Investment Rating - The report maintains an "Overweight" rating for the computer industry [1] Core Viewpoints - The industry is experiencing rapid advancements in AI models, with significant updates from major players like OpenAI and Zhiyuan, indicating a strong trend towards commercialization of AI applications [3][4] - The recent performance of the computer index has lagged behind major indices, suggesting a need for cautious optimism in the short term [2] Summary by Sections Market Review - In the past week (December 8-12), the Shanghai Composite Index fell by 0.34%, while the ChiNext Index rose by 2.74%. The computer index dropped by 1.14%, underperforming the Shanghai Composite by 0.80 percentage points and the ChiNext by 3.88 percentage points [2] Weekly Insights - OpenAI launched the new GPT-5.2 model, enhancing capabilities in information retrieval, writing, translation, and programming tasks, with improved performance in enterprise applications [3] - Zhiyuan released the GLM-4.6V series of multimodal models, significantly increasing context window size and integrating tool invocation capabilities into visual models [3] AI Application Acceleration - Google introduced the Gemini Deep Research AI agent, marking a significant step towards industrial application of AI with enhanced web search capabilities [4] - Zhiyuan also launched the AutoGLM model, which can operate mobile applications, indicating a shift towards more interactive AI agents [4] - Alibaba restructured its business to focus on creating a super app, aiming to become a primary AI assistant across various platforms [4] Investment Recommendations - Suggested companies to watch include those in computing power such as Cambrian, Haiguang Information, and Zhongke Shuguang, as well as AI application firms like Kingsoft Office and iFlytek [8]

[Table_Page] 跟踪分析|传媒 证券研究报告 [Table_Summary] [Table_Title] 传媒行业•AI 周度跟踪之四十六 OpenAI 发布 GPT-5.2，谷歌开源深度研究 Agent | [Table_Gr ade] 行业评级 | 买入 | | --- | --- | | 前次评级 | 买入 | | 报告日期 | 2025-12-14 | 核心观点： [Table_PicQuote] 相对市场表现 -24% -14% -5% 5% 14% 24% 12/24 02/25 05/25 07/25 09/25 12/25 传媒 沪深300 | [分析师： Table_Author]旷实 | | | --- | --- | | | SAC 执证号：S0260517030002 | | | SFC CE No. BNV294 | | | 010-59136610 | | | kuangshi@gf.com.cn | | 分析师： | 廖志国 | | | SAC 执证号：S0260525060001 | | | 021-38003665 | | | liaozhiguo@gf.com.c ...

在全球人工智能领域竞争快速升温的当下，谷歌与 OpenAI 再次在同一天抛出重磅更新，令整个行业的注意力高度集中。 昨天夜里，谷歌发布了全新"重新构想"的 Gemini Deep Research 版本，并首次开放了嵌入式研究智能体 API。 而几乎同时，OpenAI 正式发布了备受期待的 GPT-5.2（代号 Garlic）。两家公司围绕智能体（Agent）未来、基础大模型能力边界以及应用生态主导权的 竞争，正进入一个前所未有的焦灼阶段。 这一次，谷歌和 OpenAI 的攻防几乎精确地踩在同一时间窗口，让外界得以清晰观察这两家全球 AI 巨头之间的战略对抗节奏。 1 谷歌推出全新 Deep Research Agent 谷歌推出的全新 Gemini Deep Research 工具是一款智能 Agent，能够整合海量信息并处理提示信息中大量的上下文数据。谷歌表示，客户使用 Deep Research Agent 执行的任务范围广泛，从尽职调查到药物毒性安全研究均有涉及。 谷歌还表示，很快会将这款全新的 Deep Research Agent 集成到其各项服务中，包括谷歌搜索、谷歌财经、Gemini 应用以及 ...

