“The pre-training paradigm for AI, which gave rise to models like GPT-4, appears to be yielding diminishing returns.”