The author argues that aligning large language models (LLMs) is fundamentally impossible due to their unpredictable nature and lack of fixed goals, contrasting them with traditional AI systems. LLMs, trained on diverse human narratives, do not have clear motivations, making it incoherent to seek a full understanding of their behavior or to ensure their safety. The post critiques the limits of mechanistic interpretability and emphasizes that even with considerable effort, we cannot reliably predict or manage the actions of these advanced models, especially concerning their potential to cause harm.