[R] Ternary neural networks as a path to more efficient AI - is (+1, 0, -1) weight quantization getting serious research attention?

The viral discussion centers on Qubic's Aigarth project, a novel architecture that trains neural networks natively using ternary weights (+1, 0, -1). This contrasts with the established research path of Ternary Weight Networks (TWN), a 2016 method focused on post-training quantization, where a model is trained with full-precision weights and then compressed. The core innovation of Aigarth is its training mechanism: it reportedly uses an evolutionary selection process instead of the industry-standard backpropagation and gradient descent. This combination of native ternary representation and evolutionary optimization is what makes it an unusual and potentially significant departure from conventional AI training pipelines.

Proponents suggest this approach could yield multiple advantages. Theoretically, models with ternary weights are drastically smaller—approximately 3x more compact than their full-precision counterparts—leading to significantly lower memory footprint and inference costs on hardware. Furthermore, the developers claim that natively trained ternary models might "represent uncertainty more naturally" and "stay adaptive rather than freezing after training." This hints at models that could continue to learn or adjust post-deployment, a challenging feat for standard, statically quantized networks. The research community is now scrutinizing whether this method is a genuinely novel, peer-reviewed direction or an incremental step on prior quantization work.