生成式AI 一、 估值120亿美元的Thinking Machines，发布首篇研究博客 1. Thinking Machines发布首篇研究博客，解决LLM推理中的非确定性问题，核心是批次不变性； 2. 研究团队通过改进RMSNorm、矩阵乘法和注意力机制，实现完全可复现的推理结果，性能损失可接受； 3. 公司估值达120亿美元，创始团队多来自OpenAI，首款产品命名为Connection Machine。 https://mp.weixin.qq.com/s/2m_8ZPYBBIs3SuKoEJWiIw 二、 ChatGPT终于支持MCP了，一句Prompt即可全自动化 1. OpenAI宣布ChatGPT正式支持MCP(模型上下文协议)，Plus和Pro用户可一句Prompt实现自动化操作； 2. MCP实现了AI模型、工具和数据源的标准化交互，使不同模型能共享上下文，支持即插即用； 3. 用户可通过开启开发人员模式连接第三方服务(如Stripe)，完成复杂任务，但目前无法与其他ChatGPT功能同时使 用。 https://mp.weixin.qq.com/s/09par8_260tRn10VEEg ...

伊利亚金·奇斯列夫 《 关系5.0 》作者 孙涵晓 译者 我们爱上一个人的时候并不一定是爱这个人的全部，往往是爱他/她身上的某些特质，比如他/她的微 笑、手势、思维方式或者幽默感。有时，这样的人会让我们产生一种特殊的信任感，觉得他/她比其他 人都更值得信赖。这样的人能给我们带来愉悦的感受，与他/她聊天不仅有趣，还会受到启发。这样的 人还能传达出一种信息，即我们是被需要的、特别的、有才华的、有吸引力的、风趣的。与这样的人相 处时，我们会非常兴奋。所有这些因素结合在一起，让我们感觉到自己在这个人身边时更愉悦、更快 乐、更放松。 当遇到"伴侣候选人"时，大多数人都会像在亚马逊上购买商品一样对这个人做一些"调研"。我们会调查 这个人的背景，比如受教育程度、年龄、家庭、政治观点，甚至恋爱史等。我们会衡量这个人能否成为 一个完美的伴侣：能否在生活中保护、帮助我们，收入是否稳定，生育能力如何，以及能否与我们共筑 情感港湾，使两个人在其中都能放松身心、不断成长。 后来，我们对这个人产生了感情，因为他/她在某种程度上满足了我们的需求。你可以回想一段你或身 边的人经历的持续了至少一两年的关系，观察一下这段关系的每个阶段。一开始， ...

Group 1 - Nvidia launched the Rubin CPX GPU designed for long-context inference, capable of processing millions of tokens at once, supporting software development and video generation tasks [1] - The Rubin CPX GPU will be part of the Vera Rubin NVL144 CPX platform, providing 8 exaflops of AI computing power, which is 7.5 times that of the GB300 NVL72 system [1] - The system features 100TB of high-speed memory and 1.7 PB/s memory bandwidth, expected to be available by the end of 2026, promising unprecedented performance and efficiency for long-context tasks [1] Group 2 - Claude introduced a significant update allowing direct creation and editing of Excel, Word, PPT, and PDF files, outputting usable file formats [2] - The system is equipped with a private computing environment capable of writing code to generate various documents, supporting advanced data analysis and file operations [2] - This functionality is available to Max, Team, and Enterprise users, with Pro users to gain access in a few weeks, allowing file uploads or demand descriptions for Claude to process [2] Group 3 - Tencent released the AI CLI tool CodeBuddy Code and opened public testing for CodeBuddy IDE, supporting unlimited use of the DeepSeek model [3] - The system is designed for professional engineers, enabling natural language-driven development and operations, supporting multi-agent collaboration and deep integration with Git/CI/CD [3] - AI programming is evolving towards L4-level AI software engineering, with CLI becoming the foundational infrastructure, showing a 40% reduction in coding time and an increase in AI code review contributions from 12% to 35% [3] Group 4 - Kuaishou launched the AIGC super employee Kwali, capable of generating complete short videos from a single sentence, automating the entire process from script to publication [4] - The system is driven by a multi-agent framework, including intent parsing, script generation, shot matching, and editing, integrated with a material library [4] - Kwali allows independent manipulation of video elements on a timeline, enabling rapid video production that previously required multiple teams [4] Group 5 - Fellou CE created a "seamless continuum experience," achieving continuous interaction, task decomposition, and memory continuity [5] - The system supports cross-application execution, multimodal conversion, and dynamic workflow orchestration, successfully applied in travel planning and content creation [6] - Fellou CE introduced core features like "deep search" and "visual report generation," enhancing user control and productivity [6] Group 6 - Tencent released the open-source text-to-image model "Hunyuan Image 2.1," supporting native 2K images and achieving industry-leading performance in semantic understanding and text generation [7] - The model can handle prompts of up to 1000 tokens, generating detailed scene descriptions and supporting various styles [7] - Hunyuan Image 2.1 utilizes a 32x compression VAE and dual text encoders to improve training stability, reducing inference steps from 100 to 8 [7] Group 7 - Google launched an AI system to assist researchers in writing "expert-level" scientific software, combining large language models with tree search algorithms [8] - The system acts as a "mutation" engine during the search process, integrating and reorganizing research ideas from scientific literature [8] - It has shown exceptional performance in genomics, geospatial analysis, and neuroscience, marking a shift from one-time code generation to quantifiable scientific goal-oriented software evolution [8] Group 8 - a16z partners discussed that agents are not universal but systems composed of multiple agents, each specializing in specific tasks, leading to microservices and domain specialization [9] - Experts are becoming the biggest beneficiaries of AI, achieving a tenfold productivity increase, changing the nature of work rather than just output [9] - Each platform transition alters the abstract layer of human-computer interaction, with AI revolutionizing workflows and creating numerous vertical entrepreneurial opportunities [9] Group 9 - Elon Musk revealed that the Optimus humanoid robot will have near-human dexterity, costing around $20,000, with challenges mainly in hardware design [10] - The Tesla AI5 chip is expected to achieve a 40-fold performance leap over AI4, with software upgrades enabling Tesla cars to exhibit "awareness" [10] - The third-generation Starship will have a payload capacity exceeding 100 tons, aiming for full reusability next year, with human self-sufficiency on Mars projected within 25 years [10]

