![[7558deef-341f-4e39-bac1-8fe797dc0bb7.mp3]]

国安部：微量虚假数据即可污染AI，引发多重风险

简报：

• 国家安全部于8月5日发文指出，数据是AI的核心要素，一旦被污染，可能导致模型决策失误甚至系统失效。
• 研究显示，通过“数据投毒”（如篡改、虚构），训练数据中仅含0.01%的虚假文本，就能使模型输出的有害内容增加11.2%。
• 数据污染会造成“递归污染”，即AI生成的虚假内容成为后续模型的训练源，导致错误信息逐代累积，扭曲模型认知能力。
• 数据污染可能引发金融市场操纵、误导社会舆论、生成错误医疗建议等一系列现实风险。
• 官方提出的应对方案包括：加强源头监管、强化全生命周期风险评估、对受污染数据进行末端清洗修复。

相关链接：

• 国家安全部：警惕人工智能数据投毒，0.01% 虚假文本可致有害输出增加 11.2%

Ministry of State Security: Minute Amounts of False Data Can Contaminate AI, Leading to Multiple Risks

Briefing:

• The Ministry of State Security (MSS) stated in an article published on August 5th that data is the core element of AI, and once contaminated, it can lead to model decision errors and even system failure. ⚠️
• Research indicates that "data poisoning" (e.g., tampering, fabrication) with as little as 0.01% false text in training data can increase the harmful output of models by 11.2%. 📈
• Data contamination can lead to "recursive pollution," where AI-generated false content becomes a training source for subsequent models, causing errors to accumulate generation by generation and distorting the model's cognitive abilities.
• Data contamination could trigger a series of real-world risks, including financial market manipulation, misleading public opinion, and generating incorrect medical advice. 🚨
• Official proposed countermeasures include: strengthening source supervision, reinforcing full lifecycle risk assessment, and performing end-point cleaning and remediation of contaminated data.

Related Links:

• Ministry of State Security: Beware of AI Data Poisoning, 0.01% False Text Can Increase Harmful Output by 11.2%

ChatGPT用户量飙升：周活跃用户有望达7亿

简报：

• OpenAI于2025年8月5日宣布，其AI助手ChatGPT的周活跃用户数有望在本周达到7亿。
• OpenAI首席产品官Kevin Weil表示，这一数字与去年同期相比，增长了超过4倍。
• 此前，ChatGPT在2025年3月已达到5亿周活跃用户的里程碑。
• 公司首席执行官Sam Altman透露，未来几个月将推出一系列新模型、产品和功能，备受期待的GPT-5模型也在计划中。

相关链接：

• OpenAI：ChatGPT 本周有望达 7 亿周活跃用户，比去年增长 4 倍多

ChatGPT User Base Soars: Weekly Active Users Expected to Reach 700 Million 🚀

Briefing:

• OpenAI announced on August 5, 2025, that its AI assistant ChatGPT's weekly active user count is expected to reach 700 million this week.
• OpenAI Chief Product Officer Kevin Weil stated that this figure represents a more than four-fold increase compared to the same period last year. 📈
• Previously, ChatGPT reached the milestone of 500 million weekly active users in March 2025.
• Company CEO Sam Altman revealed that a series of new models, products, and features will be launched in the coming months, with the highly anticipated GPT-5 model also in the pipeline. ✨

Related Links:

• OpenAI：ChatGPT 本周有望达 7 亿周活跃用户，比去年增长 4 倍多

阿里发布图像生成模型Qwen-Image，主打高保真中英文渲染

简报：

• 阿里巴巴通义千问团队于8月5日宣布，开源其首个图像生成基础模型Qwen-Image。
• 该模型是一个20B的MMDiT模型，在复杂文本渲染方面表现出色，能够高保真地输出中文和英文字符，尤其在中文渲染上大幅领先现有模型。
• Qwen-Image在多个公开基准测试中取得了最先进的性能，并支持风格迁移、增删改、文字编辑、人物姿态调整等多种图像编辑功能。

