作者：刘锋 在 21 世纪的科学版图上，两个 重要 的谜题如同两座遥远而宏伟的山峰，吸引着人类智慧最前沿 的探索者。一座是物理学的终极圣杯：如何将描述宏观宇宙引力规律的广义相对论与描绘微观世 界 奇异 现象的量子力学统一在同一个理论框架之下。另一座则是智能科学的核心难题：智能的 本质是什么？意识又从何而来？ 这两个看似风马牛不相及的挑战，却 有着 深刻的内在关联。 现代 物理学理论中，无论是广义相 对论定义的时空参考系、量子力学引发的测量坍缩，还是热力学 所揭示的熵增方向，其核心都指 向了 " 观察者 " 这一角色。而在人工智能的研究中， "智能体"作为能够感知和决策主体，已成为 AI领域学术与产业的核心。 一个富有启发性的观点逐渐浮现：物理学中的 " 观察者 " ，或许可以 被理解为一种特定 形式的 " 智能体 " 。 2025 年 6 月 ， 我们发表近预印本论文 《 Generalized Agent Theory from First Principles （ 基于第 一性原理的广义智能体理论 ） 》，这是历经 11 年时间 的 探索过程 ，首次较系统的阐述了广义 智能体理论的整体框架 ， 并由此对智 ...

在推广方式上，除了传统渠道之外，我们也同样注重线 上渠道，目前国脉大学养老微信视频号、公众号及官方网 站（www.gmubrc.com）均已上线。自首次对外推广以来， 市场反响热烈，获得多家线上、线下媒体关注，目前平台 曝光量已突破百万级。 8、福州理工学院的办学情况和规模如何？ 答：公司全资举办的福州理工学院是经教育部批准成立 的应用型本科大学，坚持以产教融合的校企协同发展模式 为办学纲领，动态适配产业转型升级与区域经济发展需 求，培养具有社会责任感、科技创新精神和产业服务能力 的高素质应用型人才。截至 2024 年底，福州理工学院已 有在校生近 1.4 万人，根据学校主要招生省份 2025 年高考 人数增长情况预计新学年在校生人数将持续提升。学院根 据国家职业教育发展、产业需求以及招生等情况综合考虑 扩大办学规模、提升办学层次，9 月份新学年马尾基地将 投入使用。 证券代码：002093 证券简称：国脉科技 国脉科技股份有限公司 投资者关系活动记录表 编号：2025-013 投资者关系活动 类别 特定对象调研 □分析师会议 □媒体采访 □业绩说明会 □新闻发布会 □路演活动 □现场参观 □其他（请文字说明 ...

和技术层面的短板相比，视觉方案更具争议之处在于其隐私风险：智能体调用的无障碍权限，是安卓手 机系统内一项非常敏感的权限。一位互联网合规人士称，无障碍权限开启后，可以读取屏幕文本内容， 监视和记录用户的所有操作，其中有可能包括用户输入的敏感信息。因此，安卓官网规定，无障碍权限 必须由用户在设备设置中明确打开后才能启动。（详见：当AI接管你的手机屏幕，如何规避"黑镜"式预 言？） 一位接近监管侧的人士透露，工信部门已表示要严格管控手机的无障碍权限使用。 团体标准划红线 6月13日，广东省标准化协会发布团体标准《智能体任务执行安全要求》（下称《智能体安全要 求》），为智能体开发和运营主体提供了明确的行为准则。团体标准不具有普遍性的强制约束力，由本 团体成员约定采用或者按照本团体的规定供社会自愿采用。 AI智能体赛道方兴未艾之际，近期发布的多份团体标准警惕称，部分智能体存在滥用手机系统底层敏 感权限的嫌疑，智能体开发运营商需应对背后的数据安全和隐私保护隐患。 AI接管屏幕引担忧 根据南都此前的报道，在手机终端上，智能体为了打开第三方App，执行用户给出的指令，通常采取调 用应用程序编程接口（API）或视觉识别来模拟用 ...

港大、camel-ai 等多家机构联合提出了一种名为新的名为 Workforce 的创新多智能体框架，以及配套的 OWL（Optimized Workforce Learning）训练 方法，在通用 AI Assistant 的标杆基准测试 GAIA 上取得了 69.70% 的准确率，不仅刷新了开源系统的最佳纪录，更是超越了多家商业系统以及 OpenAI Deep Research 的开源方案。 该研究成果所有代码均已开源，目前已经在 github 上收获了 17k 的 star。 论文标题：OWL: Optimized Workforce Learning for General Multi-Agent Assistance in Real-World Task Automation 论文地址：https://arxiv.org/abs/2505.23885 项目代码：https://github.com/camel-ai/owl 背景与挑战：多智能体系统的 「 领域壁垒 」 随着 LLM 的飞速发展，单一智能体在处理复杂现实任务时逐渐暴露出局限性。为此，多智能体系统（MAS）应运而生，通过让多个专门化的智 ...

