“The process of iteratively fixing issues in models like ChatGPT via RLHF at OpenAI can be like "whack-a-mole," where solving one problem may inadvertently create a new one.”