生成式AI 一、 OpenAI强化微调终于上线，几十个样本可轻松打造AI专家 1. OpenAI正式发布RFT(强化微调)功能，通过思维链推理和专属评分机制，可用极少样本快 速提升模型在特定领域的专业表现； 2. RFT主要应用于三大场景：指令转代码、文本精华提取、复杂规则应用，已有ChipStack 等多家公司取得显著成效； 3. 实施RFT前必须创建评估体系，需要明确任务定义和强化评分方案，避免模棱两可的任务 目标。 https://mp.weixin.qq.com/s/c7RfeoWNwh3NZDeuTCXXLw 二、 Gemini 2.5实现视频理解重大突破：一口气处理6小时视频 1. Gemini 2.5 Pro突破视频处理长度限制，通过低媒体分辨率技术可处理长达6小时视频， 在多个学术基准测试中创下新纪录； 2. 实现视频内容与代码无缝结合，能将视频直接转化为交互式网页应用、p5.js动画等创新应 用形式； 3. 具备精准的视频片段检索和时序推理能力，可实现复杂场景计数、时间戳定位等高级分析 功能。 https://mp.weixin.qq.com/s/FkaOacVuVCS7wzny5l1jFQ ...

| 类别 | Top关键词 | 主体 | | --- | --- | --- | | 算力 | OpenAI for Countries | OpenAI | | 算力 | 网络提速技术 | DeepSeek、 | | | | 腾讯 | | 模型 | Gemini 2.5 Pro（I/O版） | 谷歌 | | 模型 | Medium 3 | Mistral AI | | 模型 | Nemotron开源模型 | 英伟达 | | 模型 | V2数学推理模型 | DeepSeek | | 应用 | Claude整合功能 | Anthropic | | 应用 | NotebookLM中文支持 | Google | | 应用 | 独立AI应用 | Meta | | 应用 | 合作氛围编程 | 苹果、 | | | | Anthropic | | 应用 | Omni-Reference | Midjourney | | 应用 | 参考图功能 | Runway | | 应用 | PDF渲染器 | Grok | | 应用 | V4.5正式上线 | Suno | | 应用 | Parakeet 语音识别 | 英伟达 | | 应用 ...

本期，我们非常荣幸地于4月16日邀请虞晶怡老师，为我们开启一次AI的思想远航。 徐一平 腾讯研究院 高级研究员 王强 腾讯研究院 资深专家 以生成式AI为代表的新技术浪潮日新月异，正带来一场深刻的技术、商业与社会变革，推动人类社会从 信息社会向智能社会转变。全世界热切期待AI到来的同时，也非常关心人工智能将带来哪些新机遇、新 挑战。 为此，我们发起了一项 《AI & Society 百人百问》 研讨，广泛邀请AI技术大咖、AI独角兽创始人、AI 投资人，以及社会学家、心理学家、国际关系专家、科幻作家等，用多元视角，深入研讨人工智能技术 引发的广泛影响，发掘AI时代的共识和非共识，共同推动人工智能始终朝着"助人发展，与人为善"的方 向可持续发展。 虞晶怡，上科大讲席教授、副教务长、信息学院院长。在加入上海科技大学前，他任职美国特拉华大学计算机与信息科学系正教 授。他于2000年获美国加州理工学院应用数学及计算机学士学位， 2005年获美国麻省理工大学计算机与电子工程博士学位。他 长期从事计算机视觉、计算成像、计算机图形学、生物信息学等领域的研究工作。他是IEEE Fellow、OSA Fellow、美国NSF ...

