Researchers have explained how large language models like GPT-3 are able to learn new tasks without updating their parameters, despite not being trained to perform those tasks. They found that these ...