How would we train those LLMs? Feed them binary executables along with their business requirement docs? Is the corpus of training data large enough to be effective? What happens when a new CPU architecture is developed? Where would the training data come from for the new machine code?