“The capabilities of large language models progressed from the equivalent of a preschooler (GPT-2) to a smart high-schooler (GPT-4) in four years.”