AI-Generated Kernel Code: When Only the AI Understands

Jordan Nanos from SemiAnalysis said that OpenAI engineers who were reviewing the code of their own kernel together with them did not themselves understand line by line what exactly it was doing. And it did not matter, because the AI understood.

“This is not garbage code, because it works. But essentially it is literally AI-generated assembly that is just dumped into Gluon — a low-level kernel programming language they built on top of Triton and that uses a very interesting programming model.”

“The point is that the guys who were reviewing this code with us clearly understood hardware very well, understood the concepts of the whole system very well, but at the same time had no idea what exactly this MLA kernel for DeepSeek that they were showing us does.”

“You can go through the code line by line, and it’s literally: no, don’t know, don’t know.”

“But it doesn’t matter. The AI understands it, the AI tests it, and you can see the result. It generates correct kernels that work really very well.”

“And I think this is just one more sign of what lies ahead for us.”

“In and of itself, code no longer necessarily has to be something that a person is obliged to deeply understand and dissect, if the AI knows how to manage data movement and compute elements on the hardware you’ve given it.”