all this capability was already easily available in GPT-3 with some prompt engineering, this mostly just changed it from few-shot to zero-shot.