The founding compact between Washington and the frontier labs is being drawn up, and no one has claimed the role of architect. The labs have more say in those terms than the current mood suggests. The government’s instruments of control reach the chips, the contracts, and the terms of access, but not the part of the system that sets the work’s direction. A lab can accept the constraints, fight them, or have a hand in building the arrangement itself, and the historical cases show what building takes: four moves, each depending on the one before it. The window for the first is the narrowest, because that move takes years to build and cannot be begun once relations have soured. It is open now.

The first two essays argued that the relationship between governments and the institutions building frontier AI does not fit inherited categories, and that the search for the right historical technology analogy is the wrong search. Those essays diagnosed. This one is for the side that still has room to act: the frontier lab.

The instruments meant to bring the labs into line are growing in number and force. Export restrictions, procurement conditions, security designations, limits on who may use a model: all real, and all multiplying as frustration rises on both sides. These instruments reach the chips, the contracts, and the terms of access. They do not reach the part of the system that sets the direction of the work. They press on the surface. The direction of the frontier is set below it. That gap is the lab’s opening, and it will not stay open.

A lab facing these constraints has three broad responses. It can accept them. It can fight them. Or it can build something the first two never produce: an arrangement that ties its work to public purpose on terms the lab still helps set.

The three are not equal. The difference is time. Accepting costs little now and gives up any say later, once the terms harden. Fighting costs more every quarter, because the instruments grow more numerous and harder to move the longer the fight runs. The ground shifts under the lab even when its case is good. Building costs the most now. It is also the only response that leaves the terms partly in the lab’s hands rather than wholly in the government’s.

Some will argue for a fourth course: offer restraint, collect reciprocity, the good-actor bet. It fails because the state has no single mind to bargain with. It has factions and arguments. Somewhere officials are defending the working relationship; somewhere others are pressing for restriction and control. Good conduct gives the defenders material, but it settles nothing, and the incident that arms the other side can come from any lab in the field. A lab that behaves well and waits has not found a fourth option. It is accepting, with better manners.

Accepting and fighting are reactions to a relationship someone else is defining. Building defines it. That means knowing what a workable arrangement is made of before the moment to make it is gone. The cases where private institutions held frontier capability the state could not control, and the relationship held anyway, show what it is made of.

Four moves recur across those cases. None is a rule. Each is something a lab can do. They come in order because each depends on the one before it. The first is the slowest to build, which is what matters most now.

The first is an anchor: one central public relationship the lab keeps by choice rather than need. Armstrong, the private firm whose Elswick works on the Tyne built naval guns and warships for the Royal Navy and for foreign buyers alike, kept the Admiralty as its principal customer while selling widely. The early semiconductor firms stayed close to defense and space demand even as commercial markets opened under them. In each, the public tie was primary by choice, not dependence. Held long enough, an anchor like that earns a place among the architects of the founding compact. It gives the lab a hand in the terms while they are still being set, before they are written down.

Bell Labs had that place through the regulated settlement it worked inside and helped shape. It cannot be bought in the moment it is wanted. The AI labs’ institutes and foundations matter, but they play a supporting role. They build the scaffolding around the compact: convenings, research programs, evaluation practices, policy language, and trusted channels. They can make the room easier to enter. They cannot, on their own, create the seat. That comes from the lab’s sustained public relationship and from the judgment the lab is willing to bring into it. From that seat, a lab can offer what the government cannot get any other way.

A frontier lab sees what the state only learns later: where the capability is heading, what AI can and cannot yet do, which risks are real and which are noise. This is judgment, and it is the part of the system the instruments do not reach. A lab can offer it directly, telling the government what it sees instead of waiting to be asked. From an architect of the compact, that opinion is taken seriously. The same words from an outsider are not.

A lab can also widen the overlap on purpose. The area where the lab’s interest and public purpose coincide is usually treated as a fact to be found. It can be treated as ground to expand. A lab close enough to understand public purpose can make choices that enlarge the shared area instead of leaving its size to chance. Meiji Japan did this from the state’s side, building the channels that let it absorb what it did not yet have. A lab does the same from the other side, building the overlap rather than waiting to learn how wide it happens to be.

Widening the overlap is what makes the last move credible: declaring the edge. This means being clear about where the overlap ends: where the lab’s interest and public purpose part ways. It is the one move of the four that costs the lab anything.

After the overlap has been widened, a lab that has shown where it aligns and is honest about where it does not is believed about both. Armstrong is the cautionary case. Its interests and Britain’s were never identical, and Armstrong never said so. It sold widely and quietly, leaving the limits of alignment unspoken. No one could say where those limits were until they were tested, and by then the settlement was done by the state, not by the firm. Candor about the edge is what keeps the widened overlap from looking like a sales pitch.

The four moves take a lab to the edge of the known cases. They do not solve the harder problem: at some point a lab’s commercial or research interest and public purpose pull hard in opposite directions and genuinely diverge. Armstrong shows one boundary. The firm operated for decades inside a zone of useful overlap with British power, but World War I exposed the limit. When the risk became intolerable, the state set influence aside and took direct control of armament production. The Manhattan Project is the opposite boundary: complete alignment achieved by pulling the work inside the state under wartime command. Between influence and control, the historical cases leave the central problem unresolved.  

That empty middle is the unbuilt arrangement: something that holds when the interests behind it part. History does not hand it over ready-made. A lab that makes the four moves well arrives exactly at its edge, which is as far as any of the cases go.

The catch is in the order. The first move is the slowest, and it decays fastest once a relationship sours. A place among the architects is built slowly, by keeping a relationship central over time, and it cannot be summoned the moment the constraints arrive.

As the instruments of government control multiply and the frustration rises, the room to build that place shrinks. Once it closes, the later moves lose force. A lab that waits until the constraints have hardened finds the first step already closed and is left with the responses that hand the terms to someone else.

The arrangement that ties frontier capability to public purpose, without command on one side or capture on the other, will not come from history ready-made.

It will come from the lab’s side, if it comes at all, built by a lab that did the slow work of becoming an architect rather than waiting for the state to hand down the terms.

That is what it means to be present at the creation.

— Andrea