单词学习：![[871d09fc-6b04-409a-bef7-cd3797a0bd84.mp3]]

新闻播报：![[de844d1a-363b-46f3-8d17-6aed57224917.mp3]]

华为开源Flex

，AI算力切分技术提升硬件利用率

简报：

• 在AI容器应用落地与发展论坛上，华为发布并宣布将开源其Flex
  AI容器软件。
• 该技术通过算力切分，可将单张GPU或NPU算力卡以低至10%的粒度进行切分，实现单卡同时承载多个AI工作负载。
• Flex
  还能聚合集群内各节点的空闲算力形成“共享算力池”，并将同步在魔擎社区中开源。

相关链接：

• 华为发布并开源 Flex
  AI 容器技术，实现单卡同时承载多个 AI 工作负载

Huawei Open-sources Flex:ai: AI Computing Power Slicing Technology Boosts Hardware Utilization 🎉

Brief:

• At the AI Container Application Deployment and Development Forum, Huawei announced and launched its Flex
  AI container software, declaring it open-source.
• This technology, through computing power slicing, can divide a single GPU or NPU computing card into granular units as low as 10%, enabling a single card to simultaneously support multiple AI workloads. ✨
• Flex
  can also aggregate idle computing power from various nodes within a cluster to form a "shared computing power pool," and will be open-sourced concurrently in the MoQing community. 🤝

Related links:

• Huawei Launches and Open-sources Flex
  AI Container Technology, Enabling Single Card to Simultaneously Support Multiple AI Workloads

腾讯开源视频生成模型HunyuanVideo1.5，消费级显卡即可部署

简报：

• 腾讯混元大模型团队发布了最新的视频生成模型HunyuanVideo1.5，该模型基于Diffusion Transformer（DiT）架构，参数量为8.3B。
• 模型支持通过文字或“图片+文字”生成5至10秒的高清视频，兼容中英文输入和多种风格。
• HunyuanVideo1.5的部署门槛显著降低，仅需14G显存的消费级显卡即可运行。
• 该模型目前已在腾讯“元宝”平台上线开放体验，并已在Hugging Face和GitHub上开源。

相关链接：

• 腾讯发布全新视频生成模型 HunyuanVideo1.5，降低视频创作门槛

Tencent Open-Sources Video Generation Model HunyuanVideo1.5, Deployable on Consumer-Grade Graphics Cards

Brief:

• Tencent's Hunyuan large model team has released its latest video generation model, HunyuanVideo1.5. The model is based on the Diffusion Transformer (DiT) architecture and boasts 8.3 billion parameters.
• The model supports generating 5 to 10-second high-definition videos via text or "image + text" prompts. It is compatible with both Chinese and English inputs and various styles. 🎬
• HunyuanVideo1.5 significantly lowers deployment barriers, requiring only a consumer-grade graphics card with 14GB of VRAM to run. ✨
• The model is currently available for open experience on Tencent's 'Yuanbao' platform and has been open-sourced on Hugging Face and GitHub. 🚀

Related Links:

• Tencent Releases New Video Generation Model HunyuanVideo1.5, Lowering the Barrier for Video Creation

OpenAI发布GPT-5，旨在加速科学与数学研究

简报：

• OpenAI近日发布了最新的GPT-5模型，该模型具备比先前版本更强的计算能力和语言理解能力。
• GPT-5能够更准确地理解和生成数学公式与科学概念，旨在帮助科研人员更高效地解决复杂问题，提升研究效率。
• 该模型在药物发现领域，可通过分析大量生物医学数据来识别潜在的药物分子；在新材料研究方面，可用于模拟和预测材料特性，加速研发进程。

相关链接：

• OpenAI 发布新 GPT-5 模型，加速数学与科学研究
• OpenAI 算力需求激增：2028 年支出将达到 1100 亿美元
• ChatGPT 群聊正式上线：最多 20 人同屏，@AI 即可拉它入脑暴

OpenAI Releases GPT-5, Aiming to Accelerate Science and Math Research ✨

Briefing:

• OpenAI recently unveiled its latest GPT-5 model, which boasts enhanced computational power and language comprehension capabilities compared to its predecessors.
• GPT-5 can more accurately understand and generate mathematical formulas and scientific concepts, aiming to help researchers solve complex problems more efficiently and boost research productivity. 🚀
• In drug discovery, the model can identify potential drug molecules by analyzing vast amounts of biomedical data; in new material research, it can be used to simulate and predict material properties, thereby accelerating the R&D process. 🧪

Related Links:

• OpenAI Releases New GPT-5 Model to Accelerate Math and Science Research
• OpenAI's Computing Power Demand Soars: Expenditures to Reach dJyH6kDkfC5RlM16.answerText10 Billion by 2028
• ChatGPT Group Chat Officially Launched: Up to 20 People On-Screen, @AI to Bring it into Brainstorms

OpenAI为ChatGPT引入群聊，支持20人协作

简报： OpenAI宣布向所有已登录用户（包括免费版）正式上线ChatGPT群聊功能，该功能最高支持20人参与，用户可通过@符号与AI进行互动，并支持网页与移动端实时同步。

相关链接：

• ChatGPT 群聊正式上线：最多 20 人同屏，@AI 即可拉它入脑暴
• OpenAI 算力需求激增：2028 年支出将达到 1100 亿美元
• OpenAI 发布新 GPT-5 模型，加速数学与科学研究

OpenAI Introduces Group Chat for ChatGPT, Supporting 20-Person Collaboration 🚀

Brief: OpenAI has officially launched the ChatGPT group chat feature for all logged-in users, including those on the free version. This new functionality supports up to 20 participants, allowing users to interact with the AI via the @ symbol. The feature is available and synchronizes in real-time across both web and mobile platforms. Get ready to brainstorm! 🗣️🤝

