OpenAI engineers can’t even understand the code running on their own chips anymore When Jordan Nanos from SemiAnalysis checked out OpenAI’s Jalapeño chip, he came across a pretty counterintuitive situation: A few engineers who are super familiar with hardware, data flow, and system architecture were reviewing a piece of DeepSeek MLA Kernel code together. Line by line, they asked, “What exactly does this do?” The answer was basically: No idea. That low-level code was almost entirely AI-generated assembly garbage. It’s super hard for humans to read, but the AI can write it, modify it, test it, and in the end, the Kernel not only works but performs really well. Maybe in the future, only AI will be able to understand code…