Core Insights - The article discusses the evolution of large model evaluation and expectations in the AI industry, emphasizing a shift from "reasoning" to "agent-based thinking" as the next phase of AI development [2][4] - The author predicts that agent-based thinking will become mainstream by 2025, highlighting the need for models to think in a way that supports effective actions through interaction with their environment [5] Group 1: Development of AI Models - The past two years have reshaped industry expectations for large models, with OpenAI's advancements demonstrating that "thinking" can be a trained capability [2] - The focus for the first half of 2025 will be on enhancing reasoning capabilities, specifically how models can engage in deeper reasoning before arriving at conclusions [2] - A significant ambition for the Qianwen team is to create a unified system that merges thinking and instruction modes, allowing for adjustable reasoning efforts based on context [2][3] Group 2: Challenges in Merging Thinking and Instruction - The main challenge lies in the data, as merging thinking and instruction involves balancing different data distributions and behavioral goals, which has led to subpar performance in both areas [3] - The Qwen 2507 version released in late 2025 will feature separate Instruct and Thinking versions, indicating a preference for maintaining distinct functionalities [3] Group 3: Shift in Training Focus - The transition from training models to training agents signifies a change in focus towards the entire system of models interacting with their environment [4] - The definition of "good thinking" has evolved to prioritize actions that can be effectively supported under real constraints, rather than merely the length or complexity of reasoning chains [4]

Core Insights - The article discusses the evolution of large model evaluation and expectations in the AI industry, highlighting a shift from reasoning-based thinking to agent-based thinking for actionable intelligence [3][7] - The author emphasizes the challenges of merging thinking and instruction modes in AI models, noting that achieving a balance between these modes is complex and often results in subpar performance [4][5] Group 1: Development of AI Models - The past two years have reshaped industry expectations for large models, with OpenAI's advancements demonstrating that "thinking" can be trained as a capability [3] - The focus in early 2025 is expected to shift towards agent-based thinking, which involves thinking for action and continuously updating plans based on environmental feedback [3][5] Group 2: Challenges in Merging Thinking and Instruction - The primary difficulty in merging thinking and instruction lies in the significant differences in data distribution and behavioral goals between the two modes [4] - The attempt to balance model merging with improved training data quality has often led to mediocre outcomes, where thinking behaviors become noisy and instruction behaviors lack clarity [4][5] Group 3: Future of AI Thinking - The article suggests that separating the instruction and thinking processes remains appealing, as demonstrated by Qwen's independent versions of Instruct and Thinking [5] - The successful merging of these processes requires a smooth spectrum of reasoning effort, allowing models to autonomously determine the appropriate level of thought required for different tasks [5][6] Group 4: Shift in Evaluation Criteria - The evaluation of models is transitioning from assessing the duration of thinking to determining if the thinking supports effective action [6] - The definition of "good thinking" has evolved to focus on the trajectory that best supports action under real constraints, rather than simply the longest or most complex reasoning chain [6]

文：蔚蓝 编辑：宋政 阿里铺路，腾讯种树——谁能在AI时代的终点，最先等到花开？ 3月18日，阿里云突然官宣AI算力、存储产品调价，平头哥算力卡最高涨幅34%，打响国内云厂商算力定价升级第一枪；几乎同期，腾讯旗下"龙虾"AI智能体开启大 范围公测，深度嵌入微信生态，靠场景渗透抢滩落地。一边是砸钱做基建、靠算力收租，一边是稳现金流、靠生态分账，阿里腾讯两条路线彻底走向分化。 阿里巴巴：全栈押注，重构AI时代"水电煤" 财务数据与短期承压 根据阿里巴巴最新发布的2026财年第三季度（截至2025年12月31日）财报， 集团收入达2848.43亿元，同比增长2% ；若不考虑高鑫零售和银泰已处置业务，同口径 收入同比增长9%。 云智能集团本季度收入同比增长36%至432.84亿元 ， AI相关产品收入连续第十个季度实现三位数增长 。值得注意的是，平头哥自研GPU芯片首 次写入财报，截至2026年2月已累计规模化交付47万片，其中60%以上服务于外部商业化客户。本季度自由现金流为113.46亿元，同比下降71%，主要由于对即时零售 和AI基础设施的持续投入。 过去四个季度，阿里在AI+云基础设施的资本开支累计约 120 ...

