Aigarth bets on ternary AI training

The evidence is strong for ternary quantization, but much thinner for Aigarth's exact evolutionary training recipe.

• –TWN, STTN, xTern, and related hardware papers show +1/0/-1 quantization has credible momentum for compact inference.
• –Native training with evolutionary search is less common, but it is not invented from scratch: TOT-Net and the 2024 ICCAD work on ternary neurons explore optimization-heavy, hardware-adjacent approaches.
• –Aigarth's differentiator is the systems story - decentralized compute, self-directed selection, and an explicit "unknown" state - which makes it more of a research manifesto than a standard ML stack.
• –If Qubic wants ML credibility, it needs reproducible benchmarks and head-to-head comparisons with conventional ternary training baselines.