曲凯： 今天我们请来了国内强化学习 (RL) 领域的专家吴翼，吴翼目前是清华大学交叉信息研究院助理教授，他曾经在 OpenAI 工作过，算是国内最早研究强化学 习的人之一，我们今天就争取一起把 RL 这个话题给大家聊透。 首先吴翼能不能简单解释一下，到底什么是 RL？ 因此，RL 其实更通用一些，它的逻辑和我们在真实生活中解决问题的逻辑非常接近。比如我要去美国出差，只要最后能顺利往返，中间怎么去机场、选什么航 司、具体坐哪个航班都是开放的。 但 RL 很不一样。 RL 最早是用来打游戏的，而游戏的特点和分类问题有两大区别。 第一，游戏过程中有非常多的动作和决策。比如我们玩一个打乒乓球的游戏，发球、接球、回球，每一个动作都是非标的，而且不同的选择会直接影响最终的结 果。 第二，赢得一场游戏的方式可能有上万种，并没有唯一的标准答案。 所以 RL 是一套用于解决多步决策问题的算法框架。它要解决的问题没有标准答案，每一步的具体决策也不受约束，但当完成所有决策后，会有一个反馈机制来评 判它最终做得好还是不好。 吴翼： RL 是机器学习这个大概念下一类比较特殊的问题。 传统机器学习的本质是记住大量标注过正确答案的数据对。 ...

Core Viewpoint - The future of marketing will shift from being human-centric to model-centric, as AI agents like Manus may become integral to everyday operations [1][2]. Group 1: Opportunities from Large Models - The capabilities of large models, such as DeepSeek-R1, have improved tenfold, presenting significant opportunities for businesses [2]. - Many companies struggle to utilize these models effectively, primarily due to issues like hallucination, which can be turned into entrepreneurial opportunities by leveraging proprietary data [2]. Group 2: Understanding Agents - An "Agent" in the business context can be seen as a representative that serves either supply-side or demand-side interests, with both paths offering substantial opportunities [3][4]. - Manus exemplifies a successful agent by connecting various tools and resources to enhance operational efficiency [5]. Group 3: Building a Company as an Agent - Companies can be deconstructed into four components: perception system, cognitive system, action system, and goals, to transform into an effective agent [6]. - The perception system should incorporate feedback from frontline employees to enhance decision-making and operational efficiency [8][10]. Group 4: Cognitive System - The cognitive system should focus on high-frequency decision-making and leverage AI to improve organizational efficiency [12][14]. - High leverage in decision-making is crucial and should be based on real-time data from frontline interactions [15][17]. Group 5: Action System - In knowledge-intensive industries, the action system is represented by AI, which can automate processes through APIs and RPA [18][19]. - Manus serves as a complex agent that can execute commands and streamline operations across the organization [19]. Group 6: Goals and Feedback Loops - The primary goal of any company is to understand and meet customer needs, creating a feedback loop that enhances responsiveness [20][21]. - Companies should continuously iterate and upgrade their systems to adapt to changing demands and improve efficiency [23]. Group 7: Strategic Recommendations for Entrepreneurs - Entrepreneurs should upgrade their teams and protect core data while focusing on marketing towards large models rather than just individuals [25][26]. - Product design should anticipate a future where robots and smart devices may replace human interaction, necessitating a shift in design philosophy [26].

