单词学习：![[54928bc1-b183-4b76-8c30-4b320308b9a5.mp3]]

新闻播报：![[868eb2f6-de96-4a00-a47c-d390aae04ade.mp3]]

Adobe携手OpenAI推ChatGPT自然语言修图功能

简报：

• Adobe宣布与OpenAI建立合作关系，将Photoshop和Adobe Express接入ChatGPT生态系统
• 用户可在ChatGPT左下角"更多"菜单中找到Adobe应用，通过自然语言对话直接下达修图指令
• Adobe Express结合GPT后，即便输入模糊文字提示也能自动生成完整传单，并支持鼠标点击快速调整参数
• Adobe副总裁强调合作旨在整合技术优势而非取代自身AI模型，其Firefly模型训练数据均来自授权内容确保版权合规
• Adobe近期还为Photoshop推出AI助手功能，并发布Firefly Image5模型支持生成400万像素高清图像

相关链接：

• Adobe与OpenAI联手：让修图变得简单有趣

Adobe Partners with OpenAI to Launch ChatGPT Natural Language Photo Editing Feature 🎨✨

Brief:

• Adobe announces partnership with OpenAI, integrating Photoshop and Adobe Express into the ChatGPT ecosystem
• Users can find Adobe apps in the "More" menu at the bottom left of ChatGPT, issuing photo editing commands directly through natural language conversations
• Adobe Express combined with GPT can automatically generate complete flyers even from vague text prompts, with support for quick parameter adjustments via mouse clicks
• Adobe's Vice President emphasizes the collaboration aims to integrate technological advantages rather than replace their own AI models; their Firefly model training data all comes from licensed content to ensure copyright compliance 🤝
• Adobe recently launched an AI assistant feature for Photoshop and released the Firefly Image 5 model, supporting generation of high-definition images up to 4 million pixels

Related Link:

• Adobe and OpenAI Join Forces: Making Photo Editing Simple and Fun

智源Emu3.5模型发布，首创"下一状态预测"实现AI具身操作突破

简报：

• 北京智源人工智能研究院10月30日发布新一代多模态世界模型Emu3.5，首次将自回归式"下一状态预测"（NSP）引入多模态序列建模
• 模型将文本、图像、动作指令等多模态输入视为连续状态序列，不仅能理解当前场景还能预判操作后结果并规划最优动作路径
• Emu3.5支持文图协同生成、智能图像编辑和时空动态推理等功能，可执行"把咖啡杯移到桌子右边并调亮色调"等复合操作
• 模型实现真正意义上的跨模态自由切换与协同推理，打破传统多模态模型仅做特征对齐的局限
• 智源表示Emu3.5将率先应用于教育、医疗、娱乐等领域，并持续开源部分能力推动多模态生态发展

相关链接：

• 智源发布Emu3.5大模型：以"下一状态预测"重构多模态智能，具身操作能力惊艳业界

Beijing Academy of Artificial Intelligence Releases Emu3.5 Model, Pioneering "Next-State Prediction" for Breakthrough in AI Embodied Operations 🤖

Brief:

• Beijing Academy of Artificial Intelligence (BAAI) released the next-generation multimodal world model Emu3.5 on October 30, introducing autoregressive "Next-State Prediction" (NSP) into multimodal sequence modeling for the first time
• The model treats multimodal inputs such as text, images, and action commands as continuous state sequences, enabling it not only to understand current scenes but also to predict post-operation results and plan optimal action paths
• Emu3.5 supports text-image co-generation, intelligent image editing, and spatiotemporal dynamic reasoning, capable of executing complex operations like "move the coffee cup to the right side of the table and brighten the color tone" 🎯
• The model achieves true cross-modal free switching and collaborative reasoning, breaking through the limitations of traditional multimodal models that only perform feature alignment
• BAAI states that Emu3.5 will be prioritized for applications in education, healthcare, entertainment, and other fields, while continuously open-sourcing partial capabilities to promote multimodal ecosystem development ✨

