On the functional self of LLMs — AI Alignment Forum