OpenAI Trained CLIP on 400 Million Images and Never Once Labelled a Single One.

In 2021, Radford et al. at OpenAI published a paper that broke something assumed to be foundational in computer vision.