北京商报讯（记者 魏蔚）9月25日，月之暗面Kimi对全新Agent（智能体）模式OK Computer（计算机） 启动灰度测试，OK Computer延续"模型即Agent"理念，通过端到端训练Kimi K2模型，进一步提升智能 体及工具调用能力。用户下达需求后，Kimi可操作自身的虚拟电脑，完成多功能网站开发、海量数据 分析、图片视频生成及高品质PPT制作等复杂任务。曾打赏过Kimi的用户将获得首批体验资格。 ...

21世纪经济报道见习记者冉黎黎 北京报道 "现在我们进入智能体的时代，整个发展还有很大的潜力，我们对它充满了期待。"9 月24日—26日，2025年中国国际信息通信展览会（PT展）在北京国家会议中心举行，中国工程院院士邬贺铨在PT展主论坛上如 是表示。热切的期待中，从运营商到产业链企业均在进行智能体的部署。中国联通人工智能首席科学家廉士国预测，未来各个 行业里都会有两类智能体应用，一类是模仿人类经验学习"职业技能"，另一类则是"超人专家"，可以比人类更早地从复杂的原 始数据里找出规律，带来从0到1的突破。 对于智能体未来将处于日常生活中的何种位置，华为公司ICT Marketing与解决方案销售总裁刘康将之与APP进行了类比。"过去 十年，超过300万款APP深度地融入到我们的生活和工作当中，未来终端会走向以Agent为中心。"刘康表示，"到2030年Agent数 量将超过传统的APP，每个人都会有自己专属的智能助手。" 9月24日—26日，2025年中国国际信息通信展览会（PT展）在北京国家会议中心举行。（冉黎黎/图） 从运营商的角度来说，中国移动集团首席专家、研究院网络与IT技术研究所所长张昊告诉21世纪 ...

Core Insights - Qualcomm is the first global tech company to provide a clear timeline for 6G pre-commercial terminal launch, expected in 2028 [1] - The rise of AI is expected to elevate various "edge" devices, including smartphones, PCs, wearables, and robots, into potential blockbuster products [1] - Qualcomm aims to enhance AI capabilities in smart terminals by integrating NPU and GPU technologies during the chip design phase [1] Group 1: AI Integration in Chip Design - Qualcomm showcased its self-developed Oryon architecture at the Snapdragon Summit, emphasizing its superior SoC capabilities for AI processing and energy efficiency [2] - The self-developed Oryon architecture allows Qualcomm to reduce dependency on other companies, particularly in licensing fees and long-term development [2] Group 2: Performance Enhancements - The fifth-generation Snapdragon 8 flagship chip, utilizing 3nm process technology, has achieved a 20% performance increase in CPU, 23% in GPU, and 37% in NPU for AI applications [3] - Major smartphone brands like Honor, Xiaomi, Vivo, and OPPO plan to adopt the fifth-generation Snapdragon 8 in their flagship products [3] Group 3: AI Applications and Collaborations - Qualcomm and Honor are collaborating on developing edge AI model solutions to enhance user experience in applications like image recognition and color extraction [4] - The partnership aims to create GPU-NPU heterogeneous AI technologies to improve gaming graphics through AI learning [4] Group 4: Future AI Applications - Qualcomm predicts that robots and various wearable smart glasses could become as prevalent as smartphones in the future [6] - The automotive sector is identified as a prime application area for edge AI due to its need for local processing capabilities [6] Group 5: Evolution of Intelligent Agents - The future will shift from mobile-centric to intelligent agent-centric systems, requiring advanced understanding and decision-making capabilities [7] - 6G is expected to support real-time collaboration between edge devices and the cloud, enhancing AI-native applications [7] Group 6: Competitive Landscape - Domestic smartphone manufacturers are developing specific AI applications, with Honor's intelligent agent supporting over 200 vertical scenarios and 3000 general scenarios [8] - Other brands like OPPO, Vivo, and Xiaomi are also enhancing their AI capabilities, integrating them with their operating systems [8] Group 7: Challenges and Requirements - The development of AI intelligent agents necessitates advancements in both hardware AI chip capabilities and software interaction methods [9] - Companies must focus on personalized services through AI while addressing cost and technical challenges in their solutions [9]