Core Viewpoint - The article discusses the emergence of OpenClaw, an AI software that enhances productivity by enabling direct computer operations, and its application in the automotive industry through the launch of the "Super Intelligent Agent IM Ultra Agent" by Zhiji Auto, which aims to revolutionize user interaction with vehicles [2][6][8]. Group 1: OpenClaw and Its Capabilities - OpenClaw is an open-source AI software that can perform tasks such as organizing files, replying to emails, and making reservations, marking a shift from simple conversational AI to a more functional digital assistant [2][5][6]. - The software's ability to operate computers directly represents a significant advancement in AI tools, transitioning from mere conversation to actionable tasks [6]. Group 2: Zhiji Auto's Super Intelligent Agent - Zhiji Auto introduced the "Super Intelligent Agent IM Ultra Agent," which aims to provide comprehensive assistance by understanding user preferences and executing tasks with minimal input [8][10]. - Unlike traditional smart cabins that only support basic dialogue, the Super Intelligent Agent can handle complex tasks like ordering coffee by understanding user habits and preferences [12][15]. Group 3: Advanced Features and User Interaction - The Super Intelligent Agent can interpret nuanced user statements, such as "the child is asleep," to adjust vehicle settings for a quieter and more comfortable environment [18][21]. - It can also respond to user emotions and tone, enhancing the driving experience by acting as a digital assistant that understands and anticipates needs [22][23]. Group 4: Technical Foundations - The system is built on a new integrated architecture that merges chassis control, intelligent driving AI, and intelligent cabin AI, allowing for a seamless interaction between these components [25][26]. - Zhiji Auto employs a fully controlled chassis system with triple redundancy to ensure safety and reliability, even without mechanical connections [31][34]. Group 5: AI Models and Collaborations - The IM AD ZETA model, developed in collaboration with Momenta, enhances driving capabilities through reinforcement learning, significantly improving performance compared to existing models [38][39]. - The integration of Alibaba's ecosystem allows for seamless data flow between various applications, enabling the Super Intelligent Agent to execute tasks efficiently [45][46]. Group 6: Future Implications and Safety Measures - The introduction of AI in vehicles signifies a pivotal shift towards smarter, more interactive automobiles that serve as intelligent partners in daily life [55][56]. - Zhiji Auto emphasizes the importance of safety and data security, implementing strict controls and a "black box" mechanism to ensure user information is protected [52][53].

Core Viewpoint - Alibaba has established the Alibaba Token Hub, led by CEO Wu Yongming, focusing on the creation, delivery, and application of Tokens, marking a strategic shift towards commercializing AI applications rather than just technological exploration [3][4]. Group 1: Alibaba Token Hub Structure - The Alibaba Token Hub includes several divisions: Tongyi Laboratory, MaaS business line, Qianwen Division, Wukong Division, and AI Innovation Division, covering a comprehensive layout from foundational model research to AI applications for both individual and enterprise users [3]. - Tongyi Laboratory is tasked with creating leading multimodal models, while the MaaS business line serves as a large model service platform for enterprise clients [3]. - The Wukong Division is positioned as a "B-end AI native work platform," indicating Alibaba's intent to deepen AI integration into enterprise workflows [3][4]. Group 2: Strategic Shift and Market Positioning - The establishment of the Alibaba Token Hub reflects a shift from a focus on foundational large models to a more application-driven approach, as Alibaba aims to compete in the rapidly evolving AI landscape [5][6]. - Wu Yongming highlighted the impending explosion of AGI, with billions of AI Agents expected to support digital work, emphasizing the historical opportunity presented by the Token-based model [4]. - The departure of Lin Junyang, the former head of Qianwen's large model technology, has sparked discussions about the alignment of strategic goals within Alibaba, suggesting a need for stronger collaboration between foundational models and application ends [5].

Investment Rating - The report does not explicitly provide an investment rating for the industry Core Insights - The AI industry is experiencing strong demand, with companies like Nvidia and Marvell reporting significant revenue growth due to AI applications [8][19][20] - Nvidia's strategic investments in silicon photonics technology aim to alleviate supply chain bottlenecks and support the scaling of AI infrastructure [18][31] - OpenAI and Anthropic have achieved substantial annual revenues, indicating a robust market for AI models [14][15] Industry Frontiers - Samsung plans to continue raising NAND product prices in Q2, following a 100% increase in Q1, indicating ongoing supply constraints in the semiconductor market [12] - Nvidia has ceased production of the H200 AI chip for the Chinese market, reallocating resources to the next-generation AI computing platform [13] - OpenAI's GPT-5.4 has demonstrated advanced capabilities, surpassing human-level performance in certain tasks [14] Capital Trends - Nvidia's $20 billion investments in Coherent and Lumentum highlight a commitment to developing next-generation silicon photonics technology [18] - Companies like Broadcom, Marvell, and Ciena have reported strong financial results, reflecting the increasing demand for AI-related products [19][20][21] - ByteDance's partnership with Century Internet for a 500 MW data center underscores the growing need for AI computing power in China [23] Weekly Perspective - The report emphasizes the strong performance of companies in the AI sector, with Broadcom and Marvell both reporting significant revenue increases [31] - Ciena's focus on AI training and co-packaged optics technology is seen as a response to the limitations of traditional electrical interconnects [31] - MiniMax's impressive revenue growth indicates a successful commercialization of AI technologies, driven by improved model efficiency [31]

