kind of wild “These results show how you can use the latest DeepSeek-R1 model to give better GPU kernels by using more computing power during inference time. This is still a new research area with early results on a promising approach that automatically generates effective attention kernels. “