“A fundamental limitation of current LLMs is that the amount of computation per generated token is constant, regardless of the complexity of the problem.”