2025年08月25日 证券研究报告 | 人工智能行业专题（11） 探究模型能力与应用的进展和边界 行业研究 · 行业专题 互联网 · 互联网II 投资评级：优于大市（维持） 证券分析师：张伦可 证券分析师：陈淑媛 证券分析师：刘子谭 证券分析师：张昊晨 0755-81982651 021-60375431 liuzitan@guosen.com.cn zhanghaochen1@guosen.com.cn zhanglunke@guosen.com.cn chenshuyuan@guosen.com.cn S0980525060001 S0980525010001 S0980521120004 S0980524030003 请务必阅读正文之后的免责声明及其项下所有内容 报告摘要 Ø 风险提示：宏观经济波动风险、广告增长不及预期风险、行业竞争加剧风险、AI技术进展不及预期风险等。 请务必阅读正文之后的免责声明及其项下所有内容 2 Ø 本篇报告主要针对海内外模型发展、探究模型能力与应用的进展和边界。我们认为当前海外模型呈现差异化发展，企业调用考虑性价比。当前 OpenAI在技术路径上相对领先，聚焦强化推理与专业 ...

Core Insights - The article discusses the ongoing layoffs in Silicon Valley and the paradox of AI's efficiency gains leading to increased costs in other areas, particularly in rework and corrections [1][2][3][4]. Group 1: AI's Impact on Employment and Costs - Many companies are adopting AI with the expectation of reducing costs and increasing efficiency, but the reality is that they are often spending more on rework due to AI-generated errors [23][24]. - A significant portion of entry-level jobs is expected to be replaced by AI, with predictions of unemployment rates in the U.S. potentially rising to 10%-20% [7]. - The initial savings from AI implementations are often negated by the costs associated with correcting AI mistakes, leading to a cycle of increased expenditure [8][10][36]. Group 2: The Rise of New Roles and Responsibilities - A new profession has emerged focused on correcting and refining AI-generated outputs, indicating a shift in job roles from creation to correction [4][13]. - Companies are increasingly hiring specialists to address issues caused by AI, such as bugs in code or errors in customer service interactions, which were previously manageable without AI [15][20][21]. - The need for human oversight in AI operations is becoming more apparent, as AI cannot fully replace the judgment and responsibility required in many work scenarios [21][48]. Group 3: Consumer and Brand Reactions - There is growing consumer backlash against companies that overly rely on AI, with brands facing negative perceptions when AI fails to meet expectations [34][36]. - High-profile cases, such as Klarna's experience with AI customer service, illustrate the risks of sacrificing quality for cost savings, leading to a reversal in staffing strategies [39][40]. - The failure of AI-driven initiatives, such as the automated store experiment, highlights the limitations of current AI capabilities and the necessity for human intervention [42][45]. Group 4: Long-term Perspectives on AI Integration - Historical patterns suggest that new technologies, including AI, often experience initial setbacks before achieving their full potential, as illustrated by the "J-curve" concept [46][47]. - Companies must recognize that while AI can enhance processes, it cannot replace the need for human oversight and accountability, especially when errors occur [48].