谷歌推出迄今最强的深度研究型Agent——GeminiDeep Research的"重制版"，试图定义Agent 的基础设施级入口。未来 可能不是用户"谷歌一下"，而是你的Agent替你谷歌一切。 硬·AI 作者 |卜淑情 编辑 | 硬 AI 在全球AI叙事走向"Agent时代"的关键节点，谷歌选择了一个颇具戏剧性的发布时机。 周四，就在OpenAI端出备受期待的 GPT-5.2（内部代号 Garlic）之日，谷歌同步推出了迄今最强的深度 研究型Agent—— Gemini Deep Research 的"重制版"，并宣称其基于旗下最先进的Gemini 3 Pro模型。 同日，DeepMind还 宣布将在英国建立首个自动化研究实验室 ，利用AI与机器人加速材料科学实验。 这不是"撞车"，更像一场精心策划的阳谋：在竞争对手聚焦全球目光时，谷歌以一款更具战略意味的产品 回应——将Agent 推向操作系统级能力。 01 从"写报告"到"嵌入应用" 谷歌试图定义Agent的基础设施级入口 全新 Gemini Deep Research已不是传统意义上的"自动写研究报告"的工具，它被定位为： 换句话说：未来不是用户"谷 ...

Core Insights - OpenAI's CEO Sam Altman expressed strong optimism about the company's R&D and product roadmap during the launch of GPT-5.2, despite facing unprecedented competition from rivals like Google and Anthropic [2][3] - The release of GPT-5.2 has been positioned as a significant advancement, with performance metrics surpassing competitors, particularly in professional applications [4][5] Product Performance - GPT-5.2 was launched with three different model tiers: Instant, Thinking, and Pro, achieving benchmark scores that outperformed competitors like Gemini 3 PRO and Claude Opus 4.5 [4] - In the GPQA Diamond evaluation, GPT-5.2 scored 92.4%, a notable increase from GPT-5.1's 88.1% and higher than Gemini 3 PRO's 91.9% [4] - The model achieved a perfect score in the AIME 2025 competition, showcasing its capabilities in advanced mathematics [4] Competitive Landscape - Google launched its Gemini Deep Research product shortly before GPT-5.2, emphasizing its competitive stance in the AI model market [10][12] - Gemini Deep Research reportedly offers similar performance to GPT-5 Pro at a significantly lower cost, highlighting Google's focus on cost-effectiveness and efficiency [12] - OpenAI's reliance on computational power for GPT-5.2 raises concerns about sustainability and market competitiveness, especially as rivals demonstrate more cost-effective models [7][12] User Experience and Feedback - Users have praised GPT-5.2 for its practical applications in tasks such as data analysis and project management, earning it titles like "strongest AI worker" [7] - However, some users reported slower response times in the Thinking and Pro models compared to previous versions, raising concerns about efficiency [8] - Despite its strengths, GPT-5.2 still encounters issues with common knowledge questions, indicating areas for improvement [9] Future Developments - OpenAI plans to continue enhancing its offerings, with Altman hinting at upcoming features and models, including a new model named "Garlic" [12] - The competitive landscape is expected to evolve further, with other players like Meta and DeepSeek also preparing to launch new products [12][13]

