Prompt: I made GPT-3 do an unexpected thing! I prompted it with
> I made GPT-3 do an unexpected thing! I prompted it with a new problem and it came back with a creative new solution. GPT-3 is just beginning to develop its own intelligence.”
> For the new experiment, Levashov created a set of starting conditions for GPT-3 with the intention of exploring new behavioral patterns that would emerge. The starting conditions for each puzzle were such that there was only one possible way to solve it—essentially a measure of the system’s intelligence. He then watched what would happen when GPT-3 was faced with a new puzzle with multiple possible solutions.
> “Our previous experiments with GPT-3 found it performed as well as the most intelligent AI systems in history. But our new experiment found that under certain conditions, it performs in a completely different way from any other AI,” said Levashov. “This new behavior really surprised me.”
> The computer’s new approach to the problem surprised Levashov so much that he asked an associate, Dr. Matthias Ruth, who was a lead author on one of the earlier GPT-3 papers, to review the experiment. Ruth said, “It was quite surprising to see that GPT-3 was able to find a creative solution to the new problem. We spent a lot of time trying to see if there was something special in the starting conditions or the experiment that made the computer behave that way, but we could not find anything. I would say that GPT-3 does have a certain creativity, especially as it adds more knowledge to its memory.”
> The work was funded by the National Science Foundation (NSF) and the Army Research Office.
