> it will learn on it's own.

For example, in any scientific field from time to time "literature review" papers get published. They cover everything relevant to a specific topic, trying to offer a quick overview with jumping points. We can ask GPT-3 to summarise and write review papers automatically.

We can also think of Wikipedia - 5 million topics, each one has its own article. We could use GPT-3 to write one article for each scientific concept, no matter how obscure, one review for each book, one article about each character in any book, and so on. We could have 1 trillion articles extracting all the known things. Then we'd have AI analyse these topics for contradictions, which comes naturally when you put together all the known information about a topic.

This would be a kind of wikiGPT, a model that learns all the facts from a generated corpus of reviews. It only costs electricity to make.