AI里最大的Bug，却也是人类文明最伟大的起点。

Core Viewpoint - The article discusses the phenomenon of "hallucination" in AI, explaining that it arises from the way AI is trained, which rewards guessing over admitting uncertainty [4][16]. Group 1: AI Hallucination Mechanism - AI generates incorrect answers when it lacks knowledge, often providing multiple wrong responses instead of admitting ignorance [4][5]. - The training process incentivizes guessing, leading to higher scores for models that guess rather than those that admit they don't know [5][7]. - OpenAI's research indicates that hallucination is a byproduct of the training system, where models are rewarded for incorrect answers if they guess [8][15]. Group 2: Statistical Insights - In a comparison of two models, o4-mini had a higher accuracy rate (24%) but a significantly higher error rate (75%) compared to gpt-5-thinking-mini, which had a lower accuracy (22%) but a much lower error rate (26%) [7][8]. - The abandonment rate of questions was also notable, with o4-mini answering almost all questions (1% unanswered) while gpt-5 had a 52% abandonment rate, indicating a preference for honesty over guessing [8][9]. Group 3: Theoretical Implications - The concept of "singleton rate" is introduced, highlighting that if an information appears only once in the training data, the AI is likely to make errors in judgment [11][12]. - OpenAI argues that hallucination is not an unavoidable flaw but can be managed if AI learns to admit uncertainty [14][15]. Group 4: Broader Reflections on Hallucination - The article draws parallels between AI hallucination and human creativity, suggesting that both arise from a need to make sense of uncertainty [17][31]. - It posits that the ability to create stories and myths is a fundamental aspect of humanity, which may also be reflected in AI's creative capabilities [23][30]. - The discussion raises questions about the future of AI, balancing the need for accuracy with the potential for creativity and imagination [39][42].