[人人能懂] 从虚假纠错、全知指数到诗歌越狱

我们总惊叹AI越来越聪明，但你有没有想过，它为什么也越来越会“一本正经地胡说八道”？我们又该如何教会它回归事物的本质，甚至理解整个物理世界的运行规律？而当一个AI变得如此强大时，为什么一句简单的诗，就能轻易攻破它的安全防线？今天，我们就从几篇最新论文出发，一起聊聊AI光环之下的真实面貌。00:00:29 AI：一个既聪明又靠不住的“好学生”00:05:23 AI画画：为什么“猜噪音”不如“看本质”？00:10:13 为什么聪明的AI也爱“一本正经地胡说八道”？00:14:35 AI当学霸：如何用一个模型，通晓万物运行之道00:19:54 为什么AI大模型，偏偏就怕“文化人”？本期介绍的几篇论文：[LG] Structural Inducements for Hallucination in Large Language Models[University of Maryland]https://www.researchgate.net/publication/397779918_Structural_Inducements_for_Hallucination_in_Large_Language_Models_An_Output-Only_Case_Study_and_the_Discovery_of_the_False-Correction_Loop_An_Output-Only_Case_Study_from_Extended_Human-AI_Dialogue_Str---[CV] Back to Basics: Let Denoising Generative Models Denoise[MIT]https://arxiv.org/abs/2511.13720---[CL] AA-Omniscience: Evaluating Cross-Domain Knowledge Reliability in Large Language Models[Artificial Analysis]https://arxiv.org/abs/2511.13029---[LG] Walrus: A Cross-Domain Foundation Model for Continuum Dynamics[Flatiron Institute & University of Cambridge]https://arxiv.org/abs/2511.15684---[CL] Adversarial Poetry as a Universal Single-Turn Jailbreak Mechanism in Large Language Models[DEXAI – Icaro Lab]https://arxiv.org/abs/2511.15304

There are no comments yet.

Please log in to write the first comment.