I believe Richard Socher may have introduced this terminology, see this tweet and these slides.

• Radford et al 2018 - Language Models are Unsupervised Multitask Learners
• McCann et al 2018 - The Natural Language Decathlon: Multitask Learning as Question Answering (slides, talks about supertasks)

• Richard Socher's tweet regarding GPT-3: “There are 3 equivalent super tasks of NLP: Language models, dialogue systems and question answering. LMs have the most training data→win.”

• Dialog
• Language Model
• Question Answering