У нас вы можете посмотреть бесплатно 用AI做复杂推理,先搞定它的“认知固执”!(RAG开发者必看) или скачать в максимальном доступном качестве, видео которое было загружено на ютуб. Для загрузки выберите вариант из формы ниже:
Если кнопки скачивания не
загрузились
НАЖМИТЕ ЗДЕСЬ или обновите страницу
Если возникают проблемы со скачиванием видео, пожалуйста напишите в поддержку по адресу внизу
страницы.
Спасибо за использование сервиса ClipSaver.ru
告诉 AI 英国首相换人了,它嘴上说“收到”,转头推理时却还在聊前任的八卦?这不仅是机器故障,更是 AI 根深蒂固的“认知固执”!本期视频,我将深度解读来自瑞士洛桑联邦理工学院 (EPFL) 和石溪分校的重磅研究《在冲突知识下的推理测试》(TRACK),揭示为何被寄予厚望的 RAG (检索增强生成) 技术在面对“新旧知识冲突”时会惨遭滑铁卢。当“便利贴”无法覆盖“肌肉记忆”,AI 在金融、编程、医疗等严肃领域的应用将面临怎样的隐形风险? You tell AI the British PM has changed, it says "Got it," but then bases its reasoning on the old one? This isn't just a glitch; it's AI's deep-seated "Cognitive Stubbornness"! In this video, I dive deep into the groundbreaking research "TRACK" from EPFL and Stony Brook University, revealing why RAG (Retrieval-Augmented Generation) fails when facing "Conflicting Knowledge." When "sticky notes" fail to cover "muscle memory," what hidden risks does this pose for AI in finance, coding, and healthcare? ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ 📄 核心内容 & 关键词 | Key Content & Keywords: 首相悖论 (The Prime Minister Paradox): 通过这一经典案例,我们揭示了 AI 在更新单一事实(如首相名字)与进行多步推理(如配偶出生地)之间的巨大鸿沟。 We use this classic case to reveal the massive gap between AI updating a single fact (like the PM's name) and performing multi-step reasoning (like the spouse's birthplace). 参数化知识 vs 上下文知识 (Parametric vs. In-context Knowledge): 深度解析 AI 的“肌肉记忆”(预训练权重)与“便利贴”(RAG 外部信息)之间的内在冲突。为什么在关键时刻,旧习惯总是战胜新知识? Deep dive into the internal conflict between AI's "muscle memory" (pre-trained weights) and "sticky notes" (RAG external info). Why do old habits almost always beat new knowledge in crunch time? TRACK 基准测试 (TRACK Benchmark): 介绍 EPFL 开发的“冲突知识下的推理测试”,涵盖维基百科、代码 (Code) 和数学 (Math) 三大场景,全面扫描 AI 的思维盲区。 Introducing the "Testing Reasoning Amid Conflicting Knowledge" benchmark developed by EPFL, covering Wiki, Code, and Math scenarios to scan AI's reasoning blind spots. 认知固执与风险 (Cognitive Stubbornness & Risks): 探讨了 RAG 技术的局限性,以及在金融风控、医疗诊断等高风险领域,AI 无法“真正”改变主意可能带来的灾难性后果。 Exploring the limitations of RAG and the catastrophic consequences of AI's inability to "truly" change its mind in high-stakes fields like financial risk control and medical diagnosis. ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ 🔔 订阅并加入我的会员 | Subscribe & Join my membership! 你有没有遇到过 AI 即使被纠正了也依然“死性不改”的情况?在评论区分享你的经历! Have you ever experienced AI being "stubborn" or ignoring your corrections even after being told? Share your story in the comments below! 如果你喜欢本期内容,请不要忘记点赞、分享,并【订阅】我的频道,开启小铃铛,第一时间获取关于前沿科技的深度解析。 If you enjoyed this video, please like, share, and SUBSCRIBE for more deep dives into our technological future. 👉 支持我持续创作 | Support My Work: 加入我的会员频道,提前观看视频并获得专属福利! Join my channel membership to get early access to videos and exclusive perks! / @wow.insight ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ 论文链接,请点击会员贴: • Запись ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ #AIAgent #RAG #LargeLanguageModels #MachineLearning #DeepLearning #ArtificialIntelligence #Hallucination #TechAnalysis #FutureofAI #LLM #人工智能 #大语言模型 #深度学习 #认知固执 #RAG技术 #科技解析 #EPFL #思维链 #提示工程