值9.10教师节，我们向躬耕教坛、传承文明的教育工作者致以敬意。 一场智能革命正在席卷全球，教育系统作为社会的一个关键子系统，在生成式AI的驱动下，教育本质几 近被重新定义，教育现象日益复杂，教育动因更加难以解析。 AI时代的教育之问II：教育变革 这不可避免地引出了一个核心命题：AI时代，我们如何培养下一代？教育的根本命题—— "培养什么 人、怎样培养人、为谁培养人" ——在智能浪潮的冲击下，需要重新回答。 国家、学界、产业界、社会家庭与个体都在适应、求索、变革，共同寻找面向未来教育的方案与路径。 长期以来，腾讯研究院密切关注 "AI与教育" 的时代课题，联合学界、业界专家与一线实践者，通过对 话、访谈、沙龙、趋势报告等形式，持续追踪AI教育的演进发展，针对教育焦虑、学习方式、人才培 养、就业转型、应用生态等议题进行剖析探讨。 借此教师节，腾讯研究院特将过往在AI教育领域的多方洞察与成果进行梳理与沉淀，形成此份内容集 锦，汇总产学研各界对智能时代教育本质、挑战机遇与未来图景的探索与思考。 希望这些内容能为关心教育未来的朋友们，提供有价值的参考。 【 腾讯研究院"AI+教育"合集 】 访谈 胡泳：《AI时代， ...

1. OpenAI CEO Sam Altman在博客中特别介绍了两位幕后核心研究员Jakub Pachocki和Szymon Sidor，称他们 是"完美互补的传奇搭档"； 2. Pachocki作为首席科学家负责制定公司宏观研究路线图，曾领导GPT-4预训练工作，并入选今年《时代》杂志百 大AI人物； 生成式AI 一、 Altman亲自发博客点赞，OpenAI这两大杰出人才究竟是谁？ 二、 Vidu Q1上线「参考生图」功能，人物、背景、道具随意组合 1. 国产AI工具Vidu Q1推出"参考生图"功能，能同时处理7张参考图，在一致性、真实性、美学等方面超越Flux Kontext，媲美谷歌Nano Banana； 3. 两人在2023年OpenAI"宫斗"事件中发挥关键作用，他们的辞职威胁成为员工大规模抗议的导火索，最终促使董事 会妥协迎回Altman。 三、 阿里发布最新语音识别模型Qwen3-ASR-Flash，能识别rap 1. 阿里发布语音识别模型Qwen3-ASR-Flash，支持11种语言和多种口音，能自动分辨语种、过滤噪声，并通过添 加上下文信息定制识别结果； 2. 在各项基准测试中，该模 ...