Group 1: Figma Overview - Figma is a cloud-based collaborative design software that allows multiple roles such as designers, developers, and product managers to work together in real-time, disrupting traditional design software models [1] - As of March 2025, Figma has 13 million monthly active users, with two-thirds being non-traditional designers, making it the most popular UI design tool globally [1] - Figma's revenue for FY24 reached $749 million, a 48% increase year-over-year, with Q1 FY25 revenue at $228 million, up 46% [2] Group 2: Figma's Business Model and Growth - 70% of Figma's revenue comes from large customers, with the number of customers generating over $100,000 in annual recurring revenue (ARR) increasing to 1,031, a 47% growth [2] - Figma is expanding from a single design tool to a comprehensive platform covering the entire process from conception to launch, with 76% of customers using two or more products [2] - The total addressable market (TAM) for Figma is projected to be $33 billion, with strong user growth and AI integration expected to drive future revenue [2] Group 3: Figma's Valuation and Market Comparison - Figma's current revenue growth exceeds 40%, with a free cash flow margin of 28% and a 40% rule metric above 60%, suggesting a higher valuation compared to similar SaaS companies like Crowdstrike [3] - Figma's IPO pricing range was raised to $30-32 per share, valuing the company at $18.8 billion, up from an initial range of $25-28 [2] Group 4: Amazon's Financial Performance - Amazon reported Q2 FY25 revenue of $167.7 billion, a 10% year-over-year increase, and net profit of $18.2 billion, up 35% [4] - AWS revenue for Q2 FY25 was $30.87 billion, a 17% increase year-over-year, but growth was slower compared to competitors like Microsoft Azure and Google Cloud [5][6] Group 5: Amazon's Business Segments - Amazon's online store revenue for Q2 FY25 was $61.49 billion, an 11% increase year-over-year, slightly exceeding market expectations [5] - The third-party seller services segment generated $40.35 billion in revenue, up 11% year-over-year, while advertising revenue reached $15.69 billion, a 17% increase [9] Group 6: Amazon's Future Outlook - Amazon's Q3 FY25 revenue guidance is between $174 billion and $179.5 billion, indicating a 10-13% year-over-year growth, but operating profit guidance is below market expectations [5] - AWS faces supply constraints, with a backlog of $195 billion in orders as of June 30, reflecting a 25% year-over-year increase [6]

Core Viewpoint - The article discusses the limitations of AI models in accurately interpreting images, highlighting that these models rely on memory and biases rather than true visual observation [19][20][48]. Group 1: AI Model Limitations - All tested AI models, including Grok4, OpenAI o3, and Gemini, consistently miscounted the number of fingers in an image, indicating a systemic issue in their underlying mechanisms [11][40]. - A recent paper titled "Vision Language Models are Biased" explains that large models do not genuinely "see" images but instead rely on prior knowledge and memory [14][19]. - The AI models demonstrated a strong tendency to adhere to preconceived notions, such as the belief that humans have five fingers, leading to incorrect outputs when faced with contradictory evidence [61][64]. Group 2: Experiment Findings - Researchers conducted experiments where AI models were shown altered images, such as an Adidas shoe with an extra stripe, yet all models incorrectly identified the number of stripes [39][40]. - In another experiment, AI models struggled to accurately count legs on animals, achieving correct answers only 2 out of 100 times [45]. - The models' reliance on past experiences and biases resulted in significant inaccuracies, even when prompted to focus solely on the images [67]. Group 3: Implications for Real-World Applications - The article raises concerns about the potential consequences of AI misjudgments in critical applications, such as quality control in manufacturing, where an AI might overlook defects due to its biases [72][76]. - The reliance on AI for visual assessments in safety-critical scenarios, like identifying tumors in medical imaging or assessing traffic situations, poses significant risks if the AI's biases lead to incorrect conclusions [77][78]. - The article emphasizes the need for human oversight in AI decision-making processes to mitigate the risks associated with AI's inherent biases and limitations [80][82].

本位作者分别来自南京大学，罗格斯大学和马萨诸塞大学阿默斯特分校。第一作者韩廷旭与共同第一作者王震霆是分别来自南京大学和罗格斯大学的博士生，研 究方向聚焦于大模型推理以及安全负责任的生成式人工智能。通讯作者为南京大学房春荣教授。 随着大型语言模型（LLM）技术的不断发展， Chain-of-Thought（CoT） 等推理增强方法被提出，以期提升模型在数学题解、逻辑问答等复杂任务中的 表现，并通过引导模型逐步思考，有效提高了模型准确率。 然而，这类方法也带来了新的挑战：模型生成的中间推理过程往往冗长，产生了大量冗余 Token ，这显著增加了推理阶段的计算成本和资源消耗。在 LLM 日益走向实际部署的背景下，如何在保证推理能力的同时控制成本，已成为制约其大规模应用的核心问题。 为解决这一矛盾，近日来自南京大学、罗格斯大学和马萨诸塞大学阿默斯特分校的研究团队提出了一种基于 Token 预算感知 的 LLM 推理新框架 TALE ， 旨在保证推理准确率的同时，显著压缩输出长度、降低计算开销。 TALE 的核心理念是在推理过程中引入「Token 预算」这一约束机制，引导模型在限定的 Token 预算范围内完成有效推理 ...

