Either Google is ignoring that, or crossing their fingers and hoping that one LLM can produce data to train another one.