Essay 001: ON WHAT THE PANOPTICON CANNOT SEE

We understand how a baker interacts with the world. He feels the weight of the humidity from an overnight thunderstorm. He smells the warmth of the crust as it cracks open. He engages in conversation with people as different as the bread, yet as important as the geopolitics of the region. The Arab makhbaz, the French boulangerie, or the Latin panaderia. The subtlety of life.

But our reliance on technology has structured our life into a panopticon state to which we remain oblivious. Some have bent the knee to the computational powers above, mistaking mathematical logic for reason and measurement for understanding. They allow data collection to proceed unchallenged, using the convenience of computational models to justify any desired action. Others have forgone their intellect, modeling themselves to be made legible within the system. An idolatry to an unknown faith has been born.

Yet, we fail to understand that the machines cannot grasp what any particular moment holds, let alone the unquantifiable collection of moments that govern our lives. They cannot see the baker as we see the baker. Though we cannot reject the machine that is upon us, we must discipline it.

Architectural decisions determine whether the machine can lie to you about what it sees. We build the skyscrapers within a city that allow but a small square of the blue sky to be seen, a narrow aperture through which the light looks untroubled. Though beyond the construct, dark clouds form.

The same for models. They should never be a black box. Fifty years ago, Doug McIlroy set down the rule for building things that compose: make each program do one thing well, make them work together, and give them a universal interface. Adapted for measurement, that convention becomes an ethic. A model must estimate one thing, expect its output to become the input of a consumer it will never meet, and honor a strict interface so that the parts compose into a whole that can still be held accountable.

Doing one thing well means one estimand per model. It is the foundation from which we can state what is being measured, what remains unmeasured, what is being observed by telemetry, and what lives entirely outside of the machine. A model must answer a single stated question such that the answer can be held accountable to empirical reality. An engine can estimate the temperature of an oven, the origin of the grain, or the hours of fermentation.

But the engine cannot say why khubz from a clay dome are not the baguettes from the four à pain, and neither are the tortillas de maíz off the horno de leña. Nor why the conversation within lingers on terrorism, revolution, and corruption.

The moment a model attempts to estimate two things at once, we lose the ability to say what it is wrong about. The moment a model attempts to estimate what it cannot observe, it stops measuring and starts asserting.

Expect the output of each model to feed a consumer we have not met yet. It is only possible because each model estimates one part, well and plainly. A quantity computed once feeds many consumers.

Life is lived in uncertainty. A single number cannot contain a decision. What are the implications of breaking bread with the terrorists, the revolutionaries, and the politicians?

Uncertainty lives within a distribution. UNIX's universal interface was the text stream; anything that reads and writes text composes with anything else. Our contract is narrower: every model must emit a distribution. A universal architecture turns a pile of models into a system.

We rebuild a model as a node within a graph, its outputs as the edges between models. Building afresh rather than complicating the old model.

The philosophy extends to the data itself. Every data source is a physical sensor with a finite aperture and a built-in blind spot. Each does one thing, observes one signal with one specific bias. A thermometer reads the temperature within the oven but not the hardness of the crust forming. A timer counts the fermentation but knows nothing of humidity impacting the dough. A scale weighs the flours but cannot hear the weight of the war outside.

An honest architecture does not force individual models to reconcile the bias on the fly. Instead, it composes them through a single observation layer that explicitly states the boundary between what is observed and what is not.

But data and small models are easy. The architecture is hard. The value is not the smallness; it is the pipe, and the pipe has to be universal and owned. An honest architecture is only honest if the party holding it is honest. Do we trust a state that counts its own dead?

Glue rots when no one keeps the contract honest. The discipline this philosophy demands lives in the architecture, not the models.

We do not design complex models. We must design a complex system whose simple parts tell the truth about their limits.

And because each quantity is estimated separately and kept as a distribution, the system reaches what we could never reach before. That the humidity and fermentation were never independent; the overnight storm was absorbed into the dough all along. We can measure the relationship.

We can trust what such a system tells us. The oven. The grain. The hours. We can trust it because it also tells us what it cannot see. The baker.