Summary of Deepseeker R1 Conference Call Company and Industry - The discussion revolves around the performance and updates of the Deepseeker R1 model, a product in the AI and machine learning industry, particularly focusing on its capabilities in data retrieval and code generation. Core Points and Arguments - **Performance Improvement**: The accuracy of Deepseeker R1 in CLion improved from 4/8 to 6/8 in version 0.528, although it still lags behind Claude 3.7 (7/8) and CosmoFlow with Claude 4 (8/8) [1][3][19]. - **Context Length Enhancement**: The new version increased the maximum context length to 128K for clients, addressing previous issues where excessive web content retrieval exceeded context limits [5][19]. - **Challenges in Data Retrieval**: The model faced difficulties using the fetch tool to retrieve China’s GDP data due to low success rates and lack of API support from the World Bank, indicating compatibility issues between MCP tools and large models [6][19]. - **Comparison with Other Models**: Readcloud 3.7, Readcloud 4, Grok 3, and Gemini 2.5 Pro demonstrated better performance in using MCP tools and parameter settings, successfully completing tasks that Deepseeker R1 struggled with [7][19]. - **Code Generation Quality**: While the new version shows improvements in reasoning and text generation quality, the code generation aspect still has flaws compared to Claude series models [4][19]. - **Error Handling in MCP Tools**: The MCP tools often encounter issues when a tool fails, and the selection of alternatives is not always ideal. Readcloud has shown the ability to quickly find substitutes when issues arise [13][14]. Other Important but Possibly Overlooked Content - **Task Complexity**: The complexity of tasks requiring multiple MCP tools can lead to cascading errors if one tool fails, emphasizing the need for careful planning and tool selection [11][19]. - **Improvements in Cloud 4**: Cloud 4 outperforms Cloud 3.7 in data scraping and webpage generation, with faster speeds and higher accuracy, showcasing advancements in the technology [10][19]. - **Devsec Error Handling**: Devsec's error handling is contingent on initial tool selection, suggesting a need for improved recognition and selection of backup options to enhance reliability [15][19]. - **Limitations in Code Generation**: Despite improvements, the new version's code generation still falls short in quality compared to Claude 3.7 and 4, particularly in achieving expected outcomes in specific projects [17][19]. - **Overall Model Comparison**: Claude 4 is noted for its superior speed and accuracy, especially in programming tasks, indicating a competitive edge over Deepseeker R1 [18][19].

Core Viewpoint - Anthropic's Claude Opus 4 is claimed to be the "world's strongest programming model," with a notable case of solving a long-standing bug faced by an experienced developer, ShelZuuz, showcasing its capabilities [1][2]. Group 1: Bug Resolution Case - ShelZuuz, a developer with over 30 years of C++ experience, struggled with a "white whale bug" for four years, which was a rendering error triggered under specific conditions [2][3][4]. - The bug was introduced during a code refactor of a 60,000-line project, leading to a silent failure that was difficult to reproduce and diagnose [4][5]. - After attempting various methods without success, ShelZuuz used Claude Opus 4, which identified the root cause of the bug in just a few hours, significantly faster than previous attempts [6][9]. Group 2: AI Capabilities and Limitations - Claude Opus 4's approach involved analyzing both old and new code versions, automatically identifying key differences and dependencies that were overlooked during the refactor [7][9]. - Despite successfully solving the bug, ShelZuuz emphasized that Claude Opus 4 functions more like a capable junior developer rather than a replacement for experienced engineers [10][12]. - The AI requires substantial guidance and oversight, akin to managing a junior programmer, rather than functioning autonomously [12][13]. Group 3: Cost Efficiency - The subscription cost for Claude Opus 4 is $100 per month, which is significantly lower than the cost of hiring a senior engineer, estimated at around $25,000 for 200 hours of work [13]. - This highlights the potential of AI to enhance development efficiency and reduce costs in the software engineering field [13].

