# LLMs and Self-Referentiality: A Fresh Take on an Old Idea
The post opens with a reflection on Douglas Hofstadter's classic book *Gödel, Escher, Bach*, which convinced a generation of readers that self-referentiality and "strange loops" would be central to understanding intelligence and building AI. The author revisits this idea with fresh eyes, prompted by thinking about how modern large language models (LLMs) actually work. This sets up an interesting tension between an older philosophical framework for thinking about minds and the very different reality of today's most powerful AI systems.
The core observation is that LLMs are, in a striking way, deeply self-referential by nature. These models are trained on vast amounts of human-generated text, which itself contains enormous quantities of writing about language, reasoning, mathematics, and even AI itself. In other words, the model is in some sense "reading about" the very processes it is performing. The author finds this worth examining carefully, since it suggests that self-referentiality may have snuck into modern AI through a very different door than Hofstadter or earlier thinkers anticipated.
The post ultimately raises more questions than it answers, which the author openly acknowledges. Rather than claiming LLMs have achieved anything like the conscious strange loops Hofstadter described, the author invites readers to think critically about what self-referentiality in AI really means today. It is a short but thought-provoking piece that bridges classic ideas in cognitive science and philosophy with contemporary machine learning, making it a rewarding read for anyone curious about the conceptual foundations of AI.