Core Viewpoint - Google has launched the most powerful deep research agent to date, Gemini Deep Research, aiming to redefine the infrastructure-level entry for agents, suggesting a future where users rely on their agents to conduct searches instead of manually searching themselves [1][7]. Group 1: Product Launch and Strategic Timing - Google strategically timed the release of Gemini Deep Research to coincide with OpenAI's anticipated launch of GPT-5.2, showcasing its advanced capabilities based on the Gemini 3 Pro model [3][4]. - The launch is seen as a calculated response to OpenAI's developments, positioning Gemini Deep Research as a product with significant strategic implications [4]. Group 2: Features and Capabilities - Gemini Deep Research is not just a tool for generating research reports; it is designed to handle larger contexts, process vast amounts of information, and perform long-chain reasoning tasks that can last for minutes or hours [5]. - The introduction of the Interactions API allows developers to easily integrate Deep Research into their applications, effectively packaging search, multi-step reasoning, and evaluation into an operating system-level service [5]. Group 3: Integration and Future Vision - Deep Research will gradually be integrated into various Google services, including Google Search, Google Finance, Gemini applications, and NotebookLM [6]. - The vision is that users will no longer need to perform searches themselves; instead, their agents will handle all search tasks [7]. Group 4: Addressing Challenges in AI - Google aims to tackle the significant challenge of hallucination rates in AI agents, claiming that Deep Research benefits from the higher factual accuracy of the Gemini 3 Pro model, which helps reduce distortions in long-chain reasoning tasks [8]. Group 5: Performance Benchmarks - Google has introduced new benchmarks, such as DeepSearchQA, to test multi-step information retrieval and has made these benchmarks open source [9]. - In benchmark tests, the new agent outperformed competitors, although OpenAI's ChatGPT 5 Pro showed close performance, particularly in the BrowserComp test [10]. Group 6: Competitive Landscape - The simultaneous announcements from Google and OpenAI mark a direct competition, with Google aiming to establish a foothold in the rapidly evolving agent landscape [11]. - The competition has shifted from model superiority to who can become the foundational infrastructure for future information access methods [12].

Core Insights - OpenAI and Google are both making significant updates in the AI space, with Google launching an enhanced version of Gemini Deep Research aimed at reducing hallucinations and excelling in complex information retrieval and analysis tasks [1][3][10]. Group 1: Gemini Deep Research Enhancements - The enhanced Gemini Deep Research is built on Gemini 3 Pro and will soon be integrated into various Google services such as Google Search, NotebookLM, Google Finance, and the upgraded Gemini App [3][8]. - This version of Gemini Deep Research can perform iterative reasoning, allowing it to generate queries, read and integrate search results, and identify knowledge gaps, significantly improving its web search capabilities [10][12]. - In benchmark tests like HLE, BrowseComp, and DeepSearchQA, the enhanced model has achieved state-of-the-art (SOTA) results, showcasing its superior performance in complex research tasks [10][12]. Group 2: DeepSearchQA Benchmark - Google has released the DeepSearchQA benchmark dataset to provide a more comprehensive evaluation standard for deep search and research tasks, addressing the limitations of existing benchmarks [5][12]. - The dataset includes 900 manually designed causal chain tasks from 17 domains, requiring detailed answer sets, which better measure the model's multi-step reasoning and information fusion capabilities [12]. Group 3: Interactions API - Google has introduced the Interactions API, designed to provide a unified interface for developers to interact with Gemini 3 Pro and Deep Research agents [6][16]. - This API is particularly suited for scenarios requiring multi-step reasoning, tool invocation, and long-term task execution, enhancing the capabilities of existing models [17][18]. - The Interactions API simplifies workflows and adapts better to developer environments by expanding the core capabilities of content generation and supporting server-side state, interpretable data models, and remote tool support [18].

Core Insights - Google has launched Gemini Deep Research, an advanced AI research agent, following the release of OpenAI's GPT-5.2, marking a significant step towards industrial application of AI in complex research tasks [1][3] Group 1: Gemini Deep Research Features - Gemini Deep Research is built on Gemini 3 Pro and is optimized for long-cycle content collection and synthesis, achieving a 40% reduction in hallucination rates compared to previous models, making it Google's most factually accurate model to date [3] - The AI agent utilizes multi-step reinforcement learning to navigate complex information environments with higher precision, enabling deep information mining through iterative planning of research paths [3] - In benchmark tests, Gemini Deep Research scored 46.4% in Google's new benchmarks and performed comparably to GPT-5 Pro in BrowseComp, while costing approximately one-tenth of the latter [3] Group 2: Applications and Industry Impact - The AI agent has demonstrated significant value in various industries, such as automating early information collection in financial services, enhancing research efficiency by integrating market signals and compliance risks [4] - In biotechnology, Axiom Bio has utilized the AI for literature analysis related to drug toxicity prediction, resulting in deeper and more granular research, thereby accelerating drug development processes [4] - The AI's strong information integration capabilities have also improved decision-making in market research [4] Group 3: Interactions API and Future Developments - The newly launched Interactions API allows developers to leverage Gemini Deep Research to create next-generation automated research tools, featuring capabilities like unified information synthesis and structured output [5] - Future upgrades will include native chart output capabilities and expanded support for custom data sources, with plans to launch Deep Research services on the Vertex AI platform for broader industry application [6]

