you can invent a new language, ask LLM to translate existing code bases into it, then train on that

Just like AlphaZero ditched human Go matches and trained on synthetic ones, and got better this way