Related Links:

• ChatGPT Group Chat Officially Launched: Up to 20 People on Screen, @AI to Bring It into Brainstorming Sessions
• OpenAI's Computing Power Demand Soars: Expenses to Reach dJyH6kDkfC5RlM16.answerText10 Billion by 2028
• OpenAI Releases New GPT-5 Model, Accelerating Math and Science Research

谷歌发布基于Gemini 3的新一代图像模型Nano Banana Pro

简报：

• 谷歌发布新一代图像生成模型Nano Banana Pro，该模型基于最新发布的Gemini 3。
• 新模型支持最高4K分辨率图像生成、联网实时搜索生成信息图，并集成于NotebookLM、Slides和Vids等应用中。
• 生成的图片将同时写入SynthID隐形水印和C2PA内容凭证，用户可通过Gemini App验证图片是否由谷歌AI生成。
• 新模型定价上调，1080p/2K分辨率为0.139美元/张，4K分辨率为0.24美元/张，较前代上涨3-6倍。

相关链接：

• 谷歌推出Nano Banana Pro：基于Gemini 3，支持4K、联网搜索与C2PA验证
• 谷歌NotebookLM重磅更新：AI一键生成专业PPT，配图惊艳碾压Gemini App！
• Google Gemini 推出新功能：轻松识别 AI 生成的图片内容
• Google推Gemini 3！功能摘要一次看 大學生免費用Google AI Pro一整年

Google Releases Next-Generation Image Model Nano Banana Pro Based on Gemini 3

Brief:

• Google has unveiled its next-generation image generation model, Nano Banana Pro, built upon the newly released Gemini 3. 🖼️
• The new model supports up to 4K resolution image generation, real-time web search for infographic creation, and is integrated into applications like NotebookLM, Slides, and Vids.
• Generated images will embed both SynthID invisible watermarks and C2PA content credentials. Users can verify if an image was generated by Google AI using the Gemini App. ✅
• The new model comes with increased pricing: $0.139 per image for 1080p/2K resolution and$0.24 per image for 4K resolution, marking a 3-6x increase compared to its predecessor. 💰

Related links:

• Google Unveils Nano Banana Pro: Based on Gemini 3, Supports 4K, Web Search & C2PA Verification
• Google NotebookLM's Major Update: AI Generates Professional PPTs with Stunning Images, Outperforming Gemini App!
• Google Gemini Introduces New Feature: Easily Identify AI-Generated Image Content
• Google Launches Gemini 3! All Features at a Glance, University Students Get Google AI Pro for Free for a Year

微软升级高级粘贴功能，引入本地AI处理

简报：

• 微软宣布其“高级粘贴”功能已升级，通过本地AI模型执行任务，不再依赖云端。
• 该功能旨在通过智能识别和处理内容，简化用户在不同应用程序间的操作，提高工作效率。
• 由于在本地设备进行处理，新功能在保护用户隐私的同时，也因减少了网络延迟而使粘贴操作更加流畅。

相关链接：

• 微软推出 AI 驱动的复制粘贴功能，实现本地 AI 处理

Microsoft Upgrades Advanced Paste Feature, Introduces Local AI Processing

Brief:

• Microsoft has announced an upgrade to its "Advanced Paste" feature, which now performs tasks using local AI models, eliminating the need for cloud reliance. 🚀
• This functionality aims to streamline user operations across various applications and boost productivity by intelligently recognizing and processing content. 💡
• By leveraging on-device processing, the new feature not only safeguards user privacy but also ensures smoother paste operations due to reduced network latency. 🔒

Related Link:

• Microsoft Launches AI-Powered Copy-Paste Feature with Local AI Processing

Meta发布DreamGym框架，以模拟环境提升AI代理训练效能

简报：

• Meta与芝加哥大学及加州大学伯克利分校的研究人员共同开发了DreamGym框架，旨在解决强化学习（RL）训练大型语言模型（LLM）代理时面临的高成本、复杂性和反馈不可靠等问题。
• 该框架通过在模拟环境中训练AI代理，利用合成交互降低数据收集和环境交互的成本，从而提升训练效率与安全性。
• DreamGym框架由“基于推理的经验模型”、“经验重放缓冲区”和“课程任务生成器”三部分组成，能根据代理表现动态调整任务难度。
• 在WebArena等基准测试中，经DreamGym训练的代理成功率超过基线方法30%以上，证明了其在多个领域的优越性。

相关链接：

• Meta 推出 DreamGym 框架，让 AI 代理训练更高效、更安全

Meta Releases DreamGym Framework to Enhance AI Agent Training Efficiency Through Simulated Environments

Brief:

• Meta, in collaboration with researchers from the University of Chicago and UC Berkeley, has developed the DreamGym framework. It aims to address the high costs, complexity, and unreliable feedback encountered when training Large Language Model (LLM) agents using Reinforcement Learning (RL). 🧠
• The framework improves training efficiency and safety by training AI agents in simulated environments, utilizing synthetic interactions to reduce data collection and environment interaction costs.
• The DreamGym framework consists of three components: an "inference-based experience model," an "experience replay buffer," and a "curriculum task generator," which dynamically adjusts task difficulty based on the agent's performance. 🛠️
• In benchmarks such as WebArena, agents trained with DreamGym achieved a success rate over 30% higher than baseline methods, demonstrating its superiority across multiple domains. ✨

Related Links:

• Meta launches DreamGym framework to make AI agent training more efficient and safer