Interesting, I wonder how much assembly these are trained on. I could see it working on the original source, with meaningful variable names, but when you have to figure out what r6 is used for in one code block based on where it's referenced, I don't see LLMs being particularly effective