Emily Prince LSEG分析与人工智能部门主管 LSEG在Workspace中推出的Deep Research正是将这一能力真正落地，它在专为市场参与者设计的专 业环境中，原生地提供了结构化、符合金融行业标准的研究能力。 Deep Research正在迅速发展。从目前为 ChatGPT的Deep Research提供支持的OpenAI GPT-5.2，到 LSEG的Workspace Deep Research，用户如今可以进行更加结构化的查询、连接经过批准的应用程 序、聚焦特定数据来源、实时引导研究过程，并以全屏报告的形式查看研究成果。 但在金融服务行业，更好的研究并不等同于真正有价值且值得信赖的研究。 真正值得信赖且有意义的研究，应当能够产出可执行的洞见以及深入的信息，使金融专业人士能够在 投资委员会、风险会议或客户沟通中使用这些内容，并能够充满信心地进行辩护。要实现这一点，需 要三个关键要素： 经过验证的可靠数据来源、受管控的访问权限，以及清晰的数据来源可追溯性 ——也就是对所使用的数据是什么、何时使用、以及在什么权限下使用保持透明。 这正是为什么LSEG正通过其MCP连接器，将其获得许可并可 ...

Group 1 - Alibaba Group is enhancing its artificial intelligence capabilities by hiring Zhou Hao from Google DeepMind to lead post-training research for its Qwen AI team [1][2] - Zhou Hao replaces Yu Bowen, who also left the company recently, while Alibaba has not yet announced a successor for Lin Junyang, the former technical lead of Qwen AI [3][4] - Lin Junyang's unexpected departure has caused significant reactions within the developer community, and he announced his exit without providing further details [4] Group 2 - Alibaba's stock has decreased nearly 5% over the past 12 months, underperforming the Nasdaq Composite Index, which gained approximately 23% during the same period [5] - The decline in Alibaba's shares is part of a broader trend affecting U.S.-listed Chinese tech stocks, which fell following Beijing's announcement of a 2026 GDP growth target of 4.5%–5%, the lowest since the early 1990s [5]

一位是姚顺雨，从清华姚班到OpenAI，再到腾讯史上最年轻的首席AI科学家。 姚顺雨，腾讯史上最年轻的首席AI科学家 另一位是姚顺宇，从清华物理系特奖得主到Anthropic核心研发，又转投谷歌DeepMind。 2026年2月3日，姚顺雨发布了加入腾讯后的首个研究成果：CL-bench。这篇论文揭露了一个尴尬的事实 ——即便给全球最强的AI模型提供完整上下文，它们的任务解决率也只有17.2%。 作者 | 闰然 编辑 | 江江 视觉 | 顾芗 近期，AI圈被一个奇妙的巧合刷屏：两位名叫"Yao Shunyu"的清华人，同时站在了全球智能革命的风暴 眼。 他们都出生于1997年。 27岁的他曾说，AI接下来比拼的不是训练，而是"如何定义并评估真正有用的任务"。这句话既是他对行 业的精准研判，亦是其自身的真实写照，而这，正是腾讯对他寄予厚望的关键所在。 而此刻33岁的月之暗面创始人杨植麟，也站在聚光灯下。那个在清华组建摇滚乐队Splay、写过《一夜暴 富白日梦》的年轻人，如今更富了——最新的K2.5发布不到一个月，Kimi近20天累计收入已超过2025年 全年总收入。 这位曾经的"投流狂魔"正在证明，在DeepS ...

Core Insights - The article emphasizes the transition from traditional chatbots to Long-horizon Agents, which are capable of performing complex tasks over extended periods, thus redefining the value proposition of AI products from speed of response to quality of output [3][8][10]. Group 1: Long-horizon Agents - Long-horizon Agents are designed to operate autonomously over longer time spans, allowing for multi-step decision-making and iterative processes, which are essential for tasks like research reports and code reviews [16][20]. - The emergence of Long-horizon Agents marks a significant shift in AI capabilities, moving from simple question-answer interactions to producing high-quality deliverables that require time and context [7][8][11]. Group 2: Harness Concept - The concept of "Harness" is introduced as a runtime environment that includes best practices for building Long-horizon Agents, distinguishing it from traditional frameworks by providing integrated tools and capabilities [11][23]. - Harnesses facilitate the development of agents that can autonomously manage tasks, including planning, memory management, and sub-task coordination, thus enhancing their effectiveness [11][23][24]. Group 3: Evolution of AI Agents - The evolution of AI Agents is categorized into three phases: simple prompting and chaining, cognitive architecture, and the current Long-horizon Agent era, which began around mid-2025 [26][30][31]. - The transition to Long-horizon Agents is characterized by improved model capabilities and a focus on context engineering, which is crucial for optimizing agent performance [29][34]. Group 4: Applications and Future Directions - Long-horizon Agents are particularly effective in generating initial drafts for various applications, such as coding, research, and customer support, where they can significantly reduce the workload for human users [20][22]. - The future of AI development is expected to focus on enhancing context engineering, memory management, and the integration of file systems, which are seen as critical components for the success of Long-horizon Agents [34][42][46].