Investment Rating - The investment rating for the industry is "Positive" (首次) [8] Core Insights - The OpenClaw application continues to generate high demand for computing power, with significant interest from major cloud service providers in China [4][11] - The Seedance 2.0 model, launched by ByteDance, has entered the commercialization phase, which is expected to further drive the demand for computing power [5][12] - Domestic and international cloud service providers have raised prices, indicating strong downstream demand for computing resources [6][13] Summary by Sections Investment Opportunities - OpenClaw's popularity has led to a surge in computing power demand, with various cloud providers announcing support for the application [4][11] - Seedance 2.0's pricing structure has been established, with costs set at 28 RMB per million tokens for video input and 46 RMB per million tokens without video input, which is expected to increase computing demand [5][12] - Price increases from major cloud providers like AWS and Google Cloud validate the robust demand from downstream users [6][13] Market Review - The Shanghai and Shenzhen 300 Index fell by 1.07%, while the computer index dropped by 5.29% during the week of March 2 to March 6, 2026 [15] Investment Recommendations - The report suggests focusing on the AI infrastructure industry chain, recommending companies such as Deepin Technology and Haiguang Information, with beneficiaries including Borui Data, Dongfang Guoxin, and others [7][14]

Investment Rating - The report indicates a neutral investment rating for the industry, expecting a fluctuation range of -5% to 5% compared to the broader market over the next 3-6 months [55]. Core Insights - The coffee industry remains highly prosperous, with brands actively opening new stores and a slowdown in price competition. The decline in Arabica coffee futures prices is expected to improve cost structures [3]. - The tea beverage sector is facing slight pressure due to short-term competition in food delivery platforms, although data shows resilience [3]. - The e-commerce sector continues to be under pressure, with a lackluster performance attributed to the domestic consumption environment [3]. - Streaming platforms are seen as quality internet assets driven by domestic demand, benefiting from cost-effective self-consumption and scale effects [3]. - The virtual assets and trading platforms are experiencing macroeconomic volatility, with limited catalysts in the cryptocurrency market [3]. - The automotive service sector is witnessing a push from several major automotive service chains to expand their market presence, while OEMs are struggling with inventory pressures [3]. - The AI and cloud sectors are viewed positively, with major internet companies having the resources to implement AI-related business advancements [3]. Summary by Sections 1.1 Consumer & Internet - **Coffee and Tea**: The coffee sector is thriving with ongoing store openings and reduced price competition, while the tea sector is slightly pressured by delivery competition [3][15]. - **E-commerce**: The sector is underperforming due to a challenging consumption environment, as reflected in the performance of major players like JD and Alibaba [14]. 1.2 Platform & Technology - **Streaming Platforms**: The media index has outperformed the broader market, with notable performances from Spotify and Netflix [22]. - **Virtual Assets**: The global cryptocurrency market capitalization reached $2481.9 billion, with Bitcoin and Ethereum prices showing slight increases [27]. - **Automotive Services**: The automotive service sector is seeing a decline in market performance, with several companies adjusting their strategies to capture market share [36]. - **AI & Cloud**: The sector is experiencing growth, with significant stock performances from companies like Oracle and Microsoft [47].

Core Viewpoint - The departure of Lin Junyang, the technical head of Qwen, has raised discussions about his influence and the internal dynamics at Alibaba, particularly regarding the company's organizational structure and culture [1][2][5]. Group 1 - Lin Junyang announced his departure from Alibaba on March 7, expressing his support for the company and urging his colleagues to continue their efforts [2]. - His role as a key figure in the Qwen model has been contested, with some internal sources downplaying his significance and attributing the core leadership to Alibaba Cloud's CTO, Zhou Jingren [2][3]. - There have been reports of Lin challenging company protocols, particularly regarding his abrupt resignation announcement without prior communication with management, which was deemed unacceptable by some within the organization [5].

Group 1 - The departure of Lin Junyang, the technical head of Qianwen's large model, has sparked significant industry discussion, especially after the news was approved by Alibaba's CEO [1] - An Alibaba employee publicly criticized Lin Junyang on a third-party platform, accusing him of being immature and damaging the company's stock price [3] - Lin Junyang has made two public posts regarding his departure, one on X and another on WeChat, where he encouraged his team members [4] Group 2 - Several core members of Qianwen have recently been approached by leaders from major companies, indicating a high demand for their expertise in the industry [6]