鱼羊 雷刚 发自 纽凹非寺 量子位 | 公众号 QbitAI 172天过去，Kimi在深夜低调给出了DeepSeek冲击波后的 回应 。 全新Kimi K2基础大模型，MoE架构，总参数1T，激活参数32B，能力领先性尤其展现在代码、Agent、数学推理任务上。 Kimi援引多个基准评测数据，创造了 开源全新SOTA 。 没错，Kimi K2，这次 发布即开源 。 并且Web端、App和API服务都同步上线可以使用。 作为DeepSeek R1发布后最大的"被冲击者"，Kimi这半年几乎遭遇到了全方位质疑：技术领先性还有吗？营销投流意义几何？以及之前非技 术花边缠身——Kimi还有技术信仰吗？ 172天后，Kimi在深夜用K2大模型给出了低调但明确的回应： 游戏尚未结束，Kimi不下牌桌，Kimi不打算认输。 杨植麟算是被广东大哥梁文锋叫醒了。 Kimi K2：1T参数MoE基础模型 先来看Kimi K2的具体情况。 作为Kimi最新MoE基础模型，Kimi K2强调的是 代码能力 和 通用Agent任务 能力。 总参数量达到1T…属实是让本地部署党捏了把汗，不过激活参数是32B。 支持128K上下文。 而蛰 ...

奕然 发自 凹非寺 量子位 | 公众号 QbitAI 如何用多智能体的方法构建深度搜索？ 现在，Claude团队把自家最新的心得，对外分享了。 在这篇文章中，它详细展示了如何构建一个 有效的多智能体研究系统 ，这是一个架构，其中主代理（The Lead Agent）会生成和协调子代 理（Subagents)，以并行方式探索复杂查询，内容涵盖系统架构、提示工程以及评估方法等。 Claude数据显示了不同行业领域使用此功能的比例——专业领域软件系统开发占比10%，开发和优化专业和技术内容、开发业务增长和创收 策略皆占比8%，协助学术研究和教育材料开发占比7%，研究和审核信息占比5%。 网友们点评： Anthropic团队对AI模型的理解真是killer级别啊。 一起来看看这篇干货教程。 关键架构：协调器-工作器架构 Claude团队使用了协调器-工作器架构，专门用于管理多个智能体之间的任务分配与协作。下图展示了多智能体架构运行情况。 此外，该系统使用 多步搜索 而非静态检索，动态地查找相关信息，适应新的发现，并分析结果来形成高质量的答案。 与单个代理的Claude相比，它在内部评估中成功率达到90%更高，比如，以 ...

机器之心报道 机器之心编辑部 Tokenization，一直是实现真正端到端语言模型的最后一个障碍。 我们终于摆脱 tokenization 了吗？ 答案是：可能性无限大。 最近，Mamba 作者之一 Albert Gu 又发新研究，他参与的一篇论文《 Dynamic Chunking for End-to-End Hierarchical Sequence Modeling 》提出了一个 分层网络 H- Net，其用模型内部的动态分块过程取代 tokenization ，从而自动发现和操作有意义的数据单元。 「这一研究预示着 Tokenizers 正在退场，智能字节分块（Smart Byte Chunks）开始登场。或许无需 Tokenizer 训练的时代真的要来了 —— 可能性无限大。」X 知名 博主 Rohan Paul 表示道。 现阶段，Tokenization 仍然是语言模型和其他顺序数据不可或缺的组成部分，因为它能够压缩和缩短序列。然而 Tokenization 存在许多缺点，如可解释性差，在处 理复杂语言（如中文、代码、DNA 序列）时性能下降等。 迄今为止，尚未有任何端到端的无 tokeniz ...

