Ternary weights make the multiply almost free — and point compute at new substrates
BitNet b1.58 constrains weights to {-1,0,+1} (~1.58 bits), turning the core matrix-multiply into add/subtract/skip and cutting memory >90%; Microsoft's BitNet b1.58 2B4T (2025) was the first open native 1-bit LLM (~0.4 GB, ~6× less energy), and MediaTek's Dimensity 9500 now ships native ternary compute-in-memory. Because the operation reduces to ×{-1,0,+1}, it widens where computation can physically happen — the on-ramp to photonic and even etched-weight silicon (a signal-tier extrapolation, not yet product).
Why it matters (spine — atoms). Doing more with fewer bits, out to the edge. Track ternary-native silicon (MediaTek, and any photonic entrant) as the substrate that makes edge intelligence ubiquitous in vehicles, robots, and devices; design for on-device inference by default.
What would prove it wrong (falsifier). Ternary silicon stays niche and GPUs/FP4 remain the only economic path, or photonic/etched-weight compute shows no viable route within 3 years.
Live-status check + triangulation. BitNet b1.58 constrains weights to {-1,0,+1} (1.58 bits), turning the core matrix-multiply into add/subtract/skip and cutting memory >90%; Microsoft's BitNet b1.58 2B4T (2025) was the first open native 1-bit LLM (0.4 GB, ~6× less energy), and MediaTek's Dimensity 9500 now ships native ternary compute-in-memory. Because the operation reduces to ×{-1,0,+1}, it widens where computation can physically happen — the on-ramp to photonic and even etched-weight silicon (a signal-tier extrapolation, not yet product). Source: Paper · Microsoft (BitNet b1.58 2B4T) (2025-04-16); see source link. Every post-cutoff figure was tool-fetched against a live source during the Kimi K3 sweep (2026-07-20); see sweeps/2026-07-20-VERIFIED_DIGEST.md for the full provenance and any corrections of record.
Track ternary-native silicon (MediaTek, and any photonic entrant) as the substrate that makes edge intelligence ubiquitous in vehicles, robots, and devices; design for on-device inference by default.
domains: edge-compute · quantization · atoms