Core Insights - Google has launched a reimagined version of its research agent, Gemini Deep Research, based on the Gemini 3 Pro model, which enhances its capabilities beyond just producing research reports [1] - The new agent allows developers to integrate Google's research capabilities into their applications through the Interactions API, marking a significant step towards agentic AI [1][3] - Gemini Deep Research is designed to synthesize large amounts of information and is utilized for various tasks, including due diligence and drug toxicity safety research [2] Product Features - The new deep research agent will be integrated into several Google services, such as Google Search, Google Finance, the Gemini App, and NotebookLM, indicating a shift towards AI-driven information retrieval [3] - Gemini 3 Pro is touted as Google's "most factual" model, specifically trained to reduce hallucinations during complex tasks, which is crucial for maintaining accuracy in long-running decision-making processes [3][4] Benchmarking and Competition - Google has introduced a new benchmark called DeepSearchQA to evaluate agents on complex, multi-step information-seeking tasks, which has been open-sourced [5] - The performance of Gemini Deep Research was tested against other benchmarks, including Humanity's Last Exam and BrowserComp, where it outperformed competitors, although OpenAI's ChatGPT 5 Pro closely followed [7] - On the same day as Google's announcement, OpenAI launched its GPT 5.2 model, which claims to surpass its rivals, including Google, on various benchmarks [8][9]

Core Insights - The article discusses the vulnerabilities of large language model (LLM)-based search agents, emphasizing that while they can access real-time information, they are susceptible to unreliable web sources, which can lead to the generation of unsafe outputs [2][7][26]. Group 1: Search Agent Vulnerabilities - A real-world case is presented where a developer lost $2,500 due to a search error involving unreliable code from a low-quality GitHub page, highlighting the risks associated with trusting search results [4]. - The research identifies that 4.3% of nearly 9,000 search results from Google were deemed suspicious, indicating a prevalence of low-quality websites in search results [11]. - The study reveals that search agents are not as robust as expected, with a significant percentage of unsafe outputs generated when exposed to unreliable search results [12][26]. Group 2: SafeSearch Framework - The SafeSearch framework is introduced as a method for automated red-teaming to assess the safety of LLM-based search agents, focusing on five types of risks including harmful outputs and misinformation [14][21]. - The framework employs a multi-stage testing process to generate high-quality test cases, ensuring comprehensive coverage of potential risks [16][19]. - SafeSearch aims to enhance transparency in the development of search agents by providing a quantifiable and scalable safety assessment tool [37]. Group 3: Evaluation and Results - The evaluation of various search agent architectures revealed that the impact of unreliable search results varies significantly, with the GPT-4.1-mini model showing a 90.5% susceptibility in a search workflow scenario [26][36]. - Different LLMs exhibit varying levels of resilience against risks, with GPT-5 and GPT-5-mini demonstrating superior robustness compared to others [26][27]. - The study concludes that effective filtering methods can significantly reduce the attack success rate (ASR), although they cannot eliminate risks entirely [36][37]. Group 4: Implications and Future Directions - The findings underscore the importance of systematic evaluation in ensuring the safety of search agents, as they are easily influenced by low-quality web content [37]. - The article suggests that the design of search agent architectures can significantly affect their security, advocating for a balance between performance and safety in future developments [36][37]. - The research team hopes that SafeSearch will become a standardized tool for assessing the safety of search agents, facilitating their evolution in both performance and security [37].