Mathematical Thinking for AI, Lesson 2: Metric
Resemblance is an argument. Distance is a number. Every embedding you query rests on pretending they are the same, and on one axiom that bends whatever will not comply.
The picture
A metric is an agreement, and a brazen one: treat likeness as distance.
Likeness is a judgment. It is arguable, it shifts with context, two careful people can disagree about it forever and both be right. Ask whether a hospital is more like a hotel or more like a school and there is no answer, only a follow-up question: more like it for what purpose? For lodging, for scheduling, for staffing, for regulation, the ranking changes each time, and none of the rankings is wrong.
Distance is a number. It is cold, it is settled, it can be computed, and it has no follow-up question. The idea of the metric proposes that we behave as though the first were the second, and then get on with the work.
Accept that, and an enormous amount becomes possible at once, because a machine can compute with numbers and can do nothing at all with an informal judgment of resemblance. Refuse it, and the machinery that makes geometric retrieval practical begins to disappear: vector search, nearest-neighbor indexing, clustering, recommendation. Other routes exist, and they pay for the exemption in other ways. But almost everything you use that finds the most similar anything, at scale, is running on the pretense that likeness is distance.
Much rides on the word “agreement.” Nothing in the world hands you a distance between two sentences. Somebody builds one, by choosing what to encode and how to compare, and then everyone downstream treats the output as a measurement of the world rather than a consequence of those choices. The number arrives looking objective. It is a summary of decisions, wearing the costume of a fact.
In one sentence: a metric completes the conversion of a judgment into a number, and from that moment a machine can operate on it.
