Yup, exactly. It's a good point that for self-supervised workloads, the training set can become arbitrarily large. For a lot of other workloads in the vision space, most data needs to be labeled to be able to used for training.