min{d(problem)/d(theta)} is essentially what LLMs are doing with a prompt. Every session that a chatgpt user has leads to either a resolution or not, which is the loss function given the prompt used to reach that point. It's getting better at not hallucinating in my experience over just the past 4 months.