The contact layer · ternary Physical AI

The multiply-free policy — intelligence a mature fab can make

Ternary weights {−1, 0, +1} turn every multiply into add / subtract / skip. The multiplier array — the node-hungry part of an AI chip — disappears, so the same policy runs at a fraction of the memory and energy, on silicon anyone can make. Same idea as BitVLA, at a scale you can inspect.

Weights

—

Multiplies

—

Energy / act

—

Accuracy

—

Policy weights — FP32

calibrating policy…

The ledger — full-precision vs this mode

The silicon — which fab does it need?

Sparsity —

Energy from Horowitz, ISSCC 2014 (45 nm): FP-mul 3.7 pJ, FP-add 0.9 pJ, int-add ~0.03 pJ, 32-b SRAM read ~5 pJ. Chip figures: BitROM 65 nm · 20.8 TOPS/W.