相关链接：

• 阿里通义千问首个图像生成基础模型Qwen-Image开源，支持中文高保真输出
• 腾讯混元 0.5B、1.8B、4B、7B 模型开源发布，消费级显卡即可运行
• 小米声音理解大模型 MiDashengLM-7B 发布并全量开源，22 个公开评测集刷新最好成绩

Alibaba Launches Qwen-Image, an Image Generation Model Specializing in High-Fidelity Chinese and English Rendering 🎨

Brief:

• Alibaba's Tongyi Qianwen team announced on August 5th that it has open-sourced its first image generation foundation model, Qwen-Image.
• The model is a 20B MMDiT model, excelling in complex text rendering and capable of high-fidelity output of Chinese and English characters, significantly surpassing existing models in Chinese rendering in particular. ✨
• Qwen-Image has achieved state-of-the-art performance in multiple public benchmarks and supports various image editing functionalities, including style transfer, adding/deleting/modifying elements, text editing, and human pose adjustment. 🚀

Related Links:

• Alibaba Tongyi Qianwen's First Open-Sourced Image Generation Foundation Model Qwen-Image Supports High-Fidelity Chinese Output
• Tencent Hunyuan 0.5B, 1.8B, 4B, 7B Models Open-Sourced, Runnable on Consumer-Grade Graphics Cards
• Xiaomi's MiDashengLM-7B Large Audio Understanding Model Released and Fully Open-Sourced, Setting New Records on 22 Public Benchmarks

腾讯开源四款混元小模型，消费级硬件即可部署

简报：

• 腾讯混元于8月4日发布了四款开源小尺寸模型，参数量分别为0.5B、1.8B、4B和7B，消费级显卡即可运行。
• 模型亮点在于强大的agent能力和高达256k的原生长上下文窗口，可一次性处理约40万汉字。
• 这四款模型均为融合推理模型，可根据场景灵活选择快、慢两种思考模式，适用于手机、智能座舱等低功耗场景。
• 目前模型已在腾讯会议、微信读书、搜狗输入法等多个腾讯业务中应用。

相关链接：

• 腾讯混元 0.5B、1.8B、4B、7B 模型开源发布，消费级显卡即可运行

Tencent Open-Sources Four Hunyuan Mini-Models, Deployable on Consumer-Grade Hardware

Briefing:

• Tencent Hunyuan released four open-source small-sized models on August 4th, with parameter sizes of 0.5B, 1.8B, 4B, and 7B respectively, runnable on consumer-grade graphics cards. 🚀
• Key highlights include powerful agent capabilities and an impressive native long context window of up to 256k, capable of processing approximately 400,000 Chinese characters at once.
• These four models are all fusion inference models, able to flexibly select between fast and slow thinking modes depending on the scenario, making them suitable for low-power applications like mobile phones and smart cockpits. 📱
• Currently, the models have been applied in various Tencent businesses such as Tencent Meeting, WeChat Reading, and Sogou Input Method. ✨

Related Links:

• Tencent Hunyuan 0.5B, 1.8B, 4B, 7B Models Open-Sourced, Deployable on Consumer-Grade Graphics Cards

人形机器人订单涌现，表演导览成初期应用场景

简报：

• 近期，智元机器人、宇树科技、优必选等头部企业相继获得大额订单，如智元与宇树中标1.24亿元项目，优必选中标9051.15万元项目。
• 当前订单的主要应用场景集中于表演讲解、展馆导览等交互服务领域。
• 业内观点认为，从高关注度、低技术门槛的表演导览场景切入，是人形机器人后续拓展至工业制造及家庭服务场景的务实选择和必要起点。

相关链接：

• 人形机器人“订单潮”涌动，表演、导览仍是“主战场”
• 人形机器人“订单潮”涌动

Humanoid Robot Orders Surge, Performance and Guided Tours Emerge as Initial Application Scenarios

Briefing: 📜

• Recently, leading companies such as ZHIYUAN Robotics, Unitree Robotics, and UBTECH have successively secured large orders. For instance, ZHIYUAN and Unitree jointly won a 124 million yuan project, while UBTECH secured a 90.5115 million yuan project.
• Current orders primarily focus on interactive service sectors such as performance explanations and museum guided tours.
• Industry insiders believe that entering the market through high-visibility, low-technical-threshold scenarios like performance and guided tours is a pragmatic choice and a necessary starting point for humanoid robots to subsequently expand into industrial manufacturing and home service applications. 🤖🚀

