“The lack of high-quality, modern training data is a major bottleneck for training AI models to automatically write correct and performant GPU kernels.”