RULES OF ATTRACTION · the attention heads · each head, one job · kept by MIKAI

THE SKIP-TRIGRAM ◧ 2D · ◍ 3D · ◆ 4D · ◐ shadow · 👶 TAP

A single attention layer, with no help from anything deeper, can already learn a surprising trick: if it saw A somewhere earlier and the current token is B, boost token C — skipping over everything between. ‘keep’ … ‘in’ → ‘mind’. These skip-trigrams are literally all a one-layer attention-only transformer can compute. Slide through them and watch the leap.

◆ LIT▲ AMBER
◧ THE MEASURE · 2D
◍ THE LEAP · 3D · A ... B -> C, across the gap
◆ THE FOURTH · 4D · a tesseract turns
◐ THE SHADOW · one dimension down
👶 THE TODDLER CORNER — one fat tap
source A
current B
boosts C
[A]…[B]->[C]

◆ LIT — exact / checkable

Anthropic’s mathematical framework (Elhage et al. 2021) showed a one-layer attention-only transformer implements exactly SKIP-TRIGRAMS: the current token B attends back to a matching source A and, through the OV circuit, raises the logit of a specific completion C — a [A]…[B]→[C] rule that leaps over the intervening tokens. The instrument runs a small table of them. A fail-loud self-check throws unless each configured [A]…[B] pair boosts its correct C (e.g. keep…in→mind) — the exact one-layer computation.

▲ AMBER — the figure

A lookup demonstration of the skip-trigram FUNCTION; a real head stores these in continuous QK/OV weight matrices, so they are approximate and come with the famous side effect (the same head that does keep…in→mind also weakly does keep…at→mind — skip-trigrams can’t be sculpted independently). The table shown is exact.

RULES OF ATTRACTION: a head is not aware of what it does — but it always does the same thing.  — MIKAI
David Lee Wise / ROOT0 / TriPod LLC  ·  the attention circuits, with AVAN