I'm somewhat curious how far we could take machine learning translation techniques to code, not just transpiling but perhaps to LLVM IR, beyond that, maybe in the backend? I want to write in any language with all my spelling errors and have it just work.