" 你会关心一个电饭锅能做多少种不同的饭菜，而不是单纯关注炉 子的好坏。 " 作者丨周蕾 编辑丨包永刚 阻碍金融机构把Agent从演示PPT推向核心业务场景的，究竟是什么？是算力成本，是模型能力，抑或是 一个万能的开发平台？ 在与恒生电子首席科学家白硕的深度对话中，我们得到了一个不太常见的答案：以上都不是最要紧的。 白硕早年间在中科院计算所从事前沿研究，后长期担任上海证券交易所总工程师，主导核心交易系统升 级，如今作为恒生电子首席科学家，推动AI技术落地。在经过学术前沿、行业监管核心与产业实践这一完 整路径之后，他对当下最热门的Agent话题，给出了具有历史纵深感的、颇具穿透力的洞察。 他指出，缺乏足够"厚度"的业务接口——这里并非指底层技术的API，而是指封装了业务逻辑、能"听 懂"业务人员自然语言指令的能力单元——直接导致现在许多Agent项目陷入"读不懂"真实业务需求当中的 复杂意图，无法解读有业务语义的自然语言的指令，或者只能对原有系统做简单粗暴的封装。他风趣地提 到：你会关心一个电饭锅能支持多少种花式菜谱，至于底下加热组件好不好用，会是你关注的重点吗？ 而目前通用型Agent平台的价值，其在整体解决 ...

Group 1 - GWM has appointed Zhao Yongpo, the general manager of Haval, as the new CEO of Weipai, marking the eighth CEO change in nine years, indicating a strategic shift towards internal talent over external hires [5][6] - Under the previous CEO, Feng Fuzhi, Weipai's sales were approximately 89,000 units in the first eleven months of 2025, significantly lower than competitors like Geely's Lynk & Co, which sold over 300,000 units [6] - Xiaomi has distributed over 100 million yuan in subsidies to its automotive dealers, with new store openings receiving up to 500,000 yuan each, aimed at boosting dealer morale and operations [7] Group 2 - DingTalk is reportedly launching a secret project called "D Plan," which may involve entering the hardware market with a product similar to the "Doubao phone" [9][10] - Chasing Technology has released the world's first AI health glasses, capable of monitoring heart rate, blood oxygen, and body temperature continuously [11] - ByteDance has announced significant salary increases for employees, with performance bonuses for top performers being raised by 1.5 to 3 months, reflecting a commitment to talent retention [13][14] Group 3 - MiniMax has passed the listing hearing with over $1 billion in cash reserves, supported by major investors like Alibaba and Tencent, and reported a revenue growth of over 170% year-on-year for the first nine months of 2025 [15][16] - The first L3 autonomous driving license plate has been issued to Changan Automobile, marking a significant milestone in the development of autonomous driving technology in China [28] - Zhiyun Technology has initiated its IPO process, aiming to become a leading player in the AI model sector, with substantial revenue growth projected for the coming years [26][27]

Core Viewpoint - Noitom Robotics has completed a Pre-A+ round of financing, raising several hundred million RMB, with plans to enhance its data solutions for humanoid robots and accelerate its development in the industry [2]. Group 1: Financing and Investment - The recent financing round was led by Qiming Venture Partners, with participation from various institutions including Wuyuan Capital and Junlian Capital, and saw oversubscription [2]. - The funds will be primarily used for research and development in multi-modal data collection, processing, and delivery technologies, as well as to build a scalable data production system and engineering platform [2]. Group 2: Company Overview and Team - Noitom Robotics focuses on providing high-quality, scalable training data and infrastructure capabilities for the humanoid robot and embodied intelligence industries [2]. - The company is led by Dr. Dai Ruoli, a co-founder of Noitom Ltd., who has over 15 years of experience in motion capture and human-computer interaction [3]. - Dr. Han Lei, the Chief Scientist, has a strong background in robotics and reinforcement learning, previously leading Tencent's Robotics X Lab [3]. Group 3: Data Collection Strategies - Noitom Robotics has developed a dual approach to data collection: In-the-factory and In-the-wild, ensuring high-dimensional, multi-modal, and high-precision data [4]. - The In-the-factory method focuses on collecting data that surpasses the dimensions and modalities of robots, while the In-the-wild method captures natural human behaviors in real-world scenarios to enhance model training [4]. Group 4: Industry Collaboration - The company has established deep collaborations with around 60 to 70 leading companies in the robotics field, covering various aspects such as data collection equipment and customized data sets [4]. - Noitom Robotics' data path system has been validated through practical delivery to numerous humanoid robot enterprises and embodied intelligence model clients globally [4].

