OpenAI has released optimized GPU kernels for neural networks with block-sparse weights, a type of architecture that is less common but can be faster than traditional methods. These kernels can run significantly faster than existing libraries such as cuBLAS or cuSPARSE, depending on the level of sparsity chosen. OpenAI claims to have achieved state-of-the-art results in text sentiment analysis and generative modeling using these kernels. --- Why it matters: This matters because it provides a potential performance boost for researchers working with neural networks, particularly those focused on text and image generation tasks. By optimizing GPU kernels for block-sparse architectures, OpenAI is addressing a specific need in the field of AI research. Source: https://openai.com/index/block-sparse-gpu-kernels