（来源：天津日报） 转自：天津日报 来自越南的留学生裴明日是天津外国语大学英语口译专业硕士研究生，"昨天我去了杨柳青古镇，拍了 很多视频和图片，我想让更多人通过我的分享，体验中国传统文化、感受中华文明的独特魅力。"裴明 日说，他除了记录在中国的美好生活，也展示自己在这片热土上的奋斗历程。2025年12月31日，他收到 百度"文心导师"计划项目组的邀请，在与天津外国语大学众多师生和领域专家的竞争中脱颖而出，成为 该校唯一一名学生评测专家。"我将和团队成员合作，推动科技创新和产业创新深度融合，助力大语言 模型推动语言服务新发展，为促进中越文明互鉴贡献力量。"裴明日说。 在南开大学八里台校区举办的"雅韵清歌——南开大学2026新年音乐会"上，600余名师生及校友代表在 歌声中迎接新年。开场曲《我的未来式》通过青春洋溢的演绎，瞬间点燃全场。随后，合唱团为观众带 来合唱曲目，多元的曲风与新颖的创作风格，彰显了音乐的魅力与独特性。 本报讯（记者 姜凝）在2026年新年来临之际，我市多所高校开展丰富多彩的庆新年主题活动，为学生 带来属于新年的欢喜和"在校如在家"的温暖，在寓教于乐中弘扬中华优秀传统文化。 1月1日下午，天津 ...

Core Viewpoint - A joint team from the University of Edinburgh and NVIDIA has developed a new method to compress the memory required for AI models, enhancing their accuracy in complex tasks while maintaining response speed and significantly reducing energy consumption [1][4]. Group 1: Memory Compression Technique - The team discovered that compressing the memory used by large language models (LLMs) to one-eighth of its original size improved performance in specialized tests, such as mathematics, science, and programming, without extending inference time [4]. - This memory compression technique, named "Dynamic Memory Sparsification" (DMS), allows AI models to dynamically determine which tokens are essential for subsequent reasoning and which can be discarded, enabling deeper "thinking" within the same computational resources [4][6]. Group 2: Performance Improvements - In tests based on the American Mathematics Olympiad qualification exam (AIME 24), the compressed model scored an average of 12 points higher than the uncompressed model under the same memory read conditions [5]. - The compressed model also outperformed the original model in a professional science question bank created by PhD-level experts and improved its average score by 10 points on a platform assessing coding abilities [5]. Group 3: Implications for AI Development - The DMS memory compression technique challenges the conventional belief that more computational resources lead to stronger AI, suggesting a paradigm shift towards lightweight, high-performance AI [6]. - This approach aligns with human cognitive processes, emphasizing selective memory and key information extraction, which may accelerate the development of general AI [6].

商世东表示，两家公司追求极致体验，在工匠精神的文化比较一致。两家公司有各自的长处，腾讯公司 做生态，影石在音视频有长年的积累。影石做硬件相关的产品有成熟完善的方法论，怎么把硬件产品和 硬件品质做的更好，他们之前做运动相机有很多的积累，他们在影像比如Link这样的设备，在过往可开 发出来的产品，也经受住了市场上的复杂场景考验。"所以，大家在开发这样一款产品不同维度上，形 成了很好的互补，这也是两家公司为什么能够在这么重要的产品上，找到彼此之间的定位，双方一起紧 密联合，可以推出让市场上颇受好评的产品。" 当大厂们都在布局AI硬件时，影石 Insta360（以下简称"影石"） 也联合腾讯做了一款AI录音全向麦克 风。 今年9月，影石宣布推出Wave AI录音全向麦克风，搭载腾讯天籁Inside解决方案。近期，腾讯和影石也 向外部披露了双方合作的细节。影石和腾讯为何会一起推出这款硬件产品？本质上双方推出的产品是解 决室内会议、远程协作和专业内容创作场景中的核心音频痛点。 腾讯云副总裁、腾讯会议天籁实验室主任商世东表示，从现实来看，线上通信体验跟线下有比较大的差 距，在重要的会议活动时，大家更倾向于线下面对面沟通。在 ...

