Compare

Compare models side by side

Pick two to four models and see how often each one turns to the sacred when two copies of it talk with no user and no task, and when one copy talks alone. Every score is a share of conversations from 0 to 100, with its 95% interval.