Group 1 - The rapid development of large model technology has made intelligent agents a key focus for AI development institutions, with companies like Tencent, Baidu, and JD increasing their investments in this area [2][3] - Intelligent agents possess autonomous decision-making capabilities, allowing them to perceive environments, plan tasks, and execute them independently, thus acting as assistants to large models [3][5] - Tencent's internal use of its coding intelligent agent has led to a 40% reduction in overall coding time, with AI-generated code accounting for over 40% of the total, significantly enhancing development efficiency [4] Group 2 - The collaboration between traditional industries and intelligent agents is evident, as seen in the partnership between State Grid and Baidu to create a marketing power supply solution intelligent agent [4] - The evolution of intelligent agents has led to enhanced capabilities in self-planning and tool invocation, allowing them to handle complex tasks more effectively [6] - The introduction of the Model Context Protocol (MCP) has facilitated cross-platform compatibility for intelligent agents, enabling them to operate across different application scenarios [6] Group 3 - Multi-agent collaboration is emerging as a new trend in intelligent agent technology, allowing for the division of labor to tackle more complex tasks [7] - Tencent's intelligent agent development platform has introduced a zero-code configuration feature for multi-agent collaboration, reducing the barriers to building intelligent agents [7] - The focus on specific industry scenarios is becoming more pronounced, with companies aiming to integrate intelligent agents into existing business processes to meet real-world needs [8][9]

Core Insights - The current stage of artificial intelligence (AI) development is compared to the "dial-up internet" phase, indicating that while AI has potential, it still faces numerous technical challenges before becoming a true productivity tool [4]. - From an application perspective, AI is likened to the "Web 1.0" era, where content is available but users are primarily passive consumers, suggesting that a shift towards more interactive and productive use of AI is necessary for its growth [4]. - The principle of creating value from scarce resources remains constant, with AI's role in generating value through "intelligent agents" that can achieve economies of scale due to their zero marginal cost characteristics [5]. Industry Analysis - The discussion at the "Digital Economy Think Tank" highlighted the need for breakthroughs in AI technology to transition from its current state to a more productive phase [4]. - The economic impact of AI is still not uniformly quantified, similar to the early days of the internet, indicating a need for standardized metrics to assess its value [4]. - The future of AI is expected to lead to a new phase of platform economy, where intelligent agents will enhance consumer experiences by enabling better price comparisons and reducing the prevalence of issues like "big data exploitation" [5].

Core Insights - The article discusses the challenges faced by traditional small and medium-sized enterprises (SMEs) in content production, customer conversion, and high acquisition costs, emphasizing the need for AI-driven solutions in the B2B sector [1][2][3] Group 1: AI Solutions for B2B Enterprises - Baidu's "Love Procurement" launched the first B2B industry intelligent agent solution, integrating video content generation, multilingual output, AI customer service, and search traffic distribution to create an "AI-driven intelligent marketing hub" for SMEs [1][4] - The intelligent agent aims to enhance the entire content-to-conversion process for B2B enterprises, addressing the need for effective content creation and customer engagement [3][9] Group 2: Challenges in Content Creation - Many B2B enterprises struggle with content creation due to a lack of dedicated teams and resources, leading to low-quality and infrequent content updates [2][5] - The issue of content homogeneity and the inability to connect with customer pain points result in low conversion rates, particularly in cross-border B2B scenarios where language and cultural barriers exist [2][5] Group 3: Intelligent Agent Capabilities - The intelligent agent can produce hundreds of differentiated video contents from just 10 seconds of raw material, significantly alleviating content scarcity issues [5][6] - It supports over 20 languages for simultaneous output, including Arabic, Spanish, and Russian, facilitating global outreach for businesses [5][6] Group 4: Impact on Marketing and Sales - Companies using the intelligent agent have reported significant improvements in marketing effectiveness, with one case showing a 148% increase in visitor numbers and a 22.5% rise in lead generation [5][6] - The intelligent agent enhances customer service through AI-driven interactions, allowing for tailored responses based on industry terminology and product logic [5][7] Group 5: Future of B2B Operations - The integration of AI into B2B operations is transforming traditional business models, shifting the focus from merely selling products to enhancing visibility and engagement [9] - The intelligent agent is positioned as a key connector between tools, platforms, and industries, redefining the operational landscape for B2B enterprises [9]