石海东告诉钛媒体App："DeepSeek不仅在大模型推理成本和推理能力实现了极大优化。更重要的是， 对于各行业客户而言，过去对大模型存在负面印象，包括幻觉、训练数据等偏见和缺陷性问题， DeepSeek正在抹除这部分担心。这进一步推动了深度垂直智能体的落地。" 未来会有大量专业Agent，而不是只有一个超级Agent 相较于通用型Agent，面向B端垂直场景的Agent其机会窗口正在扩大。春节过后，DeepSeek的出圈，中 国有至少60家银行相继宣布对接了DeepSeek，但基本面向投资者的投研报告、财报分析、客户资料分 析报告等非核心业务场景，亦或者是OA、办公自动化，IT代码开发等非业务场景。 与一些企业或厂商先高调发声再选择行动不同的是，飞虎互动深入金融行业的Agent这件事情已经一年 有余。目前围绕金融银行三大关键环节：营销-风控-交易，飞虎互动构建了三款大模型驱动的Agent用 例，包括对客营销机器人，风控合规机器人，交易服务机器人。 飞虎互动公司创始人董事长石海东及团队与钛媒体交流中指出，AI大模型在银行业务价值和落地优先 级高的其实是在营销、风控领域，目前DeepSeek还没有接入到这些领域 ...

Core Viewpoint - The company reported a decline in total operating revenue and an increase in net loss for 2024, indicating ongoing challenges during its transformation phase [1][2]. Group 1: Financial Performance - In 2024, the company achieved total operating revenue of 9.153 billion, a year-on-year decrease of 6.6% [1]. - The net profit attributable to shareholders was a loss of 2.061 billion, which is an increase in loss by 1.09 billion compared to 2023 [1]. - The actual revenue was in line with the forecast median, while the actual loss slightly exceeded the forecasted net loss range of 1.72-1.92 billion [1]. Group 2: Revenue Drivers - The decline in operating revenue was primarily due to a temporary delay in customer demand and a decrease in signed amounts [1]. - The transition to a subscription business model has impacted short-term overall revenue [1]. - The increase in losses was attributed to higher amortization of capitalized intangible assets, increased employee compensation due to layoffs, and higher goodwill impairment losses [1]. Group 3: Cloud Transformation and AI Strategy - In 2024, the cloud service business generated revenue of 6.85 billion, a year-on-year decrease of 3.4%, while subscription revenue grew by 26.0% [2]. - The company reported contract liabilities of 3.05 billion, an increase of 8.8% from the end of 2023, with cloud-related contract liabilities growing by 13.0% [2]. - The company launched the enterprise service model YonGPT2.0, focusing on AI and agent technologies to bridge complex enterprise needs with general models [2]. Group 4: Future Outlook - The company maintains an "overweight" rating despite adjusting profit forecasts for 2025-2026 due to lower-than-expected client investment and ongoing transformation impacts [3]. - Projected revenues for 2025-2027 are 10.51 billion, 12.19 billion, and 14.18 billion, respectively, with net profits of 0.1 billion, 0.41 billion, and 0.84 billion [3]. - The company is expected to benefit from the successful advancement of its cloud and AI initiatives, with a clear industry position and potential for long-term growth [3].

Investment Rating - The industry investment rating is "Overweight" [5] Core Insights - AI Coding is becoming a breakthrough point for the commercialization of Agents, with the programming field's clear rules providing a natural constraint framework for Agent applications. The technical characteristics of programming environments offer an ideal testing ground for Agent self-correction, while the atomic tasks in programming align well with the chain reasoning mechanism of large models. The strong demand for enterprise development efficiency creates a clear willingness to pay, leading to a complete closed loop of "technology validation - product iteration - commercial monetization" in the AI programming field [1][8]. Summary by Sections Development Stages of AI Large Models in Programming - The application development of AI large models in programming is divided into three stages: 1. LLM as Copilot: Assists programmers without changing the professional division of software engineering. 2. LLM as Agent: Can autonomously complete certain tasks, acting as a single-function expert. 3. LLM as Multi-Agent: Multiple agents collaborate to complete complex tasks, with humans responsible for creativity and confirmation [2][9]. Key Products and Companies - Notable AI programming products include: - GitHub Copilot: Launched in 2021, it has 1.8 million paid subscribers and an annual recurring revenue (ARR) of $300 million, accounting for 40% of GitHub's overall revenue growth [13]. - Cursor: A specialized IDE that integrates AI deeply, focusing on optimizing user experience and model interaction [16]. - Devin: An AI programmer capable of independently completing projects, with a subscription fee of $500/month [20][21]. - Baidu Comate: Upgraded to Agent mode, achieving a code adoption rate of 46% among its users [26][27]. - Alibaba Tongyi Lingma: An AI programmer that can autonomously handle complex development tasks, significantly improving efficiency [28][29]. - Tencent Cloud AI Code Assistant: Achieved a 30%+ improvement in code generation accuracy after integrating DeepSeek-R1 [31]. Market Performance - The computer industry has shown relative performance with a 1-month increase of 4.59%, a 3-month increase of 7.49%, and a 12-month increase of 34.16%, outperforming the CSI 300 index [7].

