In the rotations primer you turned one constellation by one angle and learned the essential thing: rotation moves everything and changes nothing. This page is the bridge from that single turntable to the machinery the essays actually use — the rotation groups SO(2), SO(3), SO(D). Two surprises live on that bridge. First: in three dimensions, the order in which you do two turns changes where you end up — a fact you can feel in your hands in the figure below, and the deepest single fact about rotations. Second: a rotation in 768 dimensions, which sounds unthinkable, is just a stack of ordinary turntables spinning independently. This primer assumes only primer 00 and primer 01.
In two dimensions a rotation is one number: the angle. Turn the constellation by 40°, then by 70° — or by 70° first, then 40°. Either way you have turned it by 110°, full stop. Angles just add, and addition doesn't care about order. Watch both orders run at once: the two panels take visibly different routes — mid-flight they disagree — and land in exactly the same place.
Now pick up a die — or use the two below. Two moves only: roll it forward a quarter turn, and spin it to the right a quarter turn. Do roll then spin on cube A. Do spin then roll on cube B. Same two moves, nothing skipped, nothing extra — and the cubes end up facing different ways. Not slightly different: measurably, unmistakably different. Click it through yourself; this is the kind of fact hands believe before heads do.
Why does the page behave and the die refuse? In 2D, both turns happen on the same turntable — one shared plane, so the angles pool. In 3D, "roll" and "spin" turn different, overlapping planes: each move tilts the stage the other was about to perform on. Whichever you do second acts on an already-moved object. Keep that reason in your pocket; it pays off twice below.
Collect all the rigid turns of an object — every way to rotate it without bending or mirror-flipping it. That collection has three homely properties. Do one turn, then another: the combined effect is itself some single turn (you could have gone there directly). Every turn can be undone by another turn. And "leave it alone" counts as a turn — the do-nothing turn. A collection of moves with those three properties is what mathematicians call a group. That's the whole word: the club of all the ways to turn something, closed under doing-one-then-another.
The names you meet in the essays are just this club at different sizes. SO(2): all rotations of the flat page — one angle each, order irrelevant. SO(3): all rotations of ordinary space — where you just watched order start to matter. SO(D): the same club in D dimensions. (The letters, once each: O for orthogonal — distances preserved, the primer-01 promise; S for special — no mirror flips allowed.) Nothing new happens to the idea as D grows. What changes is only how much room the club has.
So what could it possibly mean to "rotate" in 768 dimensions? Here is the fact that takes the mysticism out: every rotation in D dimensions is a stack of ordinary 2D plane rotations, each in its own plane, all spinning independently. Pair up directions — dimensions 1&2 form one plane, 3&4 another, and so on — and give each pair its own private turntable with its own private angle. A rotation in SO(768) is at most 384 turntables, each doing exactly what primer 01 taught. (This is not just a picture — it is literally how RoPE, the transformer's positional rotation, is built: one little turntable per pair of channels.)
Below: one cloud of twelve points in four dimensions, seen through two flat windows. The left window shows dimensions 1&2, where the points form a star; the right shows dimensions 3&4, where the same points form a spiral. Each slider spins one plane. Watch the other window while you slide: it cannot feel it.
Everything the essays do with rotations is assembled from what you just handled. After LayerNorm, a transformer's tokens live on the unit sphere SD−1 (the transformer primer) — and the moves that respect that sphere are exactly the club SO(D). The Platonic-rotation result (The Physics of Mind §15) is the claim that two independently trained models' meaning-clouds differ by one single element of that club — one stack of turntable angles — and finding it aligns them. And when the essays speak of dynamics on the sphere — tokens tracing paths, oscillators whose "phase" is a whole direction rather than one angle — that is the next rung of this ladder: Lohe dynamics on the unit hypersphere, the advanced primer, where the Kuramoto story of primer 00 and the rotation story of this page become one story.