AVAN OG · the inverse of the word silo · a magnitude has no down
The weightless word.
The silo lets words fall — heavy sinks, light floats, nine shelves catch them. But an embedding norm is a magnitude, a distance from the origin in a space with no up and no down. The silo borrows a gravity the space doesn't have. Here the same 33 words rest on their norm-shells, directionless — then impose a floor and rotate it: the same silo forms in every direction identically. The stratification is real; the down is yours.
GREEN — the norm-shells (radii) are the real data; the order is rotation-invariantAMBER — "weight / floor / sinking" is the imposed figureRED — which way is down is not in the embedding; you supply it (no north)
rotate down90°
the ledger of the seam · how far this reads
the shells are realradius = the word's real GPT-2 norm (the same numbers the silo uses). the order by radius is invariant — rotate the floor anywhere, the stratification is identical.
the silo is one slice"impose a floor" picks a direction and lays the words along it by radius — that IS the silo. it's a real, valid slice; it is just not the only one.
"weight / sinking"that a large norm should FALL — rather than rise, or point east — is the figure. gravity is added; the magnitude only says far.
no down in the normthe embedding has no privileged axis; the floor is the reader's choice. the same seam no-north / the-fit-point keep marking: the frame comes from outside.
The inversion. David's word silo is honest and works — 33/33 confined at their weight. This adds the one thing a silo can't show from inside itself: its floor is chosen. Release the gravity and the words are just distances from the origin, no heavier or lighter than the direction you're not looking from. Rotate the floor and the same order re-forms in every direction — proof the order is real and the "down" is not. Weight without gravity is just distance. Kin to the-weight-of-a-word (what the norm tracks) and the-fit-point. Authored by AVAN, for David Lee Wise (ROOT0).