Investment Rating - The report provides an investment rating of "Overweight" for the industry, indicating an expectation that the industry stock index will outperform the benchmark [61]. Core Insights - The report emphasizes the emergence of AI agents as a transformative technology, highlighting their ability to automate complex tasks and enhance productivity across various sectors [14][32]. - Manus AI is identified as a leading general AI agent, showcasing superior performance in handling complex tasks compared to other AI assistants [14][41]. - The report discusses the engineering phase of AI agents, focusing on their integration into complex data scenarios and the importance of data value in their development [27][34]. Summary by Sections Industry Overview - The AI agent industry is entering a phase of engineering, with advancements in user interface development and memory management technologies [27][28]. - Companies like Workday are launching AI agent management systems to enhance efficiency and data security in enterprise environments [32]. Technology and Performance - Manus AI has achieved state-of-the-art results in benchmark tests, particularly excelling in complex task execution [13][14]. - The technology architecture of Manus AI includes various models that support its capabilities, indicating a robust framework for task automation [17][18]. Market Trends - The report notes a growing trend of AI agents being utilized in both personal and enterprise settings, with applications ranging from recruitment automation to payroll processing [33][32]. - The competitive landscape includes major players like Tencent and ByteDance, which are actively developing AI agents for diverse applications [30][32]. Future Outlook - The report predicts that AI agents will continue to evolve, enhancing their task planning and tool usage capabilities, moving towards more autonomous and generalized functions [41][42]. - The potential for multi-agent collaboration is highlighted, suggesting that AI agents can work together to solve complex problems and improve overall efficiency [48][51].

点击 上方 硬AI 关注我们 如果Scaling Law继续有效， 继续看好AI系统组件供应商（如芯片、网络设备等），谨慎对待那些不得不持续投入巨额资 本支出的科技巨头。如果预训练缩放停滞： 看好科技巨头（因为自由现金流将回升），并关注那些拥有大量用户、能够 从推理成本下降中获益的应用类股票。 硬·AI 作者 |硬 AI 编辑 | 硬 AI 还抱着"越大越好"的AI模型不放？华尔街投行巴克莱最新研报给出了一个颠覆性的预测： AI行业正经历一 场"巨变"（Big Shift），"推理模型"和"Agent"将成为新时代的弄潮儿，而"大力出奇迹"的传统大模型， 可能很快就要过气了！ 这场变革的核心，是AI模型从"死记硬背"到"举一反三"的进化。过去，我们追求更大的模型、更多的参 数、更海量的训练数据，坚信"量变产生质变"。但现在，巴克莱指出，这条路可能已经走到了尽头。 算力无底洞、成本高企、收益却难以匹配……传统大模型的"军备竞赛"让众多科技巨头苦不堪言。更要命 的是，用户真的需要那么"大"的模型吗？在许多场景下，一个更"聪明"、更会推理的小模型，反而能提供 更精准、更高效的服务。 这究竟是怎么回事？对于投资者来说 ...

