Written after the deployment, not before.

Improving the safety and helpfulness of large language models often comes at a cost to their raw capabilities. This trade-off, known as the 'alignment tax,' is a fundamental challenge in making AI systems that are both powerful and beneficial.