Models are trained by AI experts with significant compute resources, while final model alignment is often done by software engineering teams with less AI expertise, who also want model alignment to fit into conventional software engineering practices. 

While there are established tools and techniques for model alignment, it is still more art than science, more incantation than reproducible engineering process, especially for development teams without strong AI expertise.

This talk covers the following:

1. A quick summary of the objectives for and challenges of alignment.
2. A summary of what an ideal, repeatable process should be.
3. An examination of the techniques developed in the last few years for defining and implementing alignment “”capabilities”” in a standardized, programmatic, repeatable, and efficient way.
4. A walk through of a specific example demonstrating the details.
5. A summary of the benefits for alignment as an engineering process, as well as additional work required.