Sure the training can be made efficient, but how much better can these LLMs get in functionality?