机器之心报道 知情人士透露，这两人分别是曾在 OpenAI 从事多模态 AI 研发的 Allan Jabri 和 Lu Liu，后续两人将加入 Meta 的超级智能团队。 根据公开资料了解，Allan Jabri 博士就读于加州大学伯克利分校电子工程与计算机科学系，聚焦于自监督学习和无监督学习的可扩展目标和架构，曾在 DeepMind、Google Brain、Facebook 纽约人工智能研究院实习、就职。 编辑：Youli 扎克伯格继续「挖啊挖」，这次「又」轮到 OpenAI 了！ 据 The Information 报道，近日 Meta 首席执行官扎克伯格从 OpenAI 公司「挖走」了 2 名知名 AI 研究人员！ 目前并不清楚 Meta 是花了多大的价钱聘请到了这两位 AI 人才，不过有之前的数千万美元高额薪酬例子在前，相信价钱也不会很低。 不过这还不够，「不差钱」的 Meta 扩张之旅远没结束，招人可以用钱「砸」，还可以通过收购合并。 相比今天曝出的 OpenAI 被 DeepMind 「截胡」，收购 Windsurf 失败，Meta 在收购上也是「春风得意马蹄疾」。 据 Bloomberg 报道 ...

论文作者团队简介：本文第一作者周鑫，共同第一作者梁定康，均为华中科技大学博士生，导师为白翔教授。合作者包括华中科技大学陈楷锦、冯天瑞、林鸿 凯，旷视科技陈习武、丁宜康、谭飞杨和香港大学赵恒爽助理教授。 在 HunyuanVideo 上， EasyCache 在复杂场景下保持与原视频的一致外观，同时显著加速 1. 研究背景与动机 近年来，随着扩散模型（Diffusion Models）和扩散 Transformer（DiT）在视频生成领域的广泛应用，AI 合成视频的质量和连贯性有了飞跃式提升。像 OpenAI Sora、HunyuanVideo、Wan2.1 等大模型，已经能够生成结构清晰、细节丰富且高度连贯的长视频内容，为数字内容创作、虚拟世界和多媒体娱乐带来了巨大变 革。 但与此同时，推理慢、算力消耗高的问题也日益突出。以 HunyuanVideo 为例，生成一个 5 秒、720P 分辨率的视频，单次推理在单张 H20 上需要 2 小时。这种高 昂的资源代价，极大限制了扩散视频生成技术在实时互动、移动端和大规模生产场景的应用落地。 造成这一瓶颈的核心原因，是扩散模型在生成过程中需要多次迭代去噪，每一步都要进 ...

Core Insights - The rise of AI, exemplified by DeepSeek, presents a significant opportunity for value reassessment in China and a reshuffling of urban dynamics in the country [1][2] National Strategy - The Chinese government proposed a "three-step" strategy for AI development in 2017, aiming for a core industry scale of 1.5 trillion yuan by 2020, 4 trillion yuan by 2025, and 10 trillion yuan by 2030 [3] - As of September 2024, the core AI industry scale in China has reached nearly 600 billion yuan, surpassing the 2025 target by 150% [3][4] City Comparisons - Beijing is the undisputed leader in AI, holding over 30% of national resources in talent, patents, enterprises, and financing, with a core industry scale of 300 billion yuan, accounting for half of the national total [5][6] - Shanghai has established itself as a hub for integrated circuit industries and AI chip companies, focusing on the "chip" aspect of AI, while also leading in computing power infrastructure [10][12] - Shenzhen ranks third in AI-related enterprises, with a strong manufacturing base and major tech giants like Huawei and Tencent driving innovation and ecosystem development [12][15][16] - Hangzhou is emerging as a disruptor in the AI space, particularly with the launch of DeepSeek-R1, which has significantly impacted the global AI landscape [18][19][21] Emerging Players - The "Six Little Dragons" from Hangzhou, including DeepSeek and Yushu Technology, are gaining attention for their innovative AI solutions and market share [19][20] - Zhejiang University has contributed significantly to AI research, ranking just behind Tsinghua and Peking University in terms of AI publication output [20] Competitive Landscape - Other cities like Guangzhou, Suzhou, Hefei, and Xi'an are also exploring unique pathways to develop their AI industries, highlighting the competitive nature of China's economic landscape [23][24]