Group 1 - The article outlines a systematic approach to conducting a competitive analysis of the Deep Research feature, emphasizing the importance of strategic insights and actionable recommendations [1][2][21] - The core function of Deep Research, launched by OpenAI in February 2025, allows AI to autonomously conduct web searches, integrate information from multiple sources, and generate comprehensive research reports, distinguishing it from traditional AI search methods [5][6] - Key dimensions for analysis include core functionality, feature matrix, and content quality, while market positioning and model technology are considered less critical for this specific inquiry [8][9] Group 2 - The selection of competitors includes direct competitors with independent Deep Research capabilities, indirect competitors with research abilities, and potential competitors that may emerge in the future [10] - Data collection involves three main channels: public information retrieval, product experience through testing, and user research to gather real user feedback [11][12] - The analysis phase includes constructing a feature matrix to compare functionalities across competitors and evaluating content quality based on accuracy, completeness, depth, structure, and usability [14][15][16] Group 3 - The report structure is designed to present core conclusions upfront, followed by an overview of competitors, a feature comparison matrix, content quality assessment results, and typical case studies to illustrate findings [17][18][19][20] - The final section provides actionable recommendations, prioritizing features to follow up on and identifying potential pitfalls to avoid [21][22] - The overall process of competitive evaluation is summarized as a series of methodical steps: clarifying objectives, selecting appropriate competitors, defining dimensions, collecting data, analyzing findings, and producing the report [21]

Core Insights - The article emphasizes that competitive product analysis has become an essential skill for AI product managers due to the rapid iteration and evolution of AI products [1][3] - It highlights the unique characteristics of the AI industry that necessitate continuous competitive analysis to maintain market relevance and identify differentiation opportunities [3][4][5] Importance of Competitive Analysis - AI products evolve at a much faster pace than traditional software, making previous conclusions potentially obsolete within weeks [3] - The AI market is crowded yet distinctly segmented, requiring thorough research to identify unique positioning and differentiation strategies [4] - The low switching costs for users mean that they can easily migrate to competitors if new features are introduced, underscoring the need for constant vigilance in competitive analysis [5] Value of Competitive Analysis - The first layer of value is understanding the market landscape and identifying the company's position among competitors [6][7] - The second layer involves recognizing strengths and weaknesses to guide product iterations, ensuring that the company meets or exceeds industry standards [8] - The third layer aids in strategic decision-making by predicting competitive risks and industry trends, which is crucial for resource allocation and strategic planning [9][10] Tailoring Reports for Different Roles - Different stakeholders require different insights from competitive analysis reports, necessitating tailored presentations for executives, product teams, development teams, and marketing teams [9][10] - Executives focus on strategic positioning and opportunities, while product teams seek actionable insights for feature improvements [9] - Development teams look for technical implementation details, and marketing teams are interested in pricing strategies and promotional channels [10] Conclusion - The core purpose of competitive analysis is to gain a comprehensive understanding of the market landscape, identify product strengths and weaknesses, and formulate precise product strategies and differentiation [11]

