“Uljan Sharka predicts that future AI models will unify training and inference because the current pre-training paradigm is not sustainable or intelligent.”