直播介绍 直播时间 7 月 15 日 20:00-21:30 直播主题 Agent 落地实况：能用吗？怎么用？用到哪儿了？ 直播嘉宾 2025 年被称为"AI Agent 元年"，Agent 真的能落地商业化了吗？拆解难点、协作挑战、企业落地 KPI……腾讯云、彩讯股份、商汤科技三位专家深度对话！ 如何看直播？ 王磊 腾讯云智能体平台产品中心总经理 邹盼湘 彩讯股份 AI 产研部总经理 王志宏 商汤科技 / 研发总监 2025，AI Agent 元年，能用了吗？实战场景深度揭秘。 任务拆解难、协作难，Agent 失败真因是什么？专家直击痛点。 落地指标怎么量？ROI、风险和 KPI 一针见血。 戳直播预约按钮，预约 AI 前线视频号直播。 如何向讲师提问？ 文末留言写下问题，讲师会在直播中为你解答。 直播亮点 ...

Core Viewpoint - The Kimi K2 model, released by Moon's Dark Side, is a significant advancement in AI technology, featuring a 1 trillion parameter architecture and demonstrating exceptional performance in various benchmarks and practical applications [1][3]. Group 1: Model Specifications - The Kimi K2 model utilizes a MoE architecture with a total of 1 trillion parameters and 32 billion active parameters, showcasing strong capabilities in code generation and general agent tasks [1]. - The model achieved top scores in several benchmark tests, including SWE Bench Verified, Tau2, and AceBench, indicating its leading position in code, agent, and mathematical reasoning tasks [3]. Group 2: Technical Innovations - During the pre-training phase, the Kimi K2 model employed the MuonClip optimizer, which facilitated stable and efficient training of the trillion-parameter model, enhancing token utilization efficiency [3]. - The model's performance in core capabilities such as Agentic Coding, Tool Use, and Math & Reasoning was notably strong, reflecting its potential for practical applications [3]. Group 3: User Accessibility - Users can now experience the Kimi K2 model through the official website kimi.com or by downloading the Kimi App, with API services available that are compatible with OpenAI and Anthropic's Chat API [4]. - The Kimi K2 model supports a context length of up to 128K, providing enhanced versatility and tool usage capabilities for diverse user needs [4].

Core Viewpoint - Guangdong's artificial intelligence and robotics industry is experiencing significant growth and has established itself as a national leader, but it faces several challenges that need to be addressed for sustainable development [2][3][6]. Industry Scale and Technological Strength - Guangdong's industrial robot production has ranked first in the country for five consecutive years, with a production volume of 246,800 units, accounting for 44% of the national total [3]. - The core AI industry in Guangdong is projected to exceed 220 billion yuan in 2024, representing a year-on-year growth of approximately 25% [3]. - The region hosts over 1,500 core AI enterprises and 147 "little giant" companies in the AI sector, the highest in the nation [4]. Key Challenges and Bottlenecks - There is a notable weakness in foundational key technologies, with insufficient open-source ecosystem development and a lack of high-quality data [6][7]. - The gap between technology and application needs is evident, particularly for small and medium-sized enterprises facing high costs and low AI implementation rates [6][7]. - The high cost of humanoid robots and the lack of universal standards for components hinder large-scale deployment [7]. Recommendations for Development - The report suggests enhancing government functions to accelerate industry growth, focusing on key technology breakthroughs and supporting enterprise innovation [9][10]. - It is recommended to establish a legal framework to support the development of AI and robotics, including specific legislation and monitoring of existing policies [10][11]. - Suggestions include creating demonstration application scenarios and lowering entry barriers for SMEs to facilitate technology implementation [9][11].

Core Insights - Google has agreed to pay approximately $2.4 billion to acquire technology licensing from AI coding startup Windsurf and has hired its CEO along with some employees [1] - The deal follows stalled negotiations between OpenAI and Windsurf, indicating a competitive landscape in AI technology [1] - Alphabet, Google's parent company, is integrating a small number of Windsurf employees into its DeepMind division, focusing on "agentic coding" [1] - Google will also obtain non-exclusive licensing rights to certain technologies from Windsurf [1]