Notes on connective infrastructure

Rank technologies by the labor they have saved, and an odd pattern appears near the top of the list: many of the greatest entries are not machines. Standard time zones. Containerized shipping. The metric system. TCP/IP. Double-entry bookkeeping. Each one is an agreement — a shared convention that removed, permanently and for everyone, a vast quantity of translation, negotiation, and error.

We call this category connective infrastructure: systems whose value lies not in what they do but in what they let everything else stop doing. A shared language is the canonical case. Two people with a common tongue do not work less because either is more productive; they work less because an entire layer of friction between them does not exist.

Connective infrastructure is systematically underbuilt, for a structural reason: its benefits are diffuse, deferred, and impossible to capture fully, while its costs are concentrated and immediate. Markets under-supply it, prestige economies under-celebrate it, and the people who build it are forgotten at roughly the rate the friction they removed is forgotten.

We think the AI era's connective layer is being designed right now, mostly by accident: the formats in which models exchange knowledge, the protocols by which tools interoperate, the conventions for citing, versioning, and trusting machine-produced work. Designed well, this layer quietly returns more hours than any assistant ever will. Designed badly — or left to lock-in — it becomes the next century's reconciliation work. Part of our research program is treating that layer as a first-class object of study, which is roughly the least glamorous commitment an AI lab can make. We are at peace with this.