“Amjad Masad references a paper from NVIDIA demonstrating that the DeepSeek model could write CUDA kernels autonomously for 20 minutes by incorporating a verifier into its operational loop.”