Please read “Language Models are Few Shot Learners” https://arxiv.org/abs/2005.14165 it’s the paper behind gpt-3

Large transformers display few shot learning similar to children. It’s a major advancement, machine learning is much more than supervised learning nowadays.