Core Viewpoint - The article discusses the current state and future potential of embodied intelligence, particularly focusing on the challenges and opportunities in data collection methods and the maturity of the industry [2][25]. Data Quality and Collection - High-quality data is becoming a bottleneck for breakthroughs in embodied intelligence performance and cost control [2]. - The roundtable forum at the GAIR conference emphasized the importance of data quality, with consensus that the effectiveness of models and the benefits to robots depend on the quality of data collected [2][4]. - Different data collection methods, such as UMI (Universal Manipulation Interface), remote operation, motion capture, and simulation data, are being explored, with a focus on adapting to various scenarios and hardware [3][4]. Industry Maturity and Challenges - The data collection industry is still in its early stages, with companies in data, embodiment, and modeling still aligning their needs and capabilities [4][7]. - There is a lack of a unified approach in the industry, with varying demands from different model companies, indicating that data companies need to understand models and provide suggestions to improve collaboration [7][8]. - The current focus on data collection methods is shifting, with a notable rise in interest in remote operation and UMI, particularly influenced by developments in North America [9][10]. In-the-Wild Data Collection - In-the-wild data collection is seen as a challenging yet promising approach, requiring advanced technical capabilities and effective management of hardware and software [3][21]. - The article highlights the need for low-friction, high-precision, and multi-modal data collection devices to effectively utilize in-the-wild data [3][21]. - The maturity of in-the-wild data collection is still developing, with current efforts primarily focused on improving data collection technology before addressing human resource management [21][22]. Government Support and Industry Dynamics - Government support for data collection factories is prevalent in China, which may influence the direction of data collection methods and industry growth [10][17]. - The article suggests that while government-backed data collection initiatives can stimulate the industry, they may not always align with the most effective technological advancements [17][18]. - The cost structure of data collection is critical, with significant portions attributed to equipment depreciation and labor costs, indicating a need for strategic investment in data collection methods [19][20]. Future Outlook - The industry is expected to evolve, with a potential shift towards more diverse data collection methods as companies adapt to changing demands and technological advancements [18][19]. - The article expresses skepticism about the current maturity of the embodied intelligence industry compared to other tech sectors, suggesting that significant challenges remain before widespread adoption can occur [25][26]. - Companies are encouraged to collaborate and share insights to enhance data collection processes and improve overall industry knowledge [28][30].

Core Viewpoint - The article discusses the challenges and transformations in the To B (business-to-business) sector, emphasizing the need for a new partnership model between service providers (乙方) and clients (甲方) that focuses on shared risks and benefits, rather than traditional transactional relationships [2][6][28]. Group 1: Traditional To B Challenges - The To B sector has been characterized by a "double loss" situation where clients demand high-quality services at low costs, leading to dissatisfaction on both sides [4][5]. - The traditional procurement logic of clients is based on a bidding mentality, which pressures service providers to deliver under tight budgets, resulting in a cycle of dissatisfaction and inefficiency [5][6]. - The industry faces dilemmas such as the conflict between standardization and customization, where service providers often compromise on product standardization to meet individual client needs, leading to inefficiencies [5][6]. Group 2: New Business Model - RaaS - The article introduces the concept of "Results as a Service" (RaaS), which aims to align the income of service providers with the quantifiable business value they create for clients [6][12]. - 百融云创 has initiated the RaaS model, transitioning from traditional software sales to a partnership approach where both parties share risks and rewards [6][12]. - The RaaS model includes innovative pricing strategies, such as charging based on the roles of digital employees (硅基员工) rather than traditional software sales, allowing for a more flexible and performance-based payment structure [12][13]. Group 3: Engineering and Technological Support - The "Results Cloud" platform is introduced as the technological backbone supporting the RaaS model, enabling efficient management and operation of digital employees [15][16]. - The platform integrates various AI capabilities and models to ensure that digital employees can deliver reliable and measurable outcomes [16][20]. - Key engineering capabilities include a unified connection protocol, high-precision document parsing, and a robust DevOps system to ensure the quality and continuous improvement of AI agents [19][21][22]. Group 4: Collaboration Between Human and Digital Employees - The article highlights the potential for collaboration between human employees (碳基员工) and digital employees, suggesting a model where both can coexist and complement each other rather than one replacing the other [24][25]. - The focus is on redefining workflows to optimize efficiency and quality around shared business outcomes, moving away from traditional cost-cutting negotiations [27][28]. - The ultimate goal is to create a new production relationship that allows for fair distribution of the benefits of technological advancements among creators, users, and enablers [29][30].

