“Brendan Foody argues that current AI model evaluations are too focused on academic, zero-shot tasks and should instead focus on evaluating performance on complex, economically valuable work.”