3 ms·Inverse scaling prize for finding a task where larger language models do worse3 points by echen 4y ago