Not enough! It is insane how many enterprises I meet that are spending $100M / year on inference and don't have offline evals to determine which model to use.