彭博社近日报道指出，在全球人工智能（AI）竞赛全面加速的背景下，"能源"正在成为左右竞争格局的关 键因素。报道援引汇丰控股集团首席可持续发展官朱利安·温策尔（Julian Wentzel）分析称，中国凭借在清 洁能源领域的领先布局，在支撑人工智能高速发展的能源体系建设上，具备了"其他国家难以复制的独特 优势"。 △彭博社网站报道截图 可再生能源的结构性红利 报道指出，中国风电、光伏装机规模持续刷新纪录，清洁能源的大规模推广不仅重塑了能源结构，更持续 压低全社会的长期用能成本。 报道援引温策尔的观点称，一旦可再生能源基础设施建成，其核心成本主要来自前期投入，而新增发电几 乎不再增加支出；相较之下，化石能源则需持续承担开采、运输、炼制及分销的成本。 编译 | 李修莉 签审 | 贾延宁 蔡耀远 温策尔还表示，当基础设施的架构搭建好后，后续新增需求可接近零成本满足需求。他强调："随着新增 电力的单位成本占GDP的比例越来越低"，这将成为"拉动经济增长潜能的强大杠杆"。 美国的算力瓶颈：从芯片转向电力不足 彭博社指出，美国能源部官员认为，快速推进能源转型将推高成本、拖累经济成长。但现实层面，美国人 工智能产业的能源供给 ...

Core Insights - The 27th China International High-tech Achievements Fair is being held in Shenzhen, featuring over 5,000 renowned companies and international industry organizations from more than 100 countries and regions, showcasing cutting-edge technologies, products, and achievements [1] Group 1: Drone Technology - The newly launched "Agile Bee II Fireworks Edition Drone" by Yifei Zhikong has significantly increased its payload capacity from 2 to 6 fireworks, allowing for more efficient performances with fewer drones, thus reducing costs and enhancing visual effects [3] - The modular architecture of the drone has been upgraded with high compatibility interface designs, enabling flexible switching between various components like fireworks and Kongming lanterns, catering to diverse scenarios such as celebrations and cultural tourism [3] Group 2: Artificial Intelligence - Artificial intelligence, particularly AI agents and digital employees, is a focal point at the fair, driving digital productivity into a new phase [3] - The B2B Agent from Tanjin Technology integrates sales, marketing, outreach, and analysis capabilities, helping B2B companies enhance customer acquisition efficiency and conversion rates [4] - Tanjin Technology has also launched the overseas version of "Futern," a global sales AI agent that leverages global business data and localized insights to assist Chinese manufacturers in overcoming challenges in overseas markets [5] Group 3: Industrial Intelligence - Simo Technology, an AI unicorn nurtured by the Shenzhen-Hong Kong collaboration, has developed over 60 original industrial intelligent devices, achieving large-scale production of industrial-grade embodied intelligence and serving 300 leading companies globally [5] - The IndustryGPT, developed by Simo Technology, is the world's first industrial multimodal large model, encompassing various functionalities such as AIGC defect generation and multimodal defect detection, with its application range and influence continuously expanding [5]

Core Insights - The platform revenue is the main growth driver, achieving profitability for the first time in a single quarter. The company reported a revenue of 4.402 billion yuan for the first three quarters, a year-on-year increase of 36.8% [1] - In Q3 2025, the revenue reached 1.776 billion yuan, up 31.4% year-on-year. The "Prophet" platform continued to show high growth, generating 1.543 billion yuan in revenue for Q3 2025, a year-on-year increase of 67.6% [1] - The number of benchmark customers increased to 103, with a significant rise in average revenue per user (ARPU) to 25.49 million yuan, up 71.4% year-on-year [1] Revenue Breakdown - The "Prophet" platform accounted for 86.9% of total revenue in Q3 2025, compared to 68.1% in the same period last year [1] - The SHIFT intelligent solutions generated 231 million yuan in revenue for Q3 2025, a slight decline of 1.7% year-on-year, while AIGS revenue plummeted to 2 million yuan, down 99.1% year-on-year [1] R&D Efficiency - R&D expenses for the first three quarters amounted to 1.489 billion yuan, an increase of 8.4% year-on-year, with an R&D expense ratio of 33.8%, down 8.9 percentage points year-on-year [1] - Gross profit for the first three quarters was 1.621 billion yuan, reflecting a year-on-year increase of 20.1%, with a stable gross margin of 36.8% [1] Product Development - The company launched ModelHub XC and AI Engine EngineX on September 22, enhancing the synergy between domestic AI software and hardware. This supports multi-architecture and multi-model batch adaptation [2] - The company has made progress in AI applications in energy storage and stablecoin development [2] Investment Outlook - As an enterprise-level AI service provider, the company is positioned to benefit from top-down policy trends, with the trend of enterprise intelligence expected to accelerate over the next decade [2] - The company's enhanced adaptability among various AI infrastructure is expected to facilitate further commercialization [2] - The target price is maintained at 72.80 HKD with a "Buy" rating [2]

