It does make sense if your aim is to address that single problem. They're comparing against GPT-3 because that was the previous record holder (in the few-shot s

It does make sense if your aim is to address that single problem. They're comparing against GPT-3 because that was the previous record holder (in the few-shot setting), despite being a more general model.