OpenAI and New Models: How AI Understands Code

Jordan Nanos from SemiAnalysis shared that OpenAI engineers, who were reviewing their core code alongside him, did not fully understand what it did line by line. And that didn’t matter because AI understood it.

"This is not junk code because it works. But essentially, it is literally AI-generated assembler that is just dumped into Gluon—a low-level programming language for cores that 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 entire system very well, but at the same time had no idea what this MLA core for DeepSeek that they showed us actually does."

"You can go through the code line by line, and there literally: no, I don’t know, I don’t know."

"But that doesn’t matter. AI understands it, AI tests it, and the results are visible. It generates correct cores that work really well."

"And I think this is just another sign of what lies ahead."

"The code itself no longer necessarily has to be something that a person must deeply understand and dissect if AI knows how to manage the movement of data and computational elements on the hardware you provided to it."