AI intelligence emerged without explicit self-reference, says Scott Aaronson

Scott Aaronson reflects on the role of self-reference in artificial intelligence, contrasting the ideas presented in Douglas Hofstadter's "Gödel, Escher, Bach" with the current capabilities of large language models. Aaronson notes that despite the emphasis on "strange loops" and self-reference in early AI theory, modern LLMs like GPT 5.6 Pro and Fable exhibit advanced conversational abilities and self-awareness without explicit programming of these concepts. He suggests that these self-referential traits emerged as a byproduct of the extensive pretraining process, similar to how universality in formal systems or computers doesn't require explicit self-reference to be built in. AI

IMPACT Challenges long-held theories about intelligence, suggesting current LLM capabilities emerged unexpectedly from scale and data.

RANK_REASON Blog post by a prominent AI theorist reflecting on the philosophical underpinnings of AI.

Read on Lobsters — AI tag →

• LLMs
• Scott Aaronson
• Douglas Hofstadter
• Fable
• Gödel, Escher, Bach
• Gödel's incompleteness theorems
• GPT 5.6 Pro
• Roger Penrose
• The Emperor's New Mind

• opinion
• other

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →