AFAIK the go and chess models played themselves to improve. They weren't trained on games played by junior players. So far, coding LLMs don't try to compile an

AFAIK the go and chess models played themselves to improve. They weren't trained on games played by junior players. So far, coding LLMs don't try to compile and run their code and check the output. They don't fuzz the inputs to look for errors. There aren't competitive, adversarial LLMs trying to outcode each other or fool each other.