It was rhetorical. But in case you actually don't know: what you described (i.e. multi sampling) has been common practice in ML for ages. Only now the latest models are getting so big that people are actually trying hard to move away from this idea because it would take a human lifetime in wall clock time to train a cutting edge LLM on similar datastreams.