来源：科技日报 英国爱丁堡大学与英伟达的联合团队开发出一种新方法，能够压缩人工智能（AI）模型运行时所依赖的内存，从而在保持响应速度不变的情况下，提升模 型处理复杂任务的准确性，或显著降低其能耗。这也意味着，更小的内存将带来"更强的AI"，有望打破大语言模型（LLM）性能瓶颈。 团队发现，将LLM所使用的内存压缩至原有大小的1/8后，模型在数学、科学和编程等专业测试中的表现反而更好，且推理时间并未延长。这一方法亦有助 于模型同时响应更多用户请求，从而降低单个任务的平均功耗。除了节能优势，这项改进还有望使AI更适用于处理复杂问题的系统，或存储速度较慢、内 存容量有限的终端设备，例如智能家居产品和可穿戴技术。 AI模型通常通过"思考"更复杂的假设，或同时探索更多可能性来寻找答案。在此过程中，模型需要将已生成的推理线程内容暂存于一种称为"KV缓存"的内 存中。随着线程数量增多或线程长度增加，KV缓存的体积会迅速扩大，成为性能瓶颈，拖慢模型输出响应的速度。 为突破这一限制，团队提出了一种名为"动态记忆稀疏化"（DMS）的内存压缩技术。该方法并非保留所有生成的标记（即AI模型处理的基本数据单元）， 而是动态判断哪些标记 ...

Core Viewpoint - The competition between China and the United States in the field of artificial intelligence (AI) is likened to a football game, with the U.S. currently leading 24 to 18 at halftime, but China is gaining momentum [1][3]. Group 1: Game Analogy - The Wall Street Journal uses a football game analogy to explain the AI competition, emphasizing that both sides can claim victory and the stakes are economic and military leadership rather than a trophy [3]. - The U.S. scored points through various advancements, including ChatGPT and Nvidia's contributions, while China made significant gains with DeepSeek and Huawei [4][5]. Group 2: Expert Opinions on Score - Different analysts provide varying perspectives on the score, with Chris Miller suggesting a 24 to 12 lead for the U.S., citing the U.S.'s ability to monetize AI compared to China [6]. - Deepika Giri offers a closer score of 21 to 19, highlighting China's rapid rise through innovations like DeepSeek [6]. - Other analysts, including Tarun Chhabra and Saif Khan, provide scores of 21 to 14 and 24 to 17 respectively, emphasizing the U.S.'s advantages in models and supply chains while acknowledging China's potential [6]. Group 3: Key Factors in Competition - The article identifies chips and chatbots as critical components of the AI competition, with chips likened to quarterbacks and chatbots to receivers [7][8]. - Trump's recent decision to allow Nvidia to sell older chips to China is seen as a strategic move that could impact the competitive landscape [7][8]. - Nvidia's H200 chip, although older, is still considered superior to many Chinese alternatives, potentially narrowing the U.S.'s computational advantage [8][9]. Group 4: Chatbot Developments - Chatbots are highlighted as crucial for achieving breakthroughs in AI, with U.S. companies currently dominating the leaderboard for top models [11]. - Chinese companies, including Alibaba and DeepSeek, are also making significant strides, with DeepSeek's recent achievements showcasing their competitive capabilities despite hardware limitations [11][12]. - The article raises questions about the future competitiveness of Chinese companies if they gain access to better hardware [12][13].

据公开报道，灵光上线两周后，用户就已成功创建330万个闪应用。不到1个月时间增长至1200万。用户 创建的灵光闪应用已覆盖娱乐与陪伴、生活服务、效率工具、教育与自我提升等主要场景。 责任编辑：宋雅芳 责任编辑：宋雅芳 新浪科技讯 12月26日上午消息，通用AI助手灵光宣布：灵光用户已成功创建1200万个闪应用。闪应用 是灵光三大功能之一，用户无需任何编程基础，用自然语言描述自己的需求，灵光便可最快30秒生成一 个可编辑、可交互、可分享的小应用。闪应用创建数的增长，显示出这一产品形态正在被普通用户快速 接受与持续使用。 新浪科技讯 12月26日上午消息，通用AI助手灵光宣布：灵光用户已成功创建1200万个闪应用。闪应用 是灵光三大功能之一，用户无需任何编程基础，用自然语言描述自己的需求，灵光便可最快30秒生成一 个可编辑、可交互、可分享的小应用。闪应用创建数的增长，显示出这一产品形态正在被普通用户快速 接受与持续使用。 据公开报道，灵光上线两周后，用户就已成功创建330万个闪应用。不到1个月时间增长至1200万。用户 创建的灵光闪应用已覆盖娱乐与陪伴、生活服务、效率工具、教育与自我提升等主要场景。 ...

