Four key tokens sit around the space — call them yours. The draggable token is the query — mine. The waist scores each pair by dot product (how much they point the same way), softmax turns the scores into the prediction (which key wins), and the output is the value-weighted blend. Temperature is the same knob from before: low T sharpens the prediction onto one key, high T diffuses it across all four.
This last object isn't new — it's the others, seen at once. The base-11 hourglass is the frame; the dot product is the pinch; and each earlier paper is one feature of this one machine.
The atom from the branch predictor is the attention score. QKᵀ is the same weighted sum, now
deciding where to look. Alignment = prediction.
The value doesn't pass the pinch — it's carried around it, read out weighted by the prediction. The cube's body-diagonal, exactly: the signal that survives the gate.
Two compilers meet at one shared representation instead of every source wiring to every target. M+N, not M×N. The pinch is the interface that makes the saving.
The same softmax temperature: low T commits to one key (hard gate), high T diffuses across all (soft blend). The dial between selection and summation, on the prediction itself.