Core Insights - Meta announced the acquisition of Manus, an AI Agent startup, for $2 billion, highlighting its urgent need for a capable team and technology in the AI space [1][4][8] - The acquisition reflects a broader trend among tech giants to address anxiety over AI capabilities and market competition, with many companies resorting to buying talent and technology [9][10] Group 1: Acquisition Details - Manus, founded by Chinese entrepreneurs, achieved an annualized revenue of $125 million within eight months of its product launch [1] - The deal was characterized by a quick negotiation process, with Meta's CEO Mark Zuckerberg agreeing to the founder's asking price without hesitation [1] - Following the announcement, Meta's stock price fell for two consecutive trading days, indicating skepticism from the market regarding the potential impact of the acquisition [1] Group 2: Meta's Challenges - Meta's previous AI model, Llama 4, faced significant criticism for underperformance despite initial high rankings, leading to concerns about the company's AI capabilities [2][3] - The company has struggled to produce a competitive foundational model, while rivals like Anthropic and Google continue to excel in the AI space [3][11] - Meta's revenue model is heavily reliant on advertising, which is threatened by the rise of AI Agents that change user engagement dynamics [5][11] Group 3: Market Dynamics - The acquisition of Manus is seen as a strategic move to mitigate Meta's vulnerabilities in AI, as the company faces competition from both established players and emerging startups [6][9] - Manus's reliance on third-party models for its product experience introduces risks related to cost variability and supply chain stability [5][6] - The acquisition reflects a broader pattern of tech companies seeking to secure their positions in the rapidly evolving AI landscape, often driven by fear of falling behind [9][10] Group 4: Future Considerations - The integration of Manus into Meta's ecosystem could provide opportunities for deeper product integration across platforms like Facebook and Instagram [7] - However, concerns remain about whether the acquisition will effectively address Meta's underlying issues, particularly regarding organizational culture and integration challenges [14][15] - Historical examples of successful acquisitions in the tech industry suggest that simply buying technology may not resolve deeper organizational deficiencies [12][16][18]

Core Viewpoint - The article discusses the perception of job-hopping among high-value talent versus ordinary workers, highlighting that the former is often viewed positively while the latter may face negative judgments regarding stability and capability [5][6][7]. Group 1 - High-value talent, such as former OpenAI engineer Yao Shunyu, is seen as making career moves that reflect ambition and a desire for growth, rather than instability [4][5]. - Ordinary workers, on the other hand, may be labeled as lacking stability or capability when they change jobs frequently, which can lead to negative consequences during resume screenings by employers [6][8]. - The article suggests that the rules governing perceptions of job-hopping are often biased against ordinary workers, while high achievers are not bound by the same standards [7][8].

深度研究（Deep Research），让人工智能（AI）系统从"生成文本"进化为"发现知识"，进而完成复杂的开放式任务。 目前，Deep Research 已被广泛应用于文本生成、科研、软件工程等领域中，帮助任何人完成诸如学术综述、会议海报、演示文稿（PPT）生成，科研选题 生成、实验执行和学术写作，以及地球观测、软件库补全等工作。 然而，这一领域尚处早期阶段，业内缺少关于 Deep Research 路线图、基础组件、实践技术、关键挑战及未来方向的系统描述。 日前，来自山东大学、中国人民大学和清华大学的研究团队及其合作者，在一篇最新综述中详细梳理了 Deep Research 的前世、今生和未来，为 Agentic 研究范式提供了理论基础。 Deep Research 是什么？ Deep Research 赋予了 LLM 一个端到端的研究工作流程，使其能够作为智能代理，在最少的人工干预下生成连贯且有明确来源依据的研究报告。在实际操 作中，Deep Research 系统的 LLM 代理会先规划研究问题，从多种异构来源中获取并筛选相关信息，维护并更新工作记忆，最终输出经过验证、并标注准 确来源的答案。 研究 ...

Investment Rating - The report does not explicitly state an investment rating for the industry or specific companies [2]. Core Insights - The AI application activity has seen a significant rebound, particularly with Gemini showing notable growth, while domestic applications remain stable. OpenAI has released the GPT-5.2 series, focusing on optimizing agent workflows [2][7]. - Oracle reported a total revenue of $16.1 billion for Q3 2025, marking a 13% year-over-year increase, with cloud revenue reaching $8 billion, up 33% [2][13]. Summary by Sections AI Applications - OpenAI launched multiple updates, including GPT-5.2, while Google expanded its applications significantly, enhancing productivity features for enterprise users [7][12]. - The active usage of chat assistant applications has increased, with Gemini leading the growth, while other applications like Claude and ChatGPT also saw slight recoveries [9][12]. Oracle's Performance - Oracle's cloud business continues to grow, with cloud infrastructure revenue increasing by 66% year-over-year, and GPU-related revenue soaring by 177% [13][14]. - The company's remaining performance obligations (RPO) reached $523.3 billion, a staggering 433% increase year-over-year, indicating strong future revenue potential [14][17]. - Despite robust revenue growth, Oracle faces cash flow pressures, with a free cash flow of -$10 billion due to significant capital expenditures [17][18].