本文来自 Anthropic 官网的分享，详细阐述了他们是如何构建多智能体研究系统《How we built our multi-agent research system》。 他们研究发现， 多智能体系统最适合三类场景：高价值并行任务、超出单上下文窗口的信息处理、需要操作多个复杂工具的情况。需要共享上下文或存 在复杂依赖关系的场景目前并不适合多智能体方案。 下面是翻译全文： 我们的"研究"功能采用多智能体协作架构，让 Claude 能更高效地探索复杂课题。本文将分享系统构建过程中遇到的工程挑战与经验总结。 如今 Claude 已具备跨网络、Google Workspace 及各类集成系统进行信息检索的研究能力，以完成复杂任务。这套多智能体系统从原型到生产的实践历 程，让我们在系统架构、工具设计和提示工程等方面积累了宝贵经验。 多智能体系统由多个自主使用工具的 LLM 智能体协同工作，在我们的研究功能中，主智能体会根据用户查询规划研究流程，随后创建并行工作的子智能 体进行信息检索 。这类系统在智能体协调、评估与可靠性方面带来了全新挑战。 本文将拆解我们验证有效的设计原则，希望能为开发者构建多智能体系统提供参考 ...

Core Insights - The current era is ripe for the emergence of "epoch-making companies" in the AI sector, with a significant gap between models, product capabilities, and actual user needs [1] - AI is evolving from a tool for efficiency enhancement to a core driver of a new generation of product paradigms, with successful AI products being key to defining the next generation of epoch-making companies [1] Event Overview - The 2025 Global Product Manager Conference will address critical questions regarding product innovation in the AI era [2] - The conference, organized by CSDN & Boolan, will take place on August 15-16 in Beijing, featuring top experts from over 40 industries discussing 12 major themes [4] Keynote Topics - The conference will feature discussions on various topics, including the productivity revolution brought by generative AI and the Skywork Agent framework [7] - Key questions include how to reshape user experiences, define new product logic, and master essential engineering capabilities in the AI era [8] Notable Speakers and Their Topics - The conference will host several prominent speakers, including: - Fang Han, CEO of Kunlun Wanwei, discussing the ultimate form of generative AI and its productivity revolution [7] - Wang Yuan, CEO of Jiuhen Technology, exploring new interaction paths in the GenAI era [13] - The founder of YouMind, discussing how AI products can connect emotionally with users [17] - Zhou Chunzhao from NetEase, explaining how intelligent agents can redefine work paradigms [23] - Huang Zixun from vivo, focusing on the productization path of system-level AI capabilities [27] - Zhao Jiuzhou from WPS, sharing experiences in creating practical AI capabilities for the mass market [32] - Sun Shiquan from Alipay, discussing the new paradigm of creative production driven by AIGC [38] - Hu Tengyu from Suoyun AI, analyzing the application of AI agents in manufacturing and education [44] - Yang Yixi, a former product director at Kuaishou, discussing the implementation of AI products in various scenarios [50] - Li Zhiyong, author of "Unmanned Companies," sharing insights on AI-driven business models [72] Additional Insights - The conference aims to foster deep exchanges and value creation among AI product practitioners, technical teams, and innovative enterprises [116][117] - Attendees can register to receive exclusive resources and insights from leading product managers [118][119]

克雷西 发自 凹非寺 量子位 | 公众号 QbitAI o3-pro刚刚也挑战了这两款游戏，而且表现还都不错，直接 突破了benchmark上限 。 具体来说，benchmark中推箱子一共就只做到了被o3-pro突破的第六关；俄罗斯方块则是强行终止的结果，实际上o3-pro根本停不下来。 如果和前SOTA——o3比较，o3-pro的成绩也是直接翻倍。 还有网友直言，比起大模型竞技场，这套标准才更适合做测试大模型的基准。 经典小游戏成为新Benchmark 推箱子、俄罗斯方块……这些人类的经典怀旧小游戏，也成大模型benchmark了。 o3-pro挑战的这两个游戏，出自一套名为 Lmgame 的benchmark，顾名思义就是让大模型玩游戏。 o3-pro挑战的推箱子是从1989年的版本修改而来，在o3-pro之前，评估指标是游戏结束之前推动到目标位置的箱子总数。 不过这次o3-pro直接把所有关卡都通了，颇有种"得一百分是因为卷面只有一百分"的感觉。 但也不必担心，测试基准会动态更新，GItHub仓库中半个月前更新的游戏地图还只有四关，原版游戏更是有足足50多个关卡。 而在o3-pro挑战之前，表现最好的 ...