Core Insights - Kimi, a Chinese AI startup, has successfully completed a $500 million Series C funding round, with a post-money valuation of approximately $4.3 billion, supported by major investors like Alibaba and Tencent [1] - The company aims to become a leading AGI (Artificial General Intelligence) player by 2026, with the upcoming K3 model expected to significantly enhance its capabilities [5] Funding and Financial Performance - Kimi's recent funding will be used to aggressively expand GPU resources and accelerate the training of the K3 model, with a current cash reserve exceeding 10 billion RMB [8] - The company has seen a remarkable growth in its commercial performance, with a monthly average growth of over 170% in paid users from September to November, and a fourfold increase in overseas API revenue during the same period [1][7] Technological Advancements - The release of K2 and K2 Thinking models marks a significant milestone for Kimi, establishing it as a pioneer in long-text processing and achieving state-of-the-art (SOTA) results in key benchmarks [3][4] - Kimi's K2 series models have gained global recognition, surpassing leading closed-source models like GPT-5 and Claude Sonnet4 in specific tasks [4] Product Development and Market Strategy - Kimi's product strategy focuses on continuous innovation, with new agent functionalities being launched frequently since May, enhancing the overall user experience [7] - The company is not solely focused on user numbers but aims to push the limits of intelligence and create greater productivity value through its agent products [5] Future Outlook - Kimi's leadership emphasizes a commitment to curiosity-driven development, aiming to explore the upper limits of AGI capabilities and the ideal model characteristics [9] - The company plans to implement significant incentive programs for its employees in 2026, with expected average rewards being 200% of those in 2025 [8]

整理 | 美股研究社 在这个快速 变 化的 时代， 人工 智能技术正以前所未有的速度发展，带来了广泛的机会 。 《AI日 报 》致力于挖掘和分析最新的AI概念股公司和市场趋势，为您提供深度的行 业 洞察和 价 值 分析。 A I 快 报 【AI耳机销量涨超600%，智能穿戴设备功能持续提升】 实时翻译的AI耳机、一键拍摄的AI眼镜、实时监测身体数据的智能手表……从体验尝鲜到健康 生活的"数字伙伴"，AI可穿戴设备功能持续提升。 一家AI耳机相关负责人表示，伴随着大语言模型推出，推理成本下降，AI耳机的人机交互能力 大幅提升。 2025年上半年，AI耳机市场线上销量增长636%。机构数据预测，2025年，中国智能穿戴设 备市场规模有望突破3000亿元，出货量预计突破7100万台，其中智能手表、手环贡献超80% 份额。 【亚马逊：将为美国政府机构AI系统投资高达500亿美元】 日前，亚马逊网络服务（AWS）宣布了一项计划，即首次为美国政府构建并部署专门用于人工 智能和高性能计算的系统，该计划承诺将投入高达500亿美元，以增强该公司为美国联邦政府 客户提供的人工智能和超级计算能力。 【AI公司MiniMax寻求通过 ...

智谱近年呈现出"高增长、高投入、高亏损"的典型科技企业发展特征。招股书数据显示，2022年至2024年，公司营业收入从 5740万元快速攀升至3.12亿元，年复合增长率高达130%；2025年上半年营收进一步增至1.91亿元，同比增幅达325%。公司毛 利率2022年至2024年分别为54.6%、64.6%、56.3%。 智谱收入主要来自大模型收入。公司采用MaaS（模型调用服务）模式，即通过API（应用程序编程接口）调用向开发者和企 业输出通用智能能力。针对AI编程场景的订阅产品也成为智谱MaaS平台收入增长的重要来源，智谱GLMCodingplan上线两个 月，全球付费开发者用户已经超过15万。 深圳商报·读创客户端记者 邱清月 作为国内首家冲刺"大模型第一股"的企业，北京智谱华章科技股份有限公司（下称"智谱"）开启全球招股，预计2026年1月5 日结束，并计划于2026年1月8日正式以"2513"为股票代码在港交所主板挂牌上市。智谱此次发行定价116.20港元/股，发行 3741.95万股，本次IPO募资总额预计将达43亿港元，IPO市值预计超511亿港元。 研发投入高企 华鑫证券研报指出，智谱的上市进 ...