文 | 沈素明 想象一个场景。 一家制造企业花费了数十万的预算，接入了市面上最先进的大语言模型（LLM）。员工们兴奋地尝试 让这个"无所不知"的AI助手来处理日常工作。 有人问道："我们公司的 XX 产品，最新版本的设计参数是什么？" AI助手礼貌地回答："抱歉，我无法访问您公司的内部产品信息。" 另一个人问："那去年第三季度的设备故障率是多少？我想写个分析报告。" AI助手再次摊手："我无法访问您企业的内部数据库和历史数据。" 员工们感到困惑了："你不是号称最智能的AI吗？为什么连我们公司自己的事都不知道？" 这不是AI不够聪明，而是我们对通用AI的能力产生了误解。ChatGPT、文心一言这些通用大模型，它们 是基于庞大、但公开的互联网数据训练出来的。它们博学多才，能写诗、能编程、能分析宏观经济，但 它们对企业的专有知识——那些内部流程文档、产品手册、数据库记录、私人聊天记录——一无所知。 通用AI是"外人"，而企业需要的是一个"内部专家"。企业想把AI真正用起来，就必须解决这个核心矛 盾：如何让通用AI，快速、准确、且低成本地掌握企业内部不断更新的专有知识？ 解决方案就是目前在大型语言模型应用中最受欢迎的 ...

Core Viewpoint - Ant Group is focusing on the health sector as a key area of its AI strategy, with the recent upgrade of its AI health application AQ to "Ant Aifu" marking a significant evolution from an AI tool to an AI health companion [2][10]. Group 1: Product Development and Market Response - The new version of the app quickly reached the top 6 in the Apple App Store download rankings within 24 hours of its release [3]. - Ant Group has made significant moves in the AI field over the past month, including a major organizational restructuring and the launch of the general assistant "Lingguang," indicating a clear strategy of advancing both general and specialized AI simultaneously [5][6]. Group 2: Market Trends and User Demand - There is a growing structural change in health demand among the public, with a shift in focus from "treating illness" to "preventing illness," and a younger demographic increasingly interested in health management [14]. - The aging population in China is projected to reach 310 million by the end of 2024, creating a long-term demand for health management services [14]. - Ant Aifu addresses a significant gap in the market for reliable health information, answering over 5 million questions daily, with 55% of inquiries coming from lower-tier cities [14]. Group 3: Competitive Landscape and Strategic Positioning - Ant Group's approach to AI is characterized by a dual strategy of general and specialized AI, with "Lingguang" focusing on creating simple applications and "Ant Aifu" targeting the health sector [8][22]. - The rapid growth of "Ant Aifu," achieving a monthly active user growth rate of 83.4%, significantly outpaces the industry average of 13.5% [8]. Group 4: Trust and Professionalism in Health AI - The health sector is sensitive to reliability and professionalism, making user trust a critical asset for health AI applications [19][24]. - Ant Aifu's success is attributed to its deep understanding of the health industry, a solid user base, and a robust technical foundation built over years of experience [16][19].

Core Insights - The AI assistant "Lingguang" has successfully created 3.3 million "flash applications" within two weeks of its launch [1][3] - Lingguang is developed by Ant Group and features three main functions: "Lingguang Dialogue," "Lingguang Flash Applications," and "Lingguang Open Eye" [1][3] - The app achieved 2 million downloads in just 6 days, surpassing ChatGPT's first-week downloads of 606,000 and Claude's 157,000 [1][3] - Lingguang reached 1 million downloads in only 4 days, faster than Sora2, which took 5 days [1][3] Application Categories - The user-created flash applications primarily cover five categories: - Entertainment tools such as interactive games and emotional relief [1][3] - Daily tools like countdowns and to-do lists to enhance user efficiency [1][3] - Educational tools including language practice and self-assessment for learning needs [1][3] - Health management tools like calorie tracking and fitness planning [1][3] - Lifestyle tools such as food lottery and travel planning [1][3]

据悉，蚂蚁集团推出的全模态通用AI助手灵光于今年11月18日发布，其中，"灵光闪应用"功能支持最快 30秒生成一个小应用，即使是完全不懂代码的用户，也能通过简单的对话，快速创造出满足个人和家庭 需求的专属应用，网友晒出的应用五花八门，包括"辅导作业赛博功德箱""遛娃抽签器""元气满满加油 站"等，让AI助手不再只会"回答问题"，而是有了 "可交互的行动能力"。 编辑 杨娟娟 新京报贝壳财经讯（记者潘亦纯）11月24日，贝壳财经记者获悉，通用AI助手灵光上线6天总下载量已 突破200万：在首次破百万下载用时4天刷新纪录后，再破百万仅用时2天。 目前，灵光在App Store中国区免费应用榜单中维持第六位，App Store中国区免费工具榜维持第一。 校对 王心 ...