Related Links:

• Humanoid Robot "Order Wave" Surges, Performance and Guided Tours Remain the "Main Battlefield"
• Humanoid Robot "Order Wave" Surges

大疆Osmo 360在台上市，公布售价与限时优惠

简报:

• DJI推出全新8K全景相机Osmo 360，采用为全景相机定制的方形1英吋CMOS感光元件，支持原生8K/50fps影片与1.2亿像素照片。
• 该相机已在台湾正式发售，标准套装售价为新台币13,980元，畅拍套装为新台币17,800元。
• Osmo 360机身重183克，内置105GB储存空间，支持10米裸机防水，并可通过OsmoAudio生态系统直连两组DJI麦克风发射器。
• 上市同步推出限时优惠：至2025年8月31日前，购买相机可以优惠价加购DJI Mic Mini系列产品；携带任一品牌全景相机至指定体验店并分享社群，可免费获得DJI Care Refresh一年版。

相关链接:

• 全球首創方形CMOS　DJI Osmo 360 8K全景相機登台開賣

DJI Osmo 360 Launched in Taiwan, Pricing and Limited-Time Offers Announced

Brief:

• DJI has launched the new 8K panoramic camera, Osmo 360, featuring a custom-designed square 1-inch CMOS sensor built for panoramic cameras. It supports native 8K/50fps video and 120-megapixel photos. 🚀
• The camera has officially gone on sale in Taiwan. The Standard Combo is priced at NT$13,980, and the Fly More Combo is NT$17,800.
• The Osmo 360 weighs 183 grams, features 105GB of built-in storage, supports bare-camera waterproofing up to 10 meters 🌊, and can directly connect two sets of DJI microphone transmitters via the OsmoAudio ecosystem.
• Limited-time offers are being launched concurrently with the release: Until August 31, 2025, purchasers can add DJI Mic Mini series products at a discounted price. Additionally, by bringing any brand of panoramic camera to a designated experience store and sharing on social media, customers can receive a free one-year DJI Care Refresh plan. 🎁

Related Link:

• World's First Square CMOS Sensor: DJI Osmo 360 8K Panoramic Camera Launched in Taiwan

联想发布新一代工作站，聚焦AI与混合办公场景

简报：

• 联想发布全新工作站解决方案与设备，旨在支持混合办公、高效运算及AI开发等复杂任务。
• 新产品包括：为远程团队设计的Lenovo Access解决方案、新一代ThinkStation P2与P3系列桌面工作站（涵盖Tower、Ultra SFF、Tiny等多种形态）。
• 同时推出搭载AMD Ryzen AI PRO 300系列处理器的ThinkPad P14s移动工作站，作为Copilot+ PC，整合高达82 TOPS的AI算力。
• 全新工作站最高可搭载Intel Core Ultra 9处理器或AMD Ryzen AI 9 HX PRO 370处理器，并支持NVIDIA RTX PRO系列GPU。

相关链接：

• Lenovo推全新工作站解決方案、裝置陣容　助攻混合辦公、高效運算

Lenovo Unveils Next-Gen Workstations, Focusing on AI and Hybrid Work Scenarios

Briefing:

• Lenovo has launched new workstation solutions and devices designed to support complex tasks such as hybrid work, high-performance computing, and AI development. 💻
• New products include: Lenovo Access solutions tailored for remote teams, and the next-generation ThinkStation P2 and P3 series desktop workstations (covering various form factors like Tower, Ultra SFF, and Tiny).
• Simultaneously unveiled is the ThinkPad P14s mobile workstation, powered by AMD Ryzen AI PRO 300 series processors. As a Copilot+ PC, it integrates up to 82 TOPS of AI computing power. ✨
• The new workstations can be configured with up to Intel Core Ultra 9 processors or AMD Ryzen AI 9 HX PRO 370 processors, and support NVIDIA RTX PRO series GPUs. 🚀

Related Links:

• Lenovo launches new workstation solutions and device lineup to boost hybrid office and high-performance computing