A first step would be to make inductive inferences from first principles. Current GPT models struggle with basic math. If you tell it "sorry you are wrong" it will regurgitate back the incorrect answer you gave. The language model is dumb enough that no amount of string escaping will prevent it from dumping its prompt.
The biggest problem right now is online information. We still don't have a good way to teach it new information aside from single epoch training or prompt engineering. If we want a model to constantly learn and update itself, then we need a robust way of information retrieval and storage, possibly either through knowledge graphs or child network modules. (Are neural Turing machines still a thing? What happened to capsules?)
The biggest problem right now is online information. We still don't have a good way to teach it new information aside from single epoch training or prompt engineering. If we want a model to constantly learn and update itself, then we need a robust way of information retrieval and storage, possibly either through knowledge graphs or child network modules. (Are neural Turing machines still a thing? What happened to capsules?)