Core Viewpoint - OpenAI is transitioning its for-profit entity into a public benefit corporation (PBC) while maintaining its non-profit status, with the non-profit organization controlling the PBC. This shift emphasizes OpenAI's commitment to non-profit principles amidst increasing competition in the AI sector [1]. Group 1 - OpenAI's valuation is currently at $300 billion, while a new project by former employee Ilya, SSI, is valued at $20 billion, indicating a competitive landscape for AI investments [1]. - The industry is witnessing a significant shift towards open-source models, with successful examples like Llama4 and Deepseek R1, which are rapidly catching up to OpenAI's earlier models [1][2]. - The estimated gap between AI model generations is currently within 14 months, suggesting a fast-paced evolution in the AI field [2]. Group 2 - OpenAI's pricing for its models, such as O1 and O3, is more than double that of competitors like R1, which may impact its market position as application usage surges [3]. - The latest quarter saw a 4-5 times increase in API call volume for AI models, indicating a growing demand for AI applications [3]. - OpenAI is expected to face unprecedented challenges due to the rise of competitive models and changing market dynamics [4].

Core Viewpoint - Gemini 2.5 Pro has successfully completed the Pokémon Blue game, marking a significant achievement in AI capabilities, particularly in gaming contexts [1][3][18]. Group 1: Achievement and Comparison - Gemini 2.5 Pro is the first large model to become a Pokémon League Champion and enter the Hall of Fame in Pokémon Blue [3]. - In comparison, the previous model, Claude 3.5, struggled to progress in the game, only reaching the forest area, while Claude 3.7 managed to defeat gym leaders but did not complete the game [3][9]. Group 2: Gameplay Process - The gameplay process involved Gemini exploring the game world, specifically aiming to capture Mewtwo in the Cerulean Cave, which required extensive thought and planning, consuming 76,011 tokens for a single action [8][9]. - The model's decision-making process was displayed in real-time, showcasing its reasoning behind each action taken [7][8]. Group 3: Challenges Faced - Despite its success, Gemini's performance highlighted challenges in navigating the game, often getting lost, indicating that AI still struggles with spatial reasoning in low-resolution environments [9][10][12]. - The model's limitations in visual interpretation and context understanding were noted, as it had difficulty recognizing in-game structures and their interactions [11][13][16]. Group 4: Future Implications - The achievement by Gemini suggests a potential shift in benchmarks for evaluating large models, with future assessments possibly focusing on their ability to complete games like Pokémon [19]. - Google plans to continue exploring this area, indicating ongoing developments in AI gaming capabilities [18].

早上群里有个朋友说自己用 Gemini APP 里面的深度研究搞了一个特斯拉 Q1 财报的分析文档，另一个朋友 说转成网页，我就说我试试。 我直接把他的文档和我最近探索出来的提示词就放到了 Chatwise 里面，以往我都是用 Claude 3.7 生成网页 的，这次默认是 Gemini 2.5 Pro，我也没看就按下了回车。 没想到生成的网页炒鸡惊艳，Gemini 的网页内容很多同时理解了提示词提到的设计风格，非常漂亮。 可以看图也可以在这里预览： https://kueaqan0fo.app.yourware.so/ | | | | | | $0.41 | 可比 -13% YoY | 同比 -16% YoY 网比 +154% YoY | | Acknowledged uncertainty, 94 update planned. Unusual admission of political/brand impact. | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | FRITTY les a 2025 01 ...