Core Viewpoint - The article discusses the transformative impact of recommendation systems in the digital age, questioning whether these systems empower individual choice or dictate user behavior, ultimately shaping personal destinies [2][4]. Group 1: Recommendation Systems and Their Influence - Recommendation systems are pervasive in daily life, influencing choices in music, movies, and travel through personalized suggestions [3][7]. - Netflix's approach to user experience is centered around the idea that "everything is a recommendation," tailoring content based on user preferences and viewing history [3][4]. - The rise of recommendation engines is likened to a revolution in personalized choice, raising questions about autonomy and the nature of decision-making in the age of AI [4][5]. Group 2: The Role of Algorithms - Algorithms are crucial for enhancing user experience by providing tailored recommendations, which can lead to increased engagement and satisfaction [6][7]. - The effectiveness of recommendation systems is linked to the volume and quality of data they process, with more data leading to better algorithm performance [6][7]. - TikTok's recommendation algorithm has been recognized for its ability to promote diverse content, allowing lesser-known creators to gain visibility alongside popular ones [8][12]. Group 3: Evaluation Metrics for Recommendations - Key metrics for assessing recommendation systems include precision, diversity, novelty, serendipity, explainability, and fairness [9][10]. - Precision measures the relevance of recommended content to user interests, while diversity ensures a broad range of topics is covered [9][10]. - Fairness has emerged as a critical metric, addressing biases in recommendations that may disadvantage certain groups or content creators [10][11]. Group 4: Addressing Fairness and Bias - The concept of "responsible recommendation" has gained traction, focusing on eliminating systemic biases in recommendation systems and ensuring equitable treatment across different demographics [14][15]. - Companies like Amazon, Netflix, and Spotify are actively working to incorporate fairness and transparency into their algorithms to avoid biases and promote diverse content [17][18]. - The need for transparency in recommendation logic is emphasized, allowing users to understand the basis for recommendations and fostering trust in the system [14][17]. Group 5: From Recommendation to Self-Discovery - The evolution of recommendation systems into self-discovery engines is highlighted, where users can gain deeper insights into their preferences and identities through tailored suggestions [19][20]. - Empowerment through better choices and the ability to explore new interests is a key aspect of this transformation, enhancing user engagement and self-awareness [20][21]. - Ultimately, understanding oneself and one's aspirations may increasingly depend on the interactions with intelligent recommendation systems [21].

建设中国国家版本馆，是以习近平同志为核心的党中央作出的重大决策，是文明大国建设的基础工程， 是功在当代、利在千秋的标志性文化工程。中国国家版本馆（国家版本数据中心）担负着赓续中华文 脉、坚定文化自信、展示大国形象、推动文明对话的重要使命，是中华版本典藏中心、展示中心、研究 中心、交流中心和国家出版信息服务中心。 本次 "文脉之光"文创设计大赛 ，旨在让沉睡在典籍中的文化密码"活"起来：通过开发文具、数码周边 等创意产品，让古籍纹样走进现代生活；借助AR技术让版本"开口讲故事"；用当代设计语言重构传统典 籍的版式美学。活动将推动文明"基因库"成为创新"孵化器"，让中华文脉在设计师的创意中焕发新生， 擦亮国家文化名片，为文化产业注入新动能。 组织机构 主办单位： 中国国家版本馆 执行单位： 阅途文化集团有限公司 广东阅途文化传播有限公司 活动对象 面向全社会广泛征集，各高校艺术院系师生、独立设计师、具有一定艺术设计基础的社会各界人士、创 意设计团队或机构均可报名参赛。 参赛作品设计手法、表现形式、材质、工艺、造型、尺寸、品类等不限，鼓励参赛者以创新视角和多元 表达，深入挖掘版本馆文化内涵，彰显版本馆特色，充分展现 ...

Group 1: Generative AI Developments - Google Gemini 2.5 Pro has achieved top rankings in LMeana, outperforming Claude 3.7 in programming performance, with significant enhancements in coding capabilities [1] - ComfyUI has introduced native API node functionality, supporting over 10 model series and 62 new nodes, allowing direct calls to paid models like Veo2 and Flux Ultra [2] - Cognition AI has open-sourced the Kevin model with 32 billion parameters, achieving a 65% average accuracy on the KernelBench dataset and a 1.41x speedup in kernel code generation [3] Group 2: Strategic Initiatives - Cursor Pro and Gemini Pro are offering one-year free access to students, potentially saving around 2000 RMB, as part of a strategy to cultivate future user habits [4][5] - Tencent Yuanbao has launched a conversation grouping feature, allowing users to create folders by theme and set independent prompts for each group [6] - Tencent Yuanbao has upgraded its text-to-image generation capabilities, enhancing image quality and consistency with user-friendly input [7] Group 3: AI in Scientific Research - Anthropic has initiated the AI for Science program, providing up to $20,000 in API credits to selected researchers to accelerate scientific discoveries [8] - The program supports all Claude series models, focusing on applications in biological systems, genetic data, drug development, and agricultural productivity [8] Group 4: Robotics and AI Models - Tsinghua ISRLab and Star Motion Era have jointly developed the VPP robot model, which has been open-sourced and recognized for its advanced capabilities in task execution [9][10] - The VPP model can learn from human motion data and perform over 100 dexterous tasks in real-world scenarios, showcasing strong interpretability and optimization abilities [10] Group 5: Industry Insights - A warning from a University of Toronto professor highlights that AI is making humans increasingly "irrelevant" in economic, cultural, and social domains, as it becomes cheaper and more reliable [11] - Bolt.new has rapidly scaled its annual revenue from $700,000 to $20 million in two months, focusing on browser-based rapid web application development [12] - The majority of Bolt's users are not developers but product managers, designers, and entrepreneurs, indicating a shift in the user base for software development tools [12]

