How to Scale a Large Language Model With Human Feedback

The new effort to humans do ranking responses and so forth to, to then improve that model. It just seems like this human feedback is a pretty big open question there. What your customer would want it to be? What type of feedback they would want? Seems like a whole lot of playbook to be built there, possibly customer specific. But those things are very hard to have a real decisive edge on relative to anyone else.