LLMs and Self-Referentiality
The author reflects on the unexpected emergence of self-referential abilities in large language models (LLMs) like GPT 5.6 Pro and Fable, which surprised even its creator, Douglas Hofstadter. Hofstadter, known for his work "Gödel, Escher, Bach," once believed that self-referentiality was key to intelligence, a notion echoed by Roger Penrose in "The Emperor's New Mind."
However, the author argues that these abilities naturally arose from the models' general intelligence, as they were able to talk about any topic within their training data without any explicit programming of self-reference. This universality, akin to how computers can run their own code as input, was a byproduct of the models' ability to process and predict patterns within their training data.
The author suggests that the focus should be on understanding the models' predictive capabilities, which underpin their conversational intelligence, rather than on the self-referential aspects that Hofstadter and others once emphasized.
Written by urgent.news from Hacker News's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
