So, I’m guessing they coded it originally with an LLM, and then used more LLMs to optimize and fix the shit from the first pass.