[AINews] Megakernels are so dead and so back

A Meta's FAIR-originated megakernel called Mixture-of-Kittens (MoK) was open-sourced by Cursor, achieving up to a 2.37x speedup over the strongest public baselines for MoE training on NVL72s. This is notable because the AI research community has been actively debating whether 'megakernels'—hand-fused, monolithic compute kernels—are a dead end, given their extreme engineering complexity and upcoming hardware features in Nvidia's Rubin architecture that aim to solve the same launch-overhead and kernel-overlap problems. Cursor's release serves as a high-profile counterpoint that megakernels are still a viable and impactful approach today, claiming a 41% increase in overall tokens per second that translates to significant cost savings at scale. The source also notes that the team lead for MoK previously co-authored a megakernel project called ThunderKittens.