AI 訓練卡在矩陣運算？科學家找到突破口，你的 GPU 可能要升級了

The core idea is that we may be using the wrong mathematical objects to represent information in transformers. Quaternions, the algebra behind 3D rotation, offer a structurally different forward pass. Replace real valued projections in attention with Hamilton products over quaternion valued inputs, and you get the same effect at a quarter of the parameter cost. The more interesting angle is geometric. Methods like LoRA work because fine-tuning updates cluster in a low dimensional rotational sub