“An early large-scale unsupervised learning model at Google, trained on 10 million random YouTube frames, developed neurons that were sensitive to high-level concepts like cat faces and human faces without explicit labels.”