Pretty cool to see this quantified. Although they're judging each model with a different model, I'd be curious to have some accuracy on that. I'd also be really

Pretty cool to see this quantified. Although they're judging each model with a different model, I'd be curious to have some accuracy on that. I'd also be really interested to have the task done by a human (or group of humans, say MTurk) and see how the accuracy compared.