Excited for practical n^2.25 matmul algorithms to instantly make inference and training 1000x cheaper in Q1 2027