But 80% sounds far from good enough, that's 20% error rate, unusable in autonomous tasks. Why stop at 80%? If we aim for AGI, it should 100% any benchmark we give.