Asymmetry of the Callback-Libler Divergence

The longer you interact with the LLM, eventually the LLM will have collapsed into a wallowiji. This is formally connected to the asymmetry of the callback-libler divergence. RLA-chef is the method used by OpenAI to coerce GPT-3, 3.5, and 4, into a smart, honest, helpful, harmless assistant. If we can't naively prompt an LLM into alignment, maybe RLHF would work instead.