当全民目光聚焦于"双11"电商盛宴之际，科技领域迎来一记重磅发布。 在成本方面，火山引擎通过技术创新和全量透明Cache能力，经由火山方舟平台提供的API服务，使得 豆包编程模型的实际使用成本相较行业水平大幅降低62.7%。综合来看，Doubao-Seed-Code输入输出单 价已达国内低价，配合全量透明Cache能力，在多轮对话中进一步降低成本。 举例来说，创建一个美观的交互式英语学习网站，相同tokens量下，在0-32k输入区间，Claude Sonnet 4.5成本约4.05元，GLM-4.6约0.77元，而Doubao-Seed-Code仅约0.34元。 此外，对个人开发者，火山引擎推出Coding Plan订阅服务，支持Claude Code、veCLI、Cursor、Cline、 Codex CLI等主流编程工具，最低9.9元即可享受首月服务。 11月11日，火山引擎发布豆包编程模型 Doubao-Seed-Code。这是一款专为智能体编程任务深度优化而 生的编程模型，其集成开发环境TRAE CN企业版同日也开放公测，为企业级开发场景带来更强大的AI 编程体验。 此举不仅标志着AI编程工具赛道迎 ...

大家好，今天犀哥这篇财经评论，主要来聊聊中美GDP。美国29.2万亿美元的GDP，真的能代表它的真实经济实力吗？ 美国GDP的计算方法 美国29.2万亿美元的GDP水分不小，不是说这个数字是瞎编的，而是它的统计逻辑，本身就给"数字膨胀"留足了空间。 美国用的是支出法核算GDP，简单说就是不管这笔钱花得有没有创造实际价值，只要有资金流动，就往GDP里算。 比如叫一次1000美元的救护车，GDP就涨1000美元；就连特朗普起诉案里2.3亿美元的赔偿款，也得算进去。 诺贝尔经济学奖得主保罗・克鲁格曼早就说过，美国GDP增长越来越依赖统计技巧，而不是生产力的真提升。 这种统计方式下，美国GDP确实长得快，2019年还是21.5万亿美元，2024年就冲到29.2万亿美元，5年涨了7.7万亿美元，增幅35.8%，相当于每年多出来一个 西班牙的经济体量。 但问题是，支撑经济的"硬指标"根本没跟上，用电量绝对是经济活动的"照妖镜"，不管是工厂生产、商场营业，还是日常服务，哪样离得开电？ 过去5年，美国总发电量才从4.1万亿度涨到4.3万亿度，只多了2000亿度，增幅4.9%，还不到GDP增幅的七分之一，这说不过去啊。 美国的 ...

Core Insights - The article discusses the emergence of AI-generated long-form content, particularly in the form of a cooking competition show titled "Making Six Dishes from the Ancient Mosasaur," which has garnered over 7 million views on Bilibili, indicating a growing acceptance of AI in content creation [7][45]. Group 1: AI Content Creation - The AI variety show features chefs from six countries competing with dishes made from a fictional ancient creature, showcasing the potential for AI to create engaging content that can captivate audiences [9][45]. - The creator of the show, a Bilibili user known as @ Huangpujiang Salmon, dedicated significant time and effort to produce the AI-generated content, highlighting the need for high-level AI skills to create quality material [14][21]. Group 2: Audience Reception - The audience's reaction to the AI show is diverse, with groups ranging from those who enjoy finding flaws in AI-generated content to those who are surprised by the quality and creativity of the production [17][19]. - A significant portion of viewers, over 90%, expressed amazement at the quality of the AI-generated show, indicating a shift in perception towards AI content [17]. Group 3: Production Techniques - The creator utilized a collaborative approach between human creativity and AI capabilities, emphasizing the importance of maintaining the creator's vision while leveraging AI for execution [25][26]. - The production involved writing approximately 200,000 prompts to guide the AI in generating content, demonstrating the complexity and effort required to achieve high-quality results [20][28]. Group 4: Industry Trends - The article notes a trend towards AI-generated content in various formats, including music and podcasts, suggesting a potential explosion of AI content across platforms like Bilibili and Kuaishou [51]. - The increasing availability of AI tools is raising the bar for content creators, necessitating a blend of technical skills and creative vision to succeed in the evolving landscape of AI-generated media [47][48].