以下文章来源于公益时报 ，作者金锦萍 公益时报 . 《公益时报》是民政部主管的公益行业媒体，是民政部基金会年度报告信息发布的指定媒体之一。 今年的 "久久公益节" ，是中国公益慈善领域所面临的一次大考，也是一次检验其资源动员能力的重要契机。 我们期待："久久公益节"是慈善组织集体亮相的高光时刻，向所有关心相关社会议题的公众倡导对于某些社会问题的关切，传递基于公共性的对于人类命运 的"远忧"，呼吁着手来解决迫在眉睫的"近患"；"久久公益节"也是慈善组织向其众多利益相关者报告的时刻，告诉那些曾经信任支持过本组织的捐赠者、志 愿者和受益人，为了解决这些问题，慈善组织作出了哪些努力，并且在怎样的程度上解决或者缓解了现实困境，所采取的措施和方法究竟有什么样的效果， 还存在哪些不足和缺陷，有没有找到更好的方法和机制，以及还需要哪些支持和帮助。 "久久公益节"更是慈善组织审视自身的最佳时刻，从获得的信息反馈中知晓自身存在哪些弊病，从组织治理结构到项目实施能力，甚至社群构建能力等方面 获得第一手的批评与建议。 但是，我们也担忧："久久公益节"第一次尝试不动用配捐机制，那么这一次中国公益慈善组织将如何参与和有效动员社会力量共同 ...

2. 特斯拉决定从两种芯片架构切换到一种，所有芯片人才将专注于同一目标，马斯克形容为"理所当然的选择"； 3. AI5预计2025年下半年推出 ， 初期代工会在中国台湾，到后期会在美国 ，算力将 是 前 代 的10倍 ； AI6芯片 或 将由 三星在美 国 工厂生产。 https://mp.weixin.qq.com/s/XivsL8vf15x5BrcUx_yTQA 二、 Meta超级智能实验室的首篇论文来了，重新定义了RAG 生成式AI 一、 马斯克谈AI5 和 AI6 芯片的最新进展，称其为史诗般的芯片 1. 马斯克在X平台透露特斯拉AI5芯片设计团队完成评审，称其将是"史诗级"芯片，下一代AI6有望成为"迄今为止最好 的AI芯片"； 1. Meta超级智能实验室推出REFRAG框架， 提 出 重新定义RAG技术，最高将首字生成延迟(TTFT)加速30倍，突破 长上下文计算冗余瓶颈； https://mp.weixin.qq.com/s/ay0nTvxTWqevXBxLczyLYA 四、 微软开源3大突破AI Agent模型，140亿参数超越DS-R1？ 1. 微软研究院开源推理模型rStar2-Agen ...

【 精彩观点整理 】 本文为基于 胡泳 访谈的文字整理 我们可能无法完全量化人工智能对人类整体智力的影响，但是认知卸载导致特定认知技能下降的担 忧，比如个体记忆能力的减退，则是完全合理的。 人机协作有潜在的危险，就是丧失人类的思想主体性。就像"骆驼挤进帐篷"，等到骆驼整个身躯都 挤进来以后，人就被完全顶出帐篷了。 就当下发展来看，大模型还远远不能够达到我们作为一个"人"所拥有的那种全部智能和意识… （大 模型的智能） 几乎只集中在语言智能和部分逻辑推理智能上。 使用人工智能需要有一个前提条件：TA必须是一个具备超高信息素养的人。 我们的教育体系应该培养会用人工智能工具或者任何工具生产出东西的能力。再往上是沟通协作、 批判性思维、创造力，还有自信心，这些是在人工智能时代所需要的关键技能。 人工智能时代到来之后，应该设想一个概念，叫做"无分数学习"，把评分体系或者说成绩体系，置 换到更具有挑战性的学习任务，并且让学生的注意力回归到更有意义的事情上面。 要做到"人的归人，机器的归机器"，不要混淆两者的边界。 我们不该只问"人工智能能为人类做什么"，还要问"人工智能正在对人类做什么"。 AI之于人类，就像"骆驼挤进帐 ...

