One of the main things were learning from the current trajectory of large models is that this kind of supervision isn't necessary. It's better to focus on bigger models, more data, and better datasets. Models will improve faster than we can come up with clever ways to add this supervision.