How to Make GPT-3 Coherent

GPT-3 is very much an autocomplete generator like in your phone. It does not only contain low-level stuff, it contains relationships on all levels of the hierarchy. So it figures out a lot of things based on that data because it's in there. But it's very difficult to make GPT-3 coherent. Can we come up with a loss function that is not optimizing for autocomplete?