They used all of Stack Overflow's data to train their AI models at the beginning of LLMs without paying for it because it was publicly available online.