The point of various unsupervised and 'self-supervised' training methods is the ability to use unlabeled data, which enables you to use much more data than it's

The point of various unsupervised and 'self-supervised' training methods is the ability to use unlabeled data, which enables you to use much more data than it's feasible for people to evaluate. Some crude automated filtering (e.g. deduplication) is still helpful, but removing the bottleneck of having a person look at every single item can enable you to use thousand or million times larger datasets, which is what's required to give models a 'general understanding' of a wide domain, by taking an (automated) look at literally everything that is available.