Performing well across tasks is difficult, as is benchmarking these models’ performance. Performance tends to vary depending on the task at hand. How can we determine if a model is “better” than another?