On: X X^t can be faster

[6gvONxR4sf7o]

In inference that's often the case, but numerical stability in training is often a massive lift for new algorithms.