你有没有觉得，最聪明的AI有时也会犯一些“低级错误”？本期节目，我们就从几篇最新论文出发，去看看AI那些意想不到的“脆弱时刻”。我们将一起探索，为什么AI合作有时会“1+1<2”，甚至被少数派“带偏”；又为什么一个不起眼的错别字，就能让它瞬间“走神儿”。更重要的是，我们将看到科学家们如何像一位“自动马鞍匠”一样，为AI打造不断进化的外部装备，又如何通过“字斟句酌”的反馈，教会AI抵御外界的恶意指令。00:00:36 如何给AI装上一个“自动升级”的马鞍？00:05:13 为什么笼统的批评没用？从教AI“防骗”的底层逻辑说起00:10:36 1+1 < 2？合作的隐形成本00:15:48 一个好汉三个帮，AI为何越帮越忙？00:20:58 为什么一个错别字，就能让AI“走神儿”？本期介绍的几篇论文：[AI] AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces[POSTECH & KAIST & Southern University of Science and Technology]https://arxiv.org/abs/2608.23041 ---[AI] SecOPD: Mitigating Adaptive Prompt Injections by On-Policy Distillation[UC Berkeley]https://arxiv.org/abs/2608.21500 ---[CL] The Collaboration Tax: How Much LLM Multi-Agent Systems Pay to Coordinate[University of Notre Dame & Meta Superintelligence Labs]https://arxiv.org/abs/2608.22152 ---[CL] Aligned Alone, Misaligned Together: Forecasting Adversarial Capture in LLM Agent Populations[ETH Zurich & Tel Aviv University]https://arxiv.org/abs/2608.22444 ---[CL] Lexical Perturbations Disrupt LLM Reasoning: An Empirical Study of Attention Diversion[Missouri University of Science and Technology & University of North Texas]https://arxiv.org/abs/2608.22140 在小宇宙查看该单集文稿

Ask this episode

“What did the council decide on item 42?” — answered with the speaker and the moment.

Transcript

Related