Supremacy: AI, ChatGPT, and the Race that Will Change the World

by Parmy Olson

these different approaches as being like a new way of educating humans. For instance, suppose two groups of art students were being taught how to paint. The first group was given a book with pictures of paintings, each one labeled with captions like “sunrise,” “portrait,” or “abstract.” That’s how traditional AI models were learning from labeled data. It was a structured and precise method—like telling the art students exactly what each picture represented—but it also limited what machines could infer. They could only recall what had been labeled. The students in this first group would probably struggle to create a painting that hadn’t been specifically described in their book. Now suppose the second group of art students were given access to an entire art gallery, with a vast collection of paintings—and no labels. They were given the freedom to wander around, observe, and interpret the artwork themselves. This was a bit like how GPT was learning from huge reams of unlabeled text. The art students (or the AI model) would find themselves looking for patterns, styles, and techniques on their own, and they’d eventually assimilate this wide variety of examples, as well as the connections between them, without being told exactly what to infer about each one. It would be a much richer form of learning.