CONCEPT
Data as Artifact
Nyquist's most consequential epistemological gift to the AI age: the demonstration that data is not found in the world but manufactured from it—produced by sampling at a chosen resolution that irrevocably discards everything above it and permanently shapes everything built downstream.
There is a persistent illusion in the culture around artificial intelligence that data is raw, abundant, and waiting to be harvested—a natural resource that models simply consume.
Harry Nyquist's sampling theorem dismantles this illusion at its root. A dataset is not a transcript of reality. It is the output of a sampling process operating at a chosen resolution, following a filtering process that deliberately discarded everything above that resolution before sampling began. Two different sampling decisions yield two different datasets from the same underlying reality. A model can only ever learn what survived the sampling, and the survival was not passive: it was actively decided, usually without anyone noticing they were deciding it. The choice of what to capture, made at the moment of sampling, propagates through everything built on the data. The model's reality is bounded by a decision about resolution that someone made before the data existed.
Harry Collins reaches the same