记者丨雷晨 编辑丨张伟贤 近期，月之暗面三位创始人杨植麟、周昕宇和吴育昕同时现身Reddit论坛，进行了一场长达数小时的线上问答活动。 这场技术极客风格的互动，举办在Kimi K2 Thinking模型发布第五天。 该模型在"人类最后的考试"等多项基准测试中表现超越GPT-5，引发全 球AI社区关注。 在问答过程中，创始人团队不仅回应了训练成本、算力优化、开源策略等核心问题，更直面海外用户对"中国LLM"的使用疑虑。 多项测评成绩领先 11月6日晚，月之暗面推出Kimi K2 Thinking模型，并称其为"Kimi迄今能力最强的开源思考模型"。 这一模型基于"模型即Agent"理念训练，原生掌握"边思考、边使用工具"的能力。在多项权威基准测试中，K2 Thinking达到SOTA水平。 譬如， 在被称为"人类最后的考试"的HLE（Humanity's Last Exam）测试中，K2 Thinking获得了44.9%的成绩，超过GPT-5的41.7%。在 自主网络浏览能力BrowseComp基准测试中，Kimi K2 Thinking同样以60.2%的得分，领先GPT-5的54.9%。同时在复杂信息收集推理 ...

记者丨雷晨 编辑丨张伟贤 近期，月之暗面三位创始人杨植麟、周昕宇和吴育昕同时现身Reddit论坛，进行了一场长达数小时的线 上问答活动。 这场技术极客风格的互动，举办在Kimi K2 Thinking模型发布第五天。该模型在"人类最后的考试"等多 项基准测试中表现超越GPT-5，引发全球AI社区关注。 在问答过程中，创始人团队不仅回应了训练成本、算力优化、开源策略等核心问题，更直面海外用户 对"中国LLM"的使用疑虑。 多项测评成绩领先 11月6日晚，月之暗面推出Kimi K2 Thinking模型，并称其为"Kimi迄今能力最强的开源思考模型"。 这一模型基于"模型即Agent"理念训练，原生掌握"边思考、边使用工具"的能力。在多项权威基准测试 中，K2 Thinking达到SOTA水平。 譬如，在被称为"人类最后的考试"的HLE（Humanity's Last Exam）测试中，K2 Thinking获得了 44.9%的成绩，超过GPT-5的41.7%。在自主网络浏览能力BrowseComp基准测试中，Kimi K2 Thinking 同样以60.2%的得分，领先GPT-5的54.9%。同时在复杂信息收集 ...

机器之心报道 编辑：泽南、杨文 现在，只需要一个简单的、用深度光线表示训练的 Transformer 就行了。 这项研究证明了，如今大多数 3D 视觉研究都存在过度设计的问题。 本周五，AI 社区最热门的话题是一篇新论文，有关 3D 建模的。 经过一年多的探索，来自字节跳动的团队推出了 Depth Anything 3（DA3），将单目深度估计扩展到了任何 视角场景，让计算机实现了媲美人类的空间感知。 为了追求最小建模，DA3 的工作获得了 两个关键见解 ： 就是这样的方法， 在姿态估计方面比当前业界最先进的方法 (SOTA) 提升了 44%，在几何估计方面提升了 25%。 原来 3D 视觉竟然这么简单？ 纽约大学计算机科学助理教授、知名 AI 学者谢赛宁表示，论文有点像电影：第一部通常是最好的，续集往 往更复杂却并不更精彩。但这完全不适用于 DepthAnything 系列。 Bingyikang 的团队每次都能让事情变得更 简单、更易于扩展。 论文：https://arxiv.org/abs/2511.10647 项目页面：https://depth-anything-3.github.io 代码：htt ...

Core Insights - The article discusses the emergence of new AI startups at the Web Summit, highlighting the potential for innovation in various countries beyond traditional tech hubs like Silicon Valley and Beijing [2][3][7]. Group 1: AI Startup Landscape - The Web Summit has become a platform for showcasing AI startups from diverse regions, including Brazil, Turkey, and Poland, which are skipping traditional SaaS models to enter the AI-native era [5][8]. - Many startups are focusing on niche problems within their local markets, demonstrating a trend where AI applications are tailored to specific user needs [5][9]. - The rise of AI is evident in sectors like sales and programming, with AI agents becoming increasingly specialized and integrated into business processes [5][28]. Group 2: Global AI Adoption - Smaller countries are leveraging AI to revitalize their economies, with examples like Lovable in Sweden achieving rapid growth and user adoption [7][8]. - The article notes that many startups are self-funded and not as reliant on external financing, indicating a different approach to growth compared to their counterparts in the US and China [5][9]. - AI is reshaping organizational structures, with many startups operating with minimal full-time staff and relying on freelance talent [17][19]. Group 3: Innovative Applications - Several startups are developing unique AI applications, such as Hablla in Brazil, which utilizes WhatsApp for personalized marketing, and Lyway in South Korea, focusing on AI-driven language testing [14][15]. - The article highlights the importance of product understanding and user engagement in the AI space, as seen with companies like Headway, which has successfully gamified learning experiences [24][25]. - AI applications are increasingly being designed to address specific industry needs, such as manufacturing and sales, showcasing the versatility of AI technology [28].