Related Link:

• BAAI Releases Emu3.5 Large Model: Reconstructing Multimodal Intelligence with "Next-State Prediction," Embodied Operation Capabilities Impress the Industry

OpenAI 发布开源 Safeguard 模型：可展示完整推理过程，灵活自定义内容审核策略

简报：

• OpenAI 于 10 月 29 日推出 gpt-oss-safeguard-120b 和 gpt-oss-safeguard-20b 两款开源模型，专门用于根据用户提供的策略对内容进行推理、分类和标记
• 模型采用 Apache 2.0 许可证，允许开发者免费使用、修改和商业部署，支持完整的"思维链"（Chain-of-Thought）输出，可展示每一步推理过程
• 与传统分类器不同，Safeguard 将定义安全边界的权力交给开发者，能在推理阶段直接解释并应用自定义安全策略，无需重新训练即可快速适应新规则
• 模型特别适用于新兴风险、领域细微场景、缺乏训练样本以及重视可解释性的应用场景
• OpenAI 提示该模型存在两个权衡：在有充足数据时传统分类器精度可能更高；运行速度较慢且资源密集，不适合大规模实时扫描

相关链接：

• Introducing gpt-oss-safeguard
• 技术报告
• Hugging Face 下载地址

OpenAI Launches Open-Source Safeguard Models: Complete Reasoning Process Display with Flexible Custom Content Moderation Policies 🛡️

Brief:

• On October 29, OpenAI introduced two open-source models, gpt-oss-safeguard-120b and gpt-oss-safeguard-20b, specifically designed to reason about, classify, and flag content based on user-provided policies
• The models are released under the Apache 2.0 license, allowing developers to freely use, modify, and commercially deploy them. They support complete "Chain-of-Thought" output, displaying each step of the reasoning process
• Unlike traditional classifiers, Safeguard empowers developers to define safety boundaries. It can directly interpret and apply custom safety policies during inference without requiring retraining, enabling rapid adaptation to new rules 🚀
• The models are particularly well-suited for emerging risks, domain-specific nuanced scenarios, situations with limited training samples, and applications where interpretability is valued
• OpenAI notes two trade-offs with the model: traditional classifiers may achieve higher precision when sufficient data is available; the model runs slower and is resource-intensive, making it unsuitable for large-scale real-time scanning ⚖️

Related Links:

• Introducing gpt-oss-safeguard
• Technical Report
• Hugging Face Download

OpenAI 三季度巨亏115亿美元，微软持股27%承担31亿损失

简报：

• 根据微软向美国证券交易委员会提交的财报，OpenAI在截至9月30日的季度中净亏损高达115亿美元
• 微软持有OpenAI 27%股份，因该亏损导致其净收入减少31亿美元，但微软整体季度净收入仍达277亿美元，财务状况稳健
• OpenAI已正式完成向营利性公司的转型，这一亏损数字相当惊人，考虑到其2025年上半年收入仅为43亿美元
• 微软与OpenAI达成新协议，重组后的OpenAI将成为公益公司，微软持股价值约1350亿美元，为未来上市铺路
• OpenAI将采购2500亿美元的Azure云端运算服务，微软则延长智慧财产权协议至2032年

相关链接：

• 微软财报透露：OpenAI 在上季度亏损高达 115 亿美元
• OpenAI與微軟達成協議重組為公益公司　為上市鋪路

OpenAI Reports Massive Q3 Loss of dJyH6kDkfC5RlM16.answerText1.5 Billion; Microsoft's 27% Stake Results in $3.1 Billion Loss 💸📉

Brief:

