Are they crawling the internet looking for all the publicly available code that exists or are they just training their model on the code that people upload directly to their servers?