Core Viewpoint - Nvidia is actively pursuing acquisitions to strengthen its position in the AI chip market, particularly focusing on AI21 Labs and Groq, to enhance its capabilities in AI inference technology and maintain its dominant market share of 80% in the AI chip sector [1][3][11]. Group 1: Acquisition Strategy - Nvidia is in advanced negotiations to acquire AI21 Labs for between $2 billion and $3 billion, following its previous $20 billion deal with Groq [1][11]. - The acquisition of AI21 Labs, which specializes in developing large language models (LLMs), is aimed at enhancing Nvidia's ability to create customized enterprise-level generative AI applications [3][4]. - Nvidia's strategy includes not only acquiring technology but also attracting top talent from these companies, as evidenced by the inclusion of Groq's core team in Nvidia post-acquisition [3][10]. Group 2: Competitive Landscape - The AI inference market is becoming increasingly competitive, particularly with the rise of Google's TPU, which poses a significant challenge to Nvidia's dominance [7][10]. - Google's latest TPU v7 shows a substantial performance improvement, with a BF16 computing power of 4614 TFLOPS, compared to the previous generation's 459 TFLOPS, indicating a shift in the competitive dynamics of AI inference [9]. - The focus in the AI industry is shifting from training powerful language models to deploying these models at the lowest cost and latency, which is where Nvidia aims to strengthen its position through acquisitions [10][11]. Group 3: Future Developments - Nvidia is constructing a large R&D center in Kiryat Tivon, Israel, which is expected to include 160,000 square meters of office space and is set to begin operations in 2031 [6]. - The rapid growth in demand for AI inference capabilities is projected to double every six months, highlighting the urgency for Nvidia to enhance its technological offerings and ecosystem [10].

Core Viewpoint - Apple's AI strategy is notably different from competitors like OpenAI, Google, and Meta, which have heavily invested in AI infrastructure. The company has taken a more restrained approach, leading to criticism regarding its lag in AI development, particularly with Siri's performance compared to advanced conversational systems [1][2]. Group 1: Investment Strategy - Apple has opted for a conservative investment strategy in AI, in contrast to competitors who have spent billions on data centers and chip acquisitions [1]. - The company retains a team for developing its own language models but believes that these models will become commoditized in the coming years, making large investments in custom models unwise at this stage [1][2]. Group 2: Future Plans - Apple plans to utilize Google's Gemini large model to support a significant upgrade of Siri in 2026, with personalized Siri being a key focus [1]. - The anticipated launch of the upgraded Siri is expected in spring 2024, marking a major AI initiative for Apple [1]. Group 3: Market Position - The current market dynamics are shifting in favor of Apple, as investors question the short-term revenue returns from the substantial expenditures of companies like OpenAI [2]. - By limiting capital expenditures on AI, Apple maintains over $130 billion in cash and securities, positioning itself to take advantage of potential acquisition opportunities in the AI startup space if valuations decline [2].

Group 1 - The company has completed the localization deployment of the DeepSeek large language model and the Tongyi Qianwen video analysis model [1] - These models have been integrated into the company's cloud management platforms, including Boyun Chelian and Boyun Shikong [1] - The deployment enhances user experience and customer stickiness while improving R&D efficiency and reducing new product development costs [1]

Manus帮助Meta补完了属于它的AI叙事 硅谷还在圣诞假期中，美股科技"七巨头"之一的 Meta 却玩了一笔大交易。 12月30日，由华人团队创立的通用AI智能体公司——Manus AI运营者蝴蝶效应公司宣布，被Meta 公司 收购。 Manus AI 官方社交媒体和公司创始人也确认了这一消息。 "Manus正在迈入新篇章：我们加入了Meta，致力将通用AI智能体带入下个阶段。"Manus AI官方社交平 台账号表示。 来源：Manus AI官方社交媒体账号 公司CEO肖弘表示："加入 Meta 让我们能够在更强大，更可持续的基础上继续发展，而不会改变 Manus 的工作或决策方式。" 据透露，肖弘未来还将担任 Meta 公司的副总裁。但 Manus AI 团队依旧在新加坡展开业务，并独立运 营。 Meta 和 Manus AI 均未透露收购对价——不过考虑有消息指出此前 Manus AI 正在寻求估值为20亿美元 的新一轮融资，最终对价或高于以上水平。 Meta 巨资收购Manus的原因，硅谷和华人创业圈众说纷纭，但有一点是共识：Meta 暂时需要 Manus， 帮助自己完成属于 Meta 的AI叙事。 ...