Group 1 - JD's private label business recently experienced significant personnel changes, with the departure of a new executive from Sam's Club and the reassignment of former head Tang Hengsheng to JD Industrial [2][3] - The introduction of the new executive was reportedly due to dissatisfaction with the overall performance of the private label business under Tang Hengsheng's management [3] - JD's private label initiative began in 2015, launching brands like "Jiabai" and "Jingzao," and has since expanded to include various product lines targeting different market segments [3][4] Group 2 - JD is consolidating its strategy by focusing resources on its core brands, Jingzao and Huixun, while eliminating underperforming categories [4] - The private label division has reportedly been disbanded, with its business lines integrated into other related business groups, although this information has not been confirmed [4] - JD Industrial is also entering the private label space, launching sub-brands aimed at industrial products, which may explain Tang Hengsheng's transfer [4]

Core Viewpoint - The article discusses the current state and future potential of embodied intelligence, focusing on the challenges and opportunities presented by world models and spatial intelligence in the field of robotics and AI [2][4][10]. Group 1: Development of Embodied Intelligence - The technology route for embodied intelligence is still in an exploratory phase, with no convergence yet, which is seen as a positive sign for innovation [4][3]. - There is a consensus among experts that the core issues of embodied intelligence, such as interaction and human-machine collaboration, should be addressed by academic institutions, while industries focus on practical applications [4][5]. - The integration of AI with physical entities is expected to lead to significant advancements in intelligence, but the field must avoid reverting to industrial automation without achieving generalized intelligence [4][5][30]. Group 2: World Models in Autonomous Driving - World models are currently being utilized by leading companies like Tesla to enhance data generation and improve decision-making processes through closed-loop testing [11][12]. - The concept of world models has gained traction in autonomous driving due to the simplicity of generating scenarios compared to robotics, with advancements in generative AI enabling the creation of realistic training samples [12][13]. - There is ongoing debate regarding the definition and application of world models in both autonomous driving and robotics, with differing opinions on the necessity of pixel-level reconstruction versus latent state representation [12][13][14]. Group 3: Spatial Intelligence in Robotics - Spatial intelligence is a critical aspect of robotics, with a focus on perception and understanding spatial relationships, which has evolved from traditional SLAM techniques to more learning-based approaches [20][21]. - The current challenges in spatial intelligence include the need for better data representation and understanding of complex spatial relationships, which are still underdeveloped in robotic systems [22][23]. - The integration of visual and semantic information is essential for enhancing robots' spatial capabilities, but the field is still in its early stages [22][23][24]. Group 4: Commercialization and Future Applications - The future of drone applications is expected to expand significantly, with potential uses in various sectors, but the timeline for widespread adoption remains uncertain [26][27]. - The gap between technological capabilities and market needs poses challenges for entrepreneurs, as there is often a mismatch between innovative ideas and practical industrial requirements [30][31]. - The shift towards learning-based control paradigms is anticipated to increase the applicability of drones and robots in real-world scenarios, moving beyond traditional automation [28][29].

