AI is mostly a dimwit
Anything involving CUDA, or GPUs in general, is beyond AI at the moment. Even when they can work iteratively with errors etc. and have search available etc.
This is the golden litmus test for the AI coding abilities. Doing any serious work with CUDA, or GPUs in general. Not just refactoring simple generic code. There is probably 1-2 OOM of potential algorithmic efficiency gains if the AI agents ever surmount this obstacle. Until that happens, we are still far from the real takeoff.