Which method, then, is the inference built upon, if not the embeddings? And the question still stands, how does "AI" escape the inherent limits of statistical inference?