A new AI training method is designed to boost efficiency and mimic human-like reasoning in large language models. Microsoft researchers claim the approach has the potential to be groundbreaking.