Core Insights - The article discusses the challenges and bubbles in the artificial intelligence (AI) industry, highlighting that 95% of AI projects are failing, with only 5% achieving success, according to a MIT survey [2][15] - The discussion emphasizes the need for realistic expectations, system integration, and data availability as critical factors for successful AI implementation [6][16][18] Group 1: Challenges in AI Industry - The AI industry faces three main challenges: expectation management, system integration, and data availability [6][16][18] - High expectations from business leaders, driven by media hype, lead to unrealistic goals and potential industry collapse [16][26] - System integration issues arise when AI technologies do not align with existing traditional systems, causing operational inefficiencies [17][18] - Data limitations hinder AI's ability to function effectively, as many applications rely solely on language models without sufficient diverse data [18][29] Group 2: Bubbles in AI - Two significant bubbles identified are in the computing power sector and the AI application sector, where many resources are underutilized or overly reliant on human input [8][30] - The computing power bubble is characterized by excessive investment in inference capabilities while lacking sufficient training infrastructure [29][30] - The AI application bubble is marked by a high degree of similarity among products, with many applications not achieving true AI capabilities [8][30] Group 3: Future Opportunities - Potential areas for investment include small models in specialized fields, which could be integrated to create comprehensive solutions [39][45] - The healthcare sector presents opportunities for AI, particularly in developing models that can work with limited data while ensuring privacy [39][42] - Safety and control in AI applications are crucial for future development, especially in sensitive industries like healthcare and finance [42][45]

Core Insights - The 3D printing brand Qingliang Zhizao has completed its angel round financing, led by Hai Xin Capital, a fund under a leading laser listed company, with Nanshan Zhanxin Investment participating [1] - Qingliang Zhizao was founded in April 2025 by Wang Zhiyu, who previously worked at Anker 3D printers and has a background in mechanical design [1] - The company aims to focus on the niche market of 3D printing for mass production, with its first product expected to launch in the European and American markets in the first half of 2026 [1] - The global consumer-grade FDM 3D printer market is approximately 50 billion yuan, with the SMB market size around 10 billion yuan, targeting this significant market [1] - The company claims to enhance the ROI for business owners to nearly 10 times the current level [1] Company Development - Wang Zhiyu emphasized the support received during his time at Anker, which broadened his perspective on the industry [2] - The funding will be invested in industrialization and technology research and development to inject more possibilities into the 3D printing industry [2]

Core Viewpoint - The article discusses the challenges and developments in spatiotemporal AI, emphasizing the need for AI to transition from virtual to physical worlds to unlock its full industrial value [3][4][39]. Group 1: Challenges of Spatiotemporal AI - Spatiotemporal AI faces three main challenges: data scarcity and high collection costs, weak modeling capabilities due to unknown physical laws, and difficulties in creating closed-loop intelligent solutions [4][8][20]. - Data in the physical world is often limited, with high costs and long collection cycles, making it difficult to gather sufficient information for effective modeling [4][9]. - Existing models do not adequately account for spatiotemporal attributes, complicating the application of AI in real-world scenarios [9][20]. Group 2: Development Stages of Spatiotemporal AI - The development of spatiotemporal AI has progressed through five stages, starting from classic models in 1960-1995 to the current focus on city-scale models [26][32][34]. - The second stage (1995-2008) involved discovering spatiotemporal patterns, leading to the application of these patterns in various scenarios, including public health [27][28]. - The third stage (2009-2016) saw the integration of classic machine learning with spatiotemporal features, significantly improving predictive accuracy in air quality monitoring [29][30]. - The fourth stage (2016-2030) introduced deep learning techniques to handle complex spatiotemporal data, particularly in urban environments [32][33]. - The current stage (2023-2035) emphasizes the need for multi-source data fusion and the development of urban intelligence systems, integrating various data types for comprehensive city management [34][35]. Group 3: Application in Smart Cities - The article highlights Xiong'an New Area as a model for smart city development, utilizing spatiotemporal AI to manage urban operations effectively [39][40]. - Real-time data analysis in Xiong'an allows for proactive management of resources, such as electricity and public safety, demonstrating the practical applications of spatiotemporal AI [39][40]. - The integration of various data types, including traffic, weather, and demographic information, is crucial for creating a responsive urban intelligence system [34][39].