大语言模型的成功挑战自指性理论

一位评论者近日反思了大语言模型（LLM）的崛起对传统AI理论的冲击，特别是对《哥德尔、埃舍尔、巴赫》中自指性和奇异循环理论的挑战1。该作品曾强调自指性和奇异循环是智能与AI的秘密所在1。

现代LLM的发展表明这一观点需要重新审视。作者指出，GPT 5.6 Pro和Fable等当代LLM在技术栈的任何层级都没有显式构建自指能力，而其自我反思能力却是预训练的自然副产品，而非工程设计的结果1。这表明自指性并非实现对话智能的必要条件1。

作者认为，智能的核心应该理解为"预测是关于压缩，压缩是关于找到Kolmogorov复杂性的更好上界"，而非必须依赖自指机制1。

原《哥德尔、埃舍尔、巴赫》的作者因LLM的成功而感到震惊和沮丧，因为这挑战了其原有的世界观1。评论者建议将"需要显式自指性才能实现对话智能"的想法列入人类历史上最重要的错误观念中1。

Modern large language models have achieved conversational intelligence without explicitly building self-referential capabilities into their architecture, contradicting long-held theoretical assumptions about the nature of artificial intelligence 1. An author reflecting on this development argues that the widespread success of systems like GPT and similar models fundamentally undermines the framework presented in Douglas Hofstadter's influential work "Gödel, Escher, Bach," which posited that self-referentiality and strange loops are the essential ingredients for intelligent systems 1.

Rather than being engineered into the system, the self-reflective and theoretical reasoning abilities observed in modern LLMs emerge naturally as byproducts of pretraining on large datasets 1. This phenomenon suggests that explicit self-referential mechanisms are not necessary prerequisites for achieving conversational intelligence 1. The author contends that the true foundation of intelligence lies instead in prediction, compression, and finding better upper bounds on Kolmogorov complexity—a computational measure of information content 1. The successful emergence of sophisticated reasoning in LLMs without deliberate self-referential design has reportedly left Hofstadter himself surprised and discouraged, as these developments challenge the worldview central to his seminal work 1. The author proposes that the belief requiring explicit self-referentiality for conversational intelligence deserves recognition among humanity's most significant theoretical misconceptions 1.