Core Insights - The article highlights the advancements in AI-driven financial services by JD Technology, focusing on three main intelligent agents: AI Wealth Manager "Jing Xiaobei," shopping assistant "JoyGlance," and inclusive credit assessment for small and micro enterprises [1][2][3] Group 1: AI Wealth Management - "Jing Xiaobei" has undergone a personalization upgrade, allowing for friend-like conversations with users, moving beyond simple Q&A to a more engaging dialogue [2] - The new "user memory system" enables "Jing Xiaobei" to remember user needs and provide proactive, intelligent, and professional services in investment opportunity insights and research strategy analysis [2] - Future plans include transforming the interaction interface into a central hub for various financial services, including bill payments, expense tracking, and investment management [2] Group 2: Shopping Experience - "JoyGlance" is introduced as an industry-first shopping assistant that supports cross-protocol and multi-device interactions, enabling a "see and buy" experience [2] - The assistant enhances the interaction potential of smart wearable devices, providing an immersive shopping experience where users can purchase items simply by looking at them [2] - "JoyGlance" employs a dual security system of voice and voiceprint recognition, offering a higher security level than traditional password and fingerprint payments [2] Group 3: Financing for Small and Micro Enterprises - JD Technology has developed an inclusive credit assessment intelligent agent that analyzes vast amounts of unstructured data, improving the coverage of risk indicators for previously hard-to-rate enterprises [3] - The intelligent agent can identify and assess the value of various movable collateral, expanding the range of acceptable collateral types and providing standardized valuation services for traditionally hard-to-evaluate goods [3] - The collaboration of multiple financial risk control intelligent agents has resulted in a 15% increase in the number of small and micro enterprises served and a 5% increase in average loan amounts [3]

AI在金融行业的渗透已进入到重要的奇点时刻。支付作为最高频、最贴近用户的金融场景，成为优化 体验的关键环节。9月25日，在2025京东供应链金融科技大会上，京东支付发布了其在用户体验全链路 中应用智能体的全景图，为商家和用户提供更便捷的新体验。比如使"用户问题一次性解决率"提升近两 倍、助力某银行交易额同比增长120%， 国补的成单率提升17%等…… 在核心决策层的数据分析维度，JoyData分析平台基于大模型与智能算法，构建了新一代数据分析智能 体，深度融合AI事实解读引擎、异动归因分析系统和对话式数据交互的三大特点，彻底解决了数据定 位难、解读成本高、归因链条复杂等痛点。目前，该智能体已覆盖60%以上的决策场景，使数据查询与 可视化效率提升超50%，分析报告生成与归因解读耗时压缩至1分钟内，显著提升了决策质量与响应速 度。在风控维度，京东支付构建了从"态势感知"到"风险分析"，再到"归因定位"、"策略推荐"、"线上验 证"的全链路智能风控决策。创新引入支付风控AI智能体，通过对金融黑灰产团伙的信源监测、行为溯 源、风险挖掘和策略迭代，提升了风险对抗效能，重塑了原本依赖人工的长链路对抗模式，比如使得风 险发 ...

Market Overview - The A-share market showed mixed performance with the Shanghai Composite Index down by 0.01% and the Shenzhen Component Index up by 0.67% [1] - The total trading volume across Shanghai and Shenzhen markets reached 23,918 billion yuan, an increase of 446 billion yuan compared to the previous day [1] Sector Performance - AI concept stocks resumed their upward trend, with CPO, liquid cooling, and intelligent body concepts leading the gains [1] - The gaming, controllable nuclear fusion, copper metal, and film and television sectors also saw increases [1] - The semiconductor industry chain experienced a general pullback, while gold, oil and gas, and real estate sectors remained sluggish [1] Index Performance - The Shanghai Composite Index closed at 3,853.30, down 0.01% [2] - The Shenzhen Component Index closed at 13,445.90, up 0.67% [2] - The ChiNext Index rose by 1.58%, closing at 3,235.76 [2] ETF Performance - The cloud computing sector led the ETF market with significant gains, including the Xinhua Fund Cloud 50 ETF rising by 4.02% [3] - The Huabao Fund Big Data Industry ETF and the CMB Cloud Computing ETF both increased by 3.6% [3] - The semiconductor equipment sector saw a decline, with the semiconductor equipment ETF dropping by 1.42% [3]

Core Insights - Alibaba's CEO, Wu Yongming, asserts that AI will become the next generation operating system, with a focus on advancing towards Artificial Superintelligence (ASI) [1] - The market reacted positively to these statements, with Alibaba's stock rising over 6%, reaching its highest point since October 2021 [1] Group 1: AI Integration in Automotive Industry - Several automotive companies, including Li Auto, BYD, and NIO, have introduced AI agents into their smart cockpits, enabling features like voice-activated food ordering while driving [2] - The initial applications of these AI agents are relatively simple, focusing on navigation, food ordering, and ride-hailing, but the ultimate goal is to create a seamless "human-vehicle-life" interaction [2][3] - Li Auto's AI agent, "Li Xiang," aims to enhance its capabilities with environmental awareness and comprehensive memory, allowing for more complex interactions [2] Group 2: Technical Frameworks for AI Agents - Li Auto employs two frameworks for its AI agent: CUA (Cockpit Using Agent) and MCP/A2A (Multi-Channel Processing/Agent-to-Agent) [2][3] - CUA involves multi-modal large model understanding tasks and executing them through apps, while MCP/A2A allows the AI agent to delegate tasks to third-party agents for efficiency [3][4] - The accuracy of current AI agents in completing complex tasks is around 30%, indicating a need for improved predictive capabilities [3] Group 3: Future Developments in AI Capabilities - Li Auto's vision for its AI agent includes "full information memory," which encompasses user actions, environmental interactions, and semantic memory regarding relationships [5] - The AI agent is expected to not only remember user behaviors but also proactively assist by mimicking past actions, enhancing user experience [5] - Environmental perception is crucial for the AI agent, enabling it to recognize real-world cues and complete tasks autonomously [5][6] Group 4: Industry Perspectives on AI - Wu Yongming emphasizes that for AI to surpass human capabilities, it must continuously interact with the physical world to gather comprehensive data [6] - The advancement of autonomous driving technology is cited as an example of how AI learns from raw data to improve performance [6]

