I agree 100% with sentiment. It is inconceivable to me that models were not trained, either directly, or indirectly through other aggregates, on code and binaries and, of course, media, fetched from torrent or other filesharing sites.