Around 2019, a friend working at a large cloud services provider excitedly shared that his team had trained a CNN model for document text recognition with good results, but the computing cost was too high to deploy directly. They then used distillation to train a smaller model for service, and it turned out to be highly effective.