LLMs and self-referentiality

Scott Aaronson observes that, despite decades of speculation that self‑reference and “strange loops” were essential to artificial intelligence, today’s large language models achieve conversational ability without any engineered self‑referential component. He points out that the ability of models like GPT five point six Pro to discuss their own operation emerged naturally from the same pre‑training that lets them talk about Pokémon or plate tectonics. According to Shtetl‑Optimized, the notion that explicit self‑reference is a prerequisite for true conversational intelligence belongs among the once‑fashionable but now discredited ideas in the history of science.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton