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 insights shared by Denny Zhou, a leading figure in AI, regarding the reasoning capabilities of large language models (LLMs) and their optimization methods [3][4]. Group 1: Key Points on LLM Reasoning - Denny Zhou emphasizes that reasoning in LLMs involves generating a series of intermediate tokens before arriving at a final answer, which enhances the model's strength without increasing its size [6][15]. - The challenge lies in the fact that reasoning-based outputs often do not appear at the top of the output distribution, making standard greedy decoding ineffective [6]. - Techniques such as chain-of-thought prompting and reinforcement learning fine-tuning have emerged as powerful methods to enhance LLM reasoning capabilities [6][29]. Group 2: Theoretical Framework - Zhou proposes that any problem solvable by Boolean circuits can be addressed by generating intermediate tokens using a constant-sized transformer model, indicating a theoretical understanding of reasoning [16]. - The importance of intermediate tokens in reasoning is highlighted, as they allow models to solve complex problems without requiring deep architectures [16]. Group 3: Decoding Techniques - The article introduces the concept of chain-of-thought decoding, which involves checking multiple generated candidates rather than relying on a single most likely answer [22][27]. - This method requires programming effort but can significantly improve reasoning outcomes by guiding the model through natural language prompts [27]. Group 4: Self-Improvement and Data Generation - The self-improvement approach allows models to generate their own training data, reducing reliance on human-annotated datasets [39]. - The concept of reject sampling is introduced, where models generate solutions and select the correct steps based on achieving the right answers [40]. Group 5: Reinforcement Learning and Fine-Tuning - Reinforcement learning fine-tuning (RL fine-tuning) has gained attention for its ability to enhance model generalization, although not all tasks can be validated by machines [42][57]. - The article discusses the importance of reliable validators in RL fine-tuning, emphasizing that the quality of machine-generated training data can sometimes surpass human-generated data [45][37]. Group 6: Future Directions - Zhou expresses anticipation for breakthroughs in tasks that extend beyond unique, verifiable answers, suggesting a shift in focus towards building practical applications rather than solely addressing academic benchmarks [66]. - The article concludes with a reminder that simplicity in research can lead to clearer insights, echoing Richard Feynman's philosophy [68].

Core Insights - The article discusses the limitations of current Visual Language Models (VLMs) in spatial reasoning and multi-perspective understanding, highlighting the need for improved AI systems that can collaborate effectively with humans [1][3][20]. Group 1: ViewSpatial-Bench Development - A new benchmark system called ViewSpatial-Bench has been developed by research teams from Zhejiang University, University of Electronic Science and Technology of China, and The Chinese University of Hong Kong to evaluate VLMs' spatial reasoning capabilities across multiple perspectives [4][33]. - ViewSpatial-Bench includes 5 different task types and over 5700 question-answer pairs, assessing models from both camera and human perspectives [5][7]. - The benchmark aims to address the fragmented understanding of spatial information in VLMs, which often leads to performance issues in multi-perspective tasks [2][20]. Group 2: Model Performance Evaluation - The evaluation of various leading models, including GPT-4o and Gemini 2.0, revealed that their performance in understanding spatial relationships is still inadequate, with overall accuracy scores being low [19][20]. - The results indicated a significant performance gap between tasks based on camera perspectives and those based on human perspectives, suggesting a lack of unified spatial cognitive frameworks in current VLMs [22][23]. - The Multi-View Spatial Model (MVSM) was introduced to enhance cross-perspective spatial understanding, achieving a 46.24% absolute performance improvement over its backbone model [27][28]. Group 3: Future Directions - The findings highlight the structural imbalance in training data regarding perspective distribution, indicating a need for future data construction and model optimization efforts [26]. - The development of MVSM and ViewSpatial-Bench provides a feasible path for AI systems to achieve human-like spatial cognitive abilities, which is crucial for the next generation of robots and multimodal assistants [34].

Core Insights - The article emphasizes that AI, like the internet decades ago, is at the beginning of a transformative phase that could redefine human productivity and creativity, leading to a state of "super agency" where humans and machines collaborate effectively [1][4][5]. Group 1: AI's Transformative Potential - AI is seen as a powerful tool that can enhance human capabilities, acting as a "force multiplier" rather than just a tool [4][5]. - The concept of "super agency" describes how individuals can leverage AI to significantly boost their creativity, productivity, and influence [5]. - AI is expected to democratize knowledge acquisition and automate numerous tasks, provided it is developed and deployed safely and equitably [5][7]. Group 2: Historical Context and Public Perception - Historical technological advancements often faced initial skepticism, with concerns about their negative impacts overshadowing their potential benefits [3]. - The narrative around AI is influenced by dystopian themes, yet there is a call to reframe this perspective to envision positive outcomes [3][4]. Group 3: AI's Advancements and Capabilities - AI is evolving to automate cognitive functions, enabling it to adapt, plan, and make decisions autonomously, which could drive unprecedented economic growth and social change [7][8]. - Significant advancements in AI, such as large language models (LLMs), have shown remarkable performance in standardized tests, indicating a leap in reasoning capabilities [8][9]. Group 4: Autonomous AI and Its Implications - Agentic AI is emerging, capable of independent action and complex task execution, marking a shift from passive tools to proactive digital partners [11][12]. - Companies are integrating agentic AI into their core products, enhancing collaboration between humans and automated systems [13]. Group 5: Multi-modal AI Development - Current AI models are advancing towards multi-modal capabilities, processing various data types (text, audio, video) simultaneously, which enhances understanding and interaction [14][15]. - Self-supervised learning techniques are being utilized to improve multi-modal models, allowing them to learn from unlabelled data and perform better across tasks [16][17]. Group 6: Hardware Innovations and AI Performance - Innovations in hardware, such as specialized chips, are driving improvements in AI performance, enabling faster and more efficient model training and execution [18][19]. - The rise of edge computing is enhancing AI's responsiveness and efficiency, particularly in real-time applications [20][21]. Group 7: Transparency and Safety in AI - There is a growing emphasis on improving AI transparency and interpretability, which are crucial for safe deployment and reducing biases [22][23]. - Progress is being made in enhancing the transparency of AI models, with notable improvements in scores reflecting their interpretability [23]. Group 8: Challenges in AI Adoption - Companies face significant challenges in AI transformation, including leadership alignment, cost uncertainty, workforce planning, supply chain management, and the need for greater interpretability [26][27][28]. - Successful AI deployment requires strategic transformation beyond mere technology implementation, focusing on organizational structure and mindset [28][29]. Group 9: Future Directions and Leadership - The article advocates for an iterative deployment approach to AI, encouraging collaboration and gradual adaptation rather than excessive regulation [29]. - Leaders are urged to prioritize human agency in AI development, ensuring that technology serves to enhance human capabilities [30][31].

Core Viewpoint - Generative AI is set to continue transforming software development, with significant advancements expected by 2025, despite current tools not fully democratizing coding for non-engineers [1][35][67]. Group 1: Impact of Generative AI on Software Engineering - The introduction of large language models (LLMs) like ChatGPT has led to a significant increase in AI tool usage among developers, with approximately 75% utilizing some form of AI for software engineering tasks [1]. - The media has sensationalized the potential impact of AI on software engineering jobs, often lacking insights from actual software engineers [1][2]. - AI tools are reshaping software engineering but are unlikely to cause dramatic changes as previously suggested [2]. Group 2: Practical Observations and Challenges - Addy Osmani's article highlights the dual modes of AI tool usage among developers: "Accelerators" for rapid prototyping and "Iterators" for daily development tasks [3][7][10][11]. - Despite increased efficiency reported by developers using AI, the overall quality of software has not significantly improved, indicating underlying issues in software development practices [5][26]. - The "70% problem" illustrates that while AI can help complete a majority of tasks quickly, the remaining complexities often lead to frustration, especially for non-engineers [14][15][20]. Group 3: Effective AI Utilization Strategies - Successful AI integration involves methods such as "AI Drafting," "Continuous Dialogue," and "Trust and Verify" to enhance productivity [27][28][32]. - Developers are encouraged to start small, maintain modularity, and trust their own experience when using AI tools [33][32]. Group 4: Future of Software Engineering with AI - The rise of software engineering agents is anticipated, which will operate more autonomously and collaboratively with human developers [35][38][42]. - The demand for experienced software engineers is expected to increase as they are better equipped to leverage AI tools effectively and manage the complexities that arise from AI-generated code [67]. - The evolution of AI tools may lead to a resurgence in personal software development, focusing on user-centric design and quality [53][54].

Core Insights - The article discusses the hallucination problem in large language models (LLMs), particularly focusing on DeepSeek-R1, which has a high hallucination rate compared to its predecessor and other models [2][6][13] - Li Yanhong criticizes DeepSeek-R1 for its limitations, including high hallucination rates, slow performance, and high costs, sparking discussions about the broader issues of hallucinations in AI models [2][6][19] - The hallucination phenomenon is not unique to DeepSeek, as other models like OpenAI's o3/o4-mini and Alibaba's Qwen3 also exhibit significant hallucination issues [3][8][13] Summary by Sections Hallucination Rates - DeepSeek-R1 has a hallucination rate of 14.3%, significantly higher than DeepSeek-V3's 3.9%, indicating a fourfold increase in hallucination [6][7] - Other models, such as Qwen-QwQ-32B-Preview, show even higher hallucination rates at 16.1% [6][7] - OpenAI's o3 model has a hallucination rate of 33%, nearly double that of its predecessor o1, while the lightweight o4-mini model reaches 48% [8][10] Industry Response - The AI industry is grappling with the persistent issue of hallucinations, which complicates the development of more advanced models [13][19] - Companies are exploring various methods to mitigate hallucinations, including retrieval-augmented generation (RAG) and strict data quality control [20][22][23] - Despite advancements in certain areas, such as multimodal outputs, hallucinations remain a significant challenge in generating long texts or complex visual scenarios [18][19] Implications of Hallucinations - Hallucinations are increasingly seen as a common trait among advanced models, raising questions about their reliability and user trust, especially in professional or high-stakes contexts [17][27] - The phenomenon of hallucinations may also contribute to creativity in AI, as they can lead to unexpected and imaginative outputs [24][26] - The acceptance of hallucinations as an inherent characteristic of AI models suggests a need for a paradigm shift in how AI is perceived and utilized [27]

Core Insights - OpenAI is updating its ChatGPT Search tool to include shopping recommendations, enhancing the online shopping experience for users [1][3] - The new feature will initially support a limited number of product categories, including fashion, beauty, home goods, and electronics, with plans to expand in the future [1][3] Group 1: Feature Overview - ChatGPT Search will provide product recommendations, displaying images, reviews, and links to products when users search [1][3] - The recommendation mechanism is designed to understand user evaluations and discussions rather than relying on traditional algorithmic signals [3] - The service will not allow users to check out within ChatGPT; instead, it redirects users to merchant websites for transactions [3] Group 2: Competitive Landscape - The update is part of OpenAI's strategy to compete with Google, particularly with the upcoming release of Gemini 2.0 [4] - The AI search and online shopping sectors are becoming increasingly competitive, with competitors like Perplexity already offering in-app shopping features [4]

Core Insights - Meta has released a new series of AI models called Llama 4, which includes Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth, trained on a vast amount of unlabelled text, images, and video data to enhance their visual understanding [1][3] - The development of Llama models has accelerated due to the success of open-source models from China's DeepSeek, prompting Meta to establish a war room to analyze cost reductions in running and deploying models [1][2] - Llama 4 models represent a new era for the Llama ecosystem, utilizing a mixture of experts (MoE) architecture for improved computational efficiency [3] Model Performance and Capabilities - According to internal testing, Maverick excels in general assistant and chat scenarios, outperforming OpenAI's GPT-4o and Google's Gemini 2.0 in various benchmarks, although it still lags behind more advanced models like Google’s Gemini 2.5 Pro and Anthropic’s Claude 3.7 Sonnet [4] - Scout is particularly strong in document summarization and reasoning over large codebases, featuring a unique context window of 10 million tokens, allowing it to handle extremely long documents [4] - Behemoth, which is still in training, is expected to require more powerful hardware and has 288 billion active parameters, surpassing GPT-4.5 and Claude 3.7 Sonnet in STEM skill evaluations [5] Licensing and Regulatory Considerations - Developers may raise concerns regarding the licensing of Llama 4, as users and companies registered in the EU are prohibited from using or distributing these models, likely due to AI and data privacy laws [2] - Companies with over 700 million monthly active users must apply for special permission from Meta to use the models, with Meta having discretion over granting such permissions [2]

Core Viewpoint - Meta has launched the Llama 4 family of models, marking a significant advancement in multimodal AI capabilities, with Llama 4 Maverick achieving a high performance score in various benchmarks [3][4][8]. Group 1: Model Overview - The Llama 4 family includes three models: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth, with the first two already released and the latter in training [3][4]. - Llama 4 Scout features 17 billion active parameters and a context window of 1 million tokens, while Llama 4 Maverick has 17 billion active parameters with 128 experts [5][19]. - Llama 4 Behemoth is a massive model with 2 trillion parameters, currently under training, and is expected to outperform existing models like GPT-4.5 and Claude Sonnet 3.7 [5][54]. Group 2: Performance Metrics - Llama 4 Maverick scored 1417 in the latest model ranking, surpassing previous models and becoming the top open-source model [8][9]. - The model outperformed Meta's previous Llama-3-405B by 149 points, marking a significant improvement [8]. - In various benchmarks, Llama 4 Scout demonstrated superior performance compared to competitors like Gemini 2.0 Flash-Lite and Mistral 3.1 [21][42]. Group 3: Multimodal Capabilities - Llama 4 models are designed for native multimodal functionality, allowing users to upload images and ask questions about them directly [30][41]. - The models are touted as the best in their class for multimodal applications, enhancing user interaction and experience [41][42]. Group 4: Cost Efficiency - Llama 4 Maverick offers competitive pricing, with inference costs significantly lower than other models like GPT-4, making it an attractive option for developers [46][49]. - The cost per million input and output tokens for Llama 4 Maverick ranges from $0.19 to $0.495, compared to $4.38 for GPT-4 [49]. Group 5: Training Innovations - The Llama 4 series utilizes a novel MoE (Mixture of Experts) architecture, enhancing computational efficiency by activating only a subset of parameters during inference [56][60]. - The training process involved over 30 trillion tokens, more than double that of Llama 3, and included diverse data types such as text, images, and videos [64][63]. - A new training technique called MetaP was developed to optimize model hyperparameters, resulting in improved performance across various tasks [62][63].

Core Viewpoint - The release of Google's DeepMind's new AI model, Gemini Robotics, marks a significant milestone in the development of general-purpose robots, enhancing their ability to adapt to complex environments and perform challenging tasks [1][9]. Group 1: Technological Advancements - The new AI model allows robots to perform tasks such as folding paper, organizing desks, and even dunking a mini basketball, showcasing its advanced capabilities [3][4][6]. - The Gemini Robotics model is reported to have double the versatility of previous models, representing a major leap towards general-purpose robotics [9]. - The model is trained using Google's Gemini 2.0 language model, endowing robots with three key abilities: environmental adaptability, instruction comprehension, and operational flexibility [10]. Group 2: Market Potential - Analysts predict a significant market expansion for humanoid robots, with an estimated annual sales of 1 million units by 2030 and a total ownership of 3 billion units by 2060, equating to 0.3 robots per person [13]. - Major tech companies, including Tesla and OpenAI, are racing to develop AI capabilities for robots, indicating a competitive landscape in the robotics sector [13]. - NVIDIA's CEO has stated that this technology could create a market worth trillions of dollars, potentially leading to the largest tech industry in history [13].