A few hidden dials

Sixteen personality types, twelve zodiac signs, five factors, four temperaments — humanity keeps trying to compress the infinite variety of people into a handful of coordinates, and keeps half-succeeding. The half that succeeds is uncanny: you meet someone for an hour and already you are placing them, sensing that this new person rhymes with someone you knew years ago, as if both were versions of one underlying thing turned to slightly different settings.

The mathematics of embeddings makes the intuition precise. High-dimensional, various surfaces — faces, words, personalities — often lie near a low-dimensional manifold, and finding its few axes lets you place, compare, and even interpolate between things that seemed incomparable. The danger and the power are the same: a few dials really do capture a lot — and never everything. Below is an honest, miniature version of the machine, running real arithmetic on a small world you can turn in your hand.

Thirty characters, each with ten visible traits. They look wildly different — but every trait was secretly written by turning just two hidden dials and adding noise. The instrument doesn't know that. It runs real principal component analysis on the raw traits and tries to recover the dials from the surface alone. Hover a point to see the creature it stands for; near points share a look.

Ten traits give each creature a thousand ways to differ. Yet the differences are not free: raise the first dial and size, spikiness, and saturation move together, because one hidden cause writes them all. The visible variety is real but correlated — and correlated variety lives on a thin sheet inside the big trait space, a manifold with far fewer directions than it first appears. To embed is to discover that sheet: to find the handful of directions along which the data actually moves, and throw away the rest as so much redundant echo.

Once you have the sheet, nearness becomes a real, computable thing. "These two are alike" stops being a vibe and becomes a distance — and, remarkably, a distance you can do arithmetic in. That is the quiet revolution behind every modern recommendation, every face-unlock, every king − man + woman ≈ queen: meaning laid out in coordinates, where similarity is proximity and analogy is a straight line.

• Interpolate between two characters and meet the ones in between. Click two distant points, then drag t. The decoded creature at t = 0.5 was never in the data — yet it is coherent, a plausible cousin of both. That is the "vector arithmetic of meaning," done honestly on a toy.
• Drop k to 2 and see how much survives. Pull the compression dial down to a single dial, then two. Compare the reconstruction to the real one. Feel how quickly it becomes almost right — and exactly where it stays wrong.
• Find what the axes mean. Read the two recovered dials at the right: PCA hands you directions, not names. The labels ("mostly size + spikes") are read off the loadings — the same interpretive leap a typologist makes when they christen an axis Boldness or Warmth.
• Switch skins. The same machine re-runs on personalities, cuisines, writing styles. Different surfaces, same discovery: a low-dimensional order hiding under the variety.

Every archetype system — the humors, the Enneagram, the Big Five, the star sign your aunt trusts — is a lossy embedding. And the scree curve is why they are not simply nonsense: few dimensions really do capture a great deal. A good typology is a genuine compression of character, letting you carry a stranger's rough shape in a single word and be right more often than chance. This is its wisdom, and it is not small.

But the same curve that flatters the type indicts it. Every reconstruction here leaves a residual, and in the human case the residual is a person. The violence of types is the confident act of discarding it — hearing "Scorpio," "Type 4," "high neuroticism" and treating the remainder as noise rather than as the specific, unrepeatable someone in front of you. The ethics of the compression lives in one distinction: a type held as a tool is a first guess you expect to revise; a type held as a cage is a verdict you have stopped listening past.

Look again at the gap above the scree curve. That leftover variance is precisely the information a two-dial model was built to ignore — and in a person, the thing a two-dial model ignores is their individuality, the untypable remainder that makes them not an instance of a category but themselves. A good typology knows its k and publishes its error bar: it tells you how much it is throwing away, and holds its verdict loosely in proportion.

A bad one mistakes the map for the manifold. It reifies the sign, the type, the score — treats the coordinate as the cause, forgets that the axes were named by a human reading loadings, and lets a compression that was only ever most of you stand in for all of you. The instrument keeps you honest by always drawing both: the real one and the reconstruction, side by side, with the error between them printed in plain numerals.