Google AI Flubs 10% of Answers—DeepMind CEO Warns of Runaway Risk
Audio in Mandarin Chinese · English transcript below
⚡ Google AI flubs 10%, DeepMind yells "runaway!", Liquid edge soars, Tesla FSD green-lit in NL, chip cycle → structure, AI eats shopping & booking—mayhem unleashed!
The rapid advancement of AI, coexisting with potential risks, has become a central focus today. New models from Liquid AI continue to drive technological progress, and AI is also propelling the storage chip industry towards structural growth. However, Google's AI summarization feature has been reported to contain approximately ten percent errors, and DeepMind's CEO has also issued a warning that AI could 'lose control'. These developments collectively remind us that while we enjoy the benefits of AI, we must more cautiously address its reliability and safety challenges.
Today's Top 3 Headlines
- AI
每十条摘要就有一条出错：研究质疑 Google AI 概览准确性
《纽约时报》与初创公司Oumi联合发布的研究指出，谷歌AI Overviews功能在4,000道SimpleQA基准题测试中，约90%的回答准确，约10%含错误或误导信息。该功能2024年上线，已在包括乌克兰在内的多地推出，利用Gemini模型聚合多源数据生成摘要。研究列举两例：询问鲍勃·马利故居改为博物馆日期时，系统引用的来源无明确日期或信息错误；另一例中，AI称某古典音乐机构不存在，却引用了其官网。谷歌回应称，该研究使用的基准可能不准确，未反映真实搜索行为，公司采用更严格内部数据集评估。原文指出，AI Overviews并非固定模型，而是按查询动态选择系统，先进模型因资源消耗高并非总被调用。谷歌已在界面标注AI回答可能出错，但许多用户未进行二次核实。
Source ↗
+6 more headlines
- Liquid AI发布450M参数视觉语言模型LFM2.5-VL-450M
- 存储芯片“超级周期”还能持续多久？
- 荷兰率先批准特斯拉监督版全自动驾驶
- NASA iPhone 时刻、苹果折叠屏、Google AI 购物、WhatsApp 用户名
- 布莱恩·考克斯：AI未来力量未知，既兴奋又存忧
- Google AI Mode “plus”界面，全球餐厅预订功能升级 - 搜狐