Core Viewpoint - The article emphasizes that advanced driver assistance systems (ADAS) should be provided without additional usage fees, as exemplified by the Li Auto i6, which will come standard with AD Max and have no usage fees, allowing more consumers to experience the benefits of artificial intelligence in vehicles [1][2]. Group 1: Electric Vehicle Market - The first half of the electric vehicle market focused on electrification, which has now been widely adopted, providing consumers with benefits such as quieter operation, environmental friendliness, lower costs, faster acceleration, and reduced maintenance costs [1]. - The key benefit of smart technology currently lies in advanced driver assistance, with future potential in autonomous driving. However, many current ADAS require additional fees, limiting consumer experience with the latest smart technologies [1]. Group 2: AD Max System - Li Auto's AD Max system is positioned as one of the top performers in the market, with the company asserting that it is either first or second in the rankings, with ongoing improvements expected to enhance its recognition [2]. Group 3: Pricing and Business Models - Discussions around pricing for AI agents suggest that costs should be significantly lower than hiring a human driver, with estimates ranging from 2,000 to 3,000 yuan per month compared to 10,000 yuan for a human driver. This could lead to a more cost-effective solution when considering insurance and charging costs [3]. - The debate on whether to charge for advanced driver assistance hinges on whether it is considered a basic need for smart vehicles. Some argue that if it becomes a productivity tool, a subscription model could be justified [4][5]. Group 4: Competitive Landscape - The establishment of a new product's business model is likely to follow the formula of user value versus payment value, considering the incremental value of new experiences over old ones and the costs associated with switching [5]. - The discussion also highlights that subscription models often arise from service providers having marginal costs, while on-device models do not incur such costs, suggesting that charging for on-device AI features may not be necessary [7].

模型迭代正驱动新一轮能力跃迁。推理能力提升、智能体内生化、多模态融合、自主进化。但技术突破 必须与商业逻辑对齐——算力经济学成为核心考量：通过模型蒸馏、微调等技术将千亿参数优化为百亿 级应用，可在保留推理能力的同时降低成本；算力体系建设需兼顾硬件性能和软件管理指标，要考虑吞 吐性能、稳定性、百万tokens成本、软件管理等指标，实现总成本领先。 在落地实施方面，需把握战略定力与创新弹性的平衡。建议建立顶层规划与基层创新双轨机制：在战略 层面锚定技术路线图，对技术平台、算力建设、成熟场景需在战略级规划和攻坚；在战术层面要激活一 线创造力，鼓励分支机构围绕细分场景开展敏捷试验。最终形成"主干扎实、枝叶繁茂"的智能生态，让 技术创新真正转化为业务护城河。一线网点人员用大模型做科创贷款获客分析，企业客户拜访谈参等着 用"超级有用"的小场景。在智能催收领域，大模型可理解客户意图与情绪，实现短期催收的自动化，效 率接近人工水平。 经济观察网 9月24日，百度智能云金融业务部副总经理梁俊锋在"2025中国普惠金融国际论坛"上表示， 大模型在金融领域已形成可复制的落地范式。对内场景逐步完成全岗位覆盖，依托强推理和多模态能 力 ...

9月24日，"罗永浩的十字路口"账号发布罗永浩与360创始人周鸿祎对谈的录播视频。对话中，双方谈及对企业家IP与网红的观点，对Manus爆火的看法，AI 对人类社会发展的影响，以及与互联网大厂的关系缓和。 对于网红群体，周鸿祎表示，第一代网红为普通人提供了传统路径外的上升通道，核心以带货、卖课等直接变现为主。而企业家网红以俞敏洪、雷军等人为 代表，核心目的并非销售消费品，而是为企业做宣传，相当于"新一代的市场部和公关部"，通过自身影响力向社会传递企业价值。 谈及AI与Agent（智能体）时，周鸿祎表示AI 整体进化速度远超预期，但 AGI（通用人工智能）短期内不会到来。另外，相较于单一大模型，他认为智能体 才是 AI 的核心进化方向 —— 智能体能实现 "目标驱动 + 工具使用 + 推理决策"，多智能体协作可实现 "1+1>2" 的效果，类似人类社会组织的协同模式。 Agent案例中，两人谈及此前爆火的Manus。周鸿祎透露，之前因为Manus母公司Monica主做插件，还有些"看不上"。但后来Manus面世后，"我也觉得有点差 异。"他称，虽然Manus没有做基座模型，但它给行业探索出一条路：可以通过智能体 ...