THE ALPHABET'S SHAPE — token embedding geometry
The trained 65×96 token table from the STITCH backbone, put under three meters: the full cosine matrix (nothing hidden), a neighbor explorer, and a PCA silhouette that admits what it can't show.
LIT every number measured from ckpt AMBER the 3D view — PCs 1–3 carry only 14% of variance FIG colors & layout
cosine matrix · 65×65 · class-sorted LIT
hover a cell — blocks on the diagonal = classes that huddle
order:vowelVOWELconsCONSpunctspacedigit
neighbor explorer LIT
click a character…
PCA cloud · soft-GL · drag to spin AMBER: 14% of variance — silhouette only
86% of this table's structure is orthogonal to your screen. The clusters you can trust are in the matrix, not here.
Toddler corner: every letter got its own spot in a big dark room. Nobody told them where to stand — but after training, big-H drifted next to little-h, the dots and squiggles made their own corner, and the vowels lean toward each other just a little. The room has 96 directions and we can only draw 3, so the picture squishes it — the number chart is the honest map.