Group 1 - Anthropic has implemented a policy to restrict access to its Claude service for entities with majority ownership by Chinese capital, citing legal, regulatory, and security risks [1] - The restriction also applies to entities from countries considered adversaries, such as Russia, Iran, and North Korea, with expected global revenue impact in the hundreds of millions of dollars [1] Group 2 - AI Key, an external AI assistant hardware for iPhone, sold out within 7 hours of launch, priced at $89, but is seen as redundant given the existing capabilities of iPhones [2] - The trend of AI hardware startups is viewed as short-lived, with future value lying in integrating AI as a system attribute rather than a standalone function [2] Group 3 - Tencent's "Hunyuan Game" platform has launched version 2.0, introducing features like game-to-video generation and custom model training [3] - The new AI capabilities allow users to create high-quality dynamic videos from game images and descriptions, significantly lowering the barrier for custom model training [3] Group 4 - Alibaba has released the Qwen3-Max-Preview model, boasting over a trillion parameters, outperforming competitors in various benchmarks [4] - The model supports over 100 languages and offers a maximum context of 256k, with a tiered pricing model based on token usage [4] Group 5 - ByteDance's Seed team has introduced Robix, a unified "robot brain" that integrates reasoning, task planning, and human-robot interaction [5][6] - Robix employs a hierarchical architecture to separate high-level decision-making from low-level control, enabling dynamic reasoning and execution [6] Group 6 - Rokid's AR+AI glasses sold 40,000 units within 5 days of launch, highlighting their lightweight design and user-friendly features [7] - The product includes customizable audio and translation capabilities, and Rokid has opened its SDK for developers, expanding its global reach [7] Group 7 - Anthropic has agreed to a $1.5 billion settlement in a copyright lawsuit involving the illegal download of 7 million books, marking a significant moment in AI and copyright disputes [8] - The settlement involves compensation for approximately 500,000 books, averaging $3,000 per book, while the financial impact is considered manageable relative to Anthropic's recent funding and revenue [8] Group 8 - The Sensor Tower report indicates that global downloads of generative AI applications reached nearly 1.7 billion in the first half of 2025, with in-app purchase revenue of $1.9 billion, reflecting a 67% quarter-over-quarter growth [10] - The report highlights a demographic shift, with female users of AI assistants exceeding 30%, and emphasizes the competitive pressure on vertical applications [10] Group 9 - OpenAI's recent paper defines "hallucination" in AI models and identifies its root causes, suggesting that current evaluation methods encourage guessing rather than acknowledging uncertainty [11] - The paper proposes a revised evaluation approach that penalizes confident errors more than uncertainty, aiming to improve the reliability of AI responses [11]

Group 1: AI Models - Grok Code Fast 1 developed by xAI is highlighted as a significant model [3] - LongCat-Flash introduced by Meituan showcases advancements in AI modeling [3] - Claude's performance degradation rollback indicates challenges in maintaining model efficiency [3] - Shanghai AI Laboratory's Shusheng·Wanshang 3.5 represents a new iteration in AI models [3] - Kimi K2-0905 from Moonlight Dark Side is noted for its innovative features [3] - Kuaishou's new multimodal model reflects the trend towards integrating various data types [3] Group 2: AI Applications - Meta's third-party AI collaboration emphasizes the importance of partnerships in AI development [3] - OpenAI's GPT-realtime application showcases real-time AI capabilities [3] - Claude's user data utilization raises discussions on data privacy and usage [3] - Tencent's Hunyuan-MT-7B highlights advancements in machine translation [3] - Step-Audio 2 mini from Jiyue Xingchen represents innovation in audio processing [3] - Hyodol's AI doll for elderly users indicates a growing market for AI in healthcare [3] - Multi-department and platform AI content identification reflects regulatory trends [3] - Tsinghua's embodied reinforcement framework shows advancements in AI learning [3] - Google's "Detailed Webpage" feature enhances user experience through AI [3] - Tencent's 3D world model indicates a shift towards immersive AI applications [3] - Runway's cross-domain robots illustrate the versatility of AI in various fields [3] Group 3: Technology and Research - Tsinghua's robot ping pong showcases the intersection of robotics and AI [5] - UCLA's AI brain-machine interface represents cutting-edge research in human-computer interaction [5] - The machine wolf project from 93rd Military Parade indicates military applications of AI [5] - RoboScience's RoboMirage simulation reflects advancements in AI-driven simulations [5] - Tesla's "Golden Pillar" project highlights the integration of AI in automotive technology [5] - Shanghai AI Laboratory's research on AI evolution in scientific fields indicates ongoing innovation [5] Group 4: Capital and Investment - OpenAI's acquisition of Statsig signifies strategic growth through mergers [5] - Anthropic's $13 billion financing round indicates strong investor confidence in AI [5] - OpenAI's recruitment of the Alex team reflects competitive talent acquisition in the industry [5] Group 5: Events and Trends - The Werewolf game battle involving GPT-5 indicates the application of AI in entertainment [5] - xAI's engineer defection raises concerns about talent retention in AI companies [5] - Meta's new executive departure highlights challenges in leadership stability [5] - Salesforce's 4,000 layoffs reflect broader trends in workforce adjustments within tech [5] Group 6: Perspectives and Insights - a16z's insights on AI hardware entry points suggest strategic investment opportunities [5] - DeepSeek's details on V3/R1 training provide valuable information for AI model development [5] - Tesla's grand blueprint outlines ambitious future plans for AI integration [5] - The use of AI by students in U.S. universities indicates a growing acceptance of AI in education [5] - OpenAI experts' strategies on AI PM reflect evolving management practices in tech [5] - OpenAI's leadership guide offers insights into effective management in AI-driven environments [5]