That said, evaluating AI systems is challenging and there is more work to do to understand how to surface differences between models trained with different constitutions:

→ View original post on X — @anthropicai