OpenAI releases agent skill for custom CUDA kernels for LLM training

AI Impact Summary

OpenAI is releasing a new agent skill that automates the creation of custom CUDA kernels for LLM training, targeting NVIDIA GPUs like H100, A100, and T4. This capability allows coding agents like Codex and Claude to generate optimized kernels for transformers and diffusers pipelines, significantly reducing the manual effort required for developers. The skill provides domain knowledge, templates, and benchmarks, streamlining the process of integrating custom hardware acceleration into LLM workflows. This is a major step towards democratizing access to optimized GPU kernels.