（文章来源：贝壳财经） 通用AI助手灵光在上线6天总下载量突破200万：在首次破百万下载用时4天刷新纪录后，再破百万的时 间压缩到了2天，持续领跑全球AI产品的下载增速。目前，灵光在App Store中国区免费应用榜单中维持 第六位，App Store中国区免费工具榜维持第一。 ...

11月24日消息，通用AI助手灵光在上线6天总下载量突破200万：在首次破百万下载用时4天刷新纪录 后，再破百万的时间压缩到了2天，持续领跑全球AI产品的下载增速度。目前，灵光在App Store中国区 免费应用榜单中维持第六位，App Store中国区免费工具榜维持第一。 据了解，灵光首批上线三大核心功能——"灵光对话"、"灵光闪应用"和"灵光开眼"，开创性地在移动端 实现"自然语言30秒生成小应用"，并且可编辑可交互可分享，也是业内首个全代码生成多模态内容的AI 助手，支持3D、音视频、图表、动画、地图等全模态信息输出，对话更生动，交流更高效，极具信息 美感。 蚂蚁集团推出的全模态通用AI助手灵光于2025年11月18日正式发布，首周表现亮眼：在下载规模上， 灵光6天突破200万下载，远高于ChatGPT首周的60.6万和Claude的15.7万；在突破100万的时间上，灵光 仅用4天，也快于Sora的5天。 其中，备受用户喜爱的"灵光闪应用"功能支持最快30秒生成一个小应用，消除了应用开发的门槛，在社 交平台上掀起一股"全民手搓AI应用"的热潮。即使是完全不懂代码的用户，也能通过简单的对话，快速 创造出 ...

Core Insights - The article highlights Tesla's advancements in the Optimus robot project, particularly the development of the "World Simulator" technology, which enhances AI training for both autonomous driving and humanoid robots [1][3][5] - The article discusses the implications of Tesla's end-to-end AI model, which allows for rapid learning and optimization, potentially revolutionizing the robotics and automotive industries [3][6] Tesla's Technological Developments - Tesla's GEN3 version technology has reached the finalization stage, with breakthroughs from domestic suppliers in core components, accelerating factory audits and order placements [1] - The "World Simulator" is a neural network system that generates highly realistic virtual driving scenarios, enabling Tesla's AI to learn the equivalent of 500 years of human driving experience in just one day [3] - The simulator's capabilities are being applied to train the Optimus humanoid robot, aligning with Elon Musk's vision of creating a universal AI that interacts with the physical world [5][6] Supply Chain and Market Opportunities - If Tesla confirms the release of V3 in Q1 2026, it suggests that supply chain contracts could be finalized by the end of 2025, leading to rapid growth over the next five years [8] - Several companies are highlighted as key players in the supply chain, including Ningbo Zhenyu Technology, which has achieved significant revenue growth and is expanding its capabilities in precision components for humanoid robots [9][10] - Sanhua Intelligent Controls is reportedly forming a joint venture with Tesla in Mexico to focus on actuator assembly for the Optimus robot, enhancing its position in Tesla's supply chain [11][12] Company Performance and Projections - Zhenyu Technology reported a revenue of 6.593 billion yuan in the first three quarters of 2025, a year-on-year increase of 31.47%, with plans for significant investments in precision components and humanoid robot modules [10] - Sanhua Intelligent Controls achieved a revenue of 24.03 billion yuan in the first three quarters of 2025, up 16.9%, and is focusing on the bionic robot actuator manufacturing sector [12] - Top Group's revenue reached 20.928 billion yuan in the first three quarters of 2025, with a focus on supplying Tesla's humanoid robot actuators [14] Emerging Players in Robotics - Zhejiang Rongtai is actively expanding into the humanoid robot sector, with strategic acquisitions and investments aimed at enhancing its capabilities in precision components [15][16] - Beite Technology is developing various screw products for applications in humanoid robots, reporting a revenue increase of 17.5% in the first three quarters of 2025 [18] - New Spring Co., a leading automotive interior supplier, is leveraging its relationship with Tesla to explore opportunities in the robotics sector, with a revenue increase of 18.83% in the first three quarters of 2025 [20][21]