Core Viewpoint - The article discusses the rise of Model Context Protocol (MCP) as a unifying tool invocation protocol in the AI industry, highlighting its rapid adoption and the excitement surrounding it, while also addressing its limitations and the need for realistic expectations regarding its applicability across different scenarios [3][4][5]. Summary by Sections What is MCP? - MCP is an open technical protocol designed to standardize interactions between large language models (LLMs) and external tools and services, functioning as a universal translator for AI models [5][6]. Why is MCP Needed? - Prior to MCP, AI tool invocation faced two main issues: fragmented interfaces requiring custom code for each combination and inefficient development processes [6][8]. MCP's Functionality - MCP employs a universal language format (JSON - RPC) allowing developers to interact with all tools supporting this protocol after a single learning phase [8][10]. MCP's Architecture - MCP consists of three core components: MCP Host (execution environment), MCP Client (communication hub), and MCP Server (service endpoint), facilitating smooth communication between AI models and external services [11][15]. MCP's Development Challenges and Market Chaos - The rapid growth of MCP has led to a chaotic market with many tools lacking practical value, as many developers rushed to create MCP-compatible services without thorough testing [24][34]. MCP's Limitations - While MCP has been beneficial for local client applications, it faces challenges in server-side and cloud applications due to its dual-link mechanism, which complicates implementation and maintenance [28][29]. Market Confusion - The current MCP market is characterized by low usability, with many tools failing to deliver real value, leading to inefficiencies in tool selection and usage [34][35]. MCP's Role in the AI Ecosystem - MCP is not a one-size-fits-all solution; it is a communication protocol that does not dictate how tools are selected or used, emphasizing the need for a collaborative approach among various AI components [39][40]. Future Directions - The article suggests that MCP's evolution may lead to a more streamlined and valuable tool ecosystem, as the market naturally selects for quality and utility over time [36][46].

生成式AI 一、 刚刚，OpenAI放弃营利性转型！奥特曼：非营利组织继续掌控 1. OpenAI放弃完全营利性转型，将由非营利组织继续控制，同时营利性机构转为公益公司（PBC）； 2. 公司架构调整后取消利润上限制度，采用常规股权结构，非营利组织将成为PBC主要股东； 3. 承诺继续专注AGI发展造福人类使命，并计划开源部分高性能模型。 https://mp.weixin.qq.com/s/Z1bl0zfwNXeEcoDZFtpWmQ 二、 公开一切，优于DeepSeek-R1？英伟达开源Llama-Nemotron家族 1. 英伟达发布Llama-Nemotron开源模型家族，包含8B到253B三种规格，支持动态切换推理模式，遵循 开放商业许可； 2. LN-Ultra运用Puzzle框架和FFN融合技术优化部署效率，在推理性能和吞吐量上超越DeepSeek-R1； 3. 通过Qwen和DeepSeek-R1教师模型支持，结合多阶段训练和强化学习，全面提升模型推理与通用对话 能力。 https://mp.weixin.qq.com/s/Ofw7l6XPNNinXvFReGI3vw 三、 Grok 增加PD ...

引言： OpenAI为何在收购与结构调整中双线推进？ 白一 独立科技观察者 2025年5月6日，OpenAI宣布放弃全面营利化重组方案，将营利性子公司转型为公益公司（PBC），由非 营利组织继续持有控制权。 这一结构调整背后，实质上是对其快速商业化扩张节奏的制度性回应。 过 去两年，OpenAI持续通过收购和新业务布局，加速构建平台级AI应用生态，商业化步伐显著加快。 此 时宣布结构调整，既是回应监管和社会对其"逐利化"倾向的质疑，也是为下一阶段收购与扩张创造治理 前提。 尽管全面盈利化看似更有利于资本进入和商业操作，OpenAI却选择了保留非营利组织控制权的PBC结 构。原因在于，PBC制度允许公司在追求利润的同时将社会使命写入治理框架，而非营利母公司继续控 股，则进一步确保公司战略不被短期财务回报所驱动。这一治理安排既回应了外部对其使命偏移的质 疑，也保留了资本融资、员工激励和并购操作所需的灵活性。 可以说，OpenAI试图在"制度可信 度"与"商业扩张性"之间建立一种长期可持续的平衡机制。 换句话说，如果说一系列收购是OpenAI打通"从底座到入口"的平台化布局工具，那么组织架构的调整 就是对其使命合法 ...