It’s basically how large language models like GTP-3 work. They simply predict the next most probable word and suddenly you can prompt them to give you answers to rather intelligent tasks.