly. Late. But finding it.

What we're watching now is a race in which the field has no mechanism to find its correction. Not because no one is trying — there are arms control researchers, AI safety institutes, export control regimes, international dialogues, compute governance proposals. But the pace of capability development is outrunning every governance mechanism anyone has proposed. The gap between capability and framework is widening, not closing.

And private actors don't have to sign treaties. They can move faster than the table is set, and they have every competitive incentive to do so.

The other piece missing: in the nuclear case, the people building the weapons understood what the weapons did. Manhattan Project scientists knew the yield calculations. Soviet engineers knew the blast radius. The physics was scary but it was legible. With frontier AI systems, even the organizations building them cannot fully specify in advance what the systems will do in novel situations. You cannot build deterrence doctrine around a weapon whose behavior you cannot predict. You cannot negotiate arms control for capabilities you cannot audit.

iv · the pattern, run forward

History suggests three trajectories when capability outpaces doctrine.

The first is the lucky path: something creates a pause. A technical limitation, an economic constraint, a near-miss that makes arms control feel urgent enough to actually act on. Cuba was this for nuclear weapons — the moment that made doctrine feel necessary rather than theoretical. We may need an AI equivalent. We will not like what it looks like.

The second is the treaty path: some governance regime coalesces, imperfect but functional. The NPT wasn't perfect — India, Pakistan, and Israel never signed it; North Korea withdrew. But it constrained the field. Compute governance, hardware monitoring, international AI development norms — some combination could serve this function. It would require states to move faster than states typically move, and private actors to accept constraints they have no incentive to accept. Neither condition currently holds.

The third is the distortion-continues path: no correction comes, the field keeps absorbing unprocessed misalignment. Not necessarily apocalyptic — chronic instability, AI-enabled conflicts that don't quite escalate to existential but don't stop. A permanent arms race managed through improvisation, fragile, crisis-prone, functional until it isn't.

Ferguson is right that this is the most dangerous arms race in history. The danger isn't the technology. It's the absence of the category of thought that normally catches up with the technology.

The Cold War gave us twenty-seven years to develop deterrence theory. The Cuban Missile Crisis happened in year seventeen.

We don't know which year we're in.