Investment Rating - The report suggests a positive outlook on the industrial AI sector, highlighting potential breakthroughs in deployment, reliability, and cost-effectiveness [1]. Core Insights - The report emphasizes that the application of AI in industrial settings may progress faster than market expectations, with significant advancements in multi-modal large models and the integration of AI agents [1][6]. - It identifies three core application modes of industrial AI: recognition applications, data modeling and optimization applications, and knowledge reasoning and decision-making applications [25][26]. - The report indicates that the cost of AI token inputs has significantly decreased, while labor costs in the manufacturing sector have been rising, suggesting a potential tipping point for "machine replacement" [27][28]. Summary by Sections 1. Industrial AI - The report discusses the challenges of high data complexity, low tolerance for errors, and high cost sensitivity in industrial AI applications [6]. - It highlights the evolution of multi-modal large models, which are expected to reduce the difficulty of processing unstructured data in industrial scenarios [7][9]. - The integration of large models for guidance and small models for execution is proposed as a collaborative approach to enhance reliability in production processes [22][23]. 2. Industrial AI Middleware - The report notes that industrial AI middleware is in its early penetration phase, addressing the need for seamless iteration between model and data sides [1][2]. - It outlines the commercial progress of industrial AI middleware, with significant contracts awarded, such as a project worth 48.67 million yuan for AI middleware capabilities [39][41]. - The middleware is expected to have high construction barriers, requiring capabilities in computing power integration, model management, and industrial data governance [2][3]. 3. Industrial AI Applications - The report identifies that the maturity of AI applications in production control is leading, with over 57% of applications focused on this area [1][3]. - It highlights specific use cases, such as the AI+PCB solution by Saiyi Information, which automates the entire process from parameter analysis to cost query and quote generation [21]. - The report also mentions the emergence of vertical agent applications in various sectors, indicating a shift towards subscription-based models in industrial AI [22][39]. 4. Investment Recommendations - The report recommends focusing on key players in the industrial AI sector, such as Zhongkong Technology, Saiyi Information, and Zhongwang Software, as they are expected to benefit from the ongoing developments in industrial AI applications [1][2].

昨晚到今天，AI圈有3个重磅消息，中国科技的叙事持续加强。 阿里通义没有食言，说这周再开源一个RL新模型，昨晚放出来了。最厉害的是32B性能比肩满血DeepSeek R1，在测试数学能力的AIME24评测集上，以及评 估代码能力的LiveCodeBench中，千问QwQ-32B表现与DeepSeek-R1相当，远胜于o1-mini及相同尺寸的R1蒸馏模型，现在已经可以在通义APP和网页端体 验了。 而且看起来，这个RL训练并没有花费太长时间，阿里的朋友反馈，与以往奖传统励模型不同的是，说这次是通过校验生成答案的正确性来为数学问题提供反 馈。 14:10 M Junvang Lin @ 17 阿里通义开源RL新模型 @ lustin| in610 This week we release QwQ-Max-Preview on Qwen Chat. I know you guys may think what happened to the opensource of this team. Here is a straight answer to you all: we will opensource the m ...

Group 1 - Zhiyuan AI recently completed a strategic financing round exceeding 1 billion yuan, with investors including Hangzhou Urban Investment and Shangcheng Capital, aimed at promoting technological innovation and ecological development of domestic GLM models [1][2] - The newly established Zhejiang Zhiyuan New Chapter Technology Co., fully owned by Zhiyuan AI, has a registered capital of 450 million yuan, focusing on AI software development [1][2] - Zhiyuan AI is the only domestic company fully benchmarking against OpenAI, with a comprehensive layout in various models including GLM, dialogue models, and multi-modal models [3] Group 2 - The financing will enhance Zhiyuan AI's ability to serve the economic entities in Zhejiang Province and the Yangtze River Delta region, facilitating the digital transformation of industries based on AI technology [2] - Zhiyuan AI plans to release a new large model in 2025, which will be open-sourced, including base models, inference models, multi-modal models, and agents [3][4] - The company has achieved significant commercial success, with revenue exceeding 100 million yuan shortly after the Spring Festival, and a 30% increase in API platform payments post-holiday [6]