• According to Microsoft's financial report filed with the U.S. Securities and Exchange Commission, OpenAI recorded a staggering net loss of dJyH6kDkfC5RlM16.answerText1.5 billion in the quarter ending September 30
• Microsoft, holding a 27% stake in OpenAI, saw its net income reduced by $3.1 billion due to this loss; however, the company's overall quarterly net income still reached$27.7 billion, maintaining a solid financial position
• OpenAI has officially completed its transformation into a for-profit company. This loss figure is particularly striking considering the company generated only $4.3 billion in revenue during the first half of 2025 🚀
• Microsoft and OpenAI have reached a new agreement to restructure OpenAI as a public benefit corporation, with Microsoft's stake valued at approximately dJyH6kDkfC5RlM16.answerText35 billion, paving the way for a future IPO
• OpenAI will procure $250 billion worth of Azure cloud computing services, while Microsoft extends its intellectual property rights agreement through 2032

Related Links:

• Microsoft's Financial Report Reveals: OpenAI Lost as Much as dJyH6kDkfC5RlM16.answerText1.5 Billion Last Quarter
• OpenAI and Microsoft Reach Agreement to Restructure as Public Benefit Corporation, Paving the Way for IPO

英伟达市值逼近5万亿美元，黄仁勋身价暴增成"五兆男"

简报：

• 英伟达市值在GTC大会后一度飙升5%，达到近4.95万亿美元，创历史新高，超越英国股市总市值
• 英伟达宣布新一代Grace Blackwell系统累计订单已达5000亿美元，预计2026年底前将出货2000万颗GPU
• 创始人黄仁勋身价攀升至1740亿美元（约5.33万亿新台币），跃居全球富豪第九名，今年累计暴增约1.84万亿新台币
• 英伟达的GPU在过去三年里成为推动AI技术快速增长的关键，分析师认为"AI的开始和结束都与英伟达息息相关"
• 英伟达台湾总部选址北士科T17、T18地块陷入膠着，黄仁勋首度回应称"不知道问题卡在哪里"

相关链接：

• 英伟达市值突破5万亿美元，推动AI行业蓬勃发展
• 富可敵國！輝達股價飆漲創新高　AI教父黃仁勳榮升「五兆男」
• 輝達台灣總部北士科T17、T18一度卡關　黃仁勳首度回應了

NVIDIA Market Cap Approaches $5 Trillion, Jensen Huang's Wealth Surges to Become "Five Trillion Man" 💰🚀

Brief:

• NVIDIA's market capitalization soared 5% following the GTC conference, reaching nearly $4.95 trillion—a historic high that surpasses the total market value of the UK stock market 📈
• NVIDIA announced cumulative orders for its new-generation Grace Blackwell system have reached $500 billion, with an expected shipment of 20 million GPUs by the end of 2026
• Founder Jensen Huang's net worth climbed to dJyH6kDkfC5RlM16.answerText74 billion (approximately NT$5.33 trillion), propelling him to ninth place on the global billionaire list with a year-to-date increase of about NTdJyH6kDkfC5RlM16.answerText.84 trillion
• NVIDIA's GPUs have been instrumental in driving rapid AI technology growth over the past three years, with analysts noting that "the beginning and end of AI are closely tied to NVIDIA"
• NVIDIA's Taiwan headquarters site selection at Beishi Technology Park plots T17 and T18 has hit a snag; Huang responded for the first time, stating "I don't know where the problem lies"

Related Links:

• NVIDIA Market Cap Breaks Through $5 Trillion, Propelling AI Industry Boom
• Wealthier Than Nations! NVIDIA Stock Soars to Record High, AI Godfather Jensen Huang Ascends to "Five Trillion Man" Status
• NVIDIA Taiwan Headquarters at Beishi Tech Park T17, T18 Once Stalled—Jensen Huang Responds for First Time

黄仁勋身价飙升至5.33万亿新台币，英伟达市值逼近5万亿美元

简报：

• 英伟达于华盛顿特区GTC大会宣布重磅消息后，市值一度飙升5%至近5万亿美元，创历史新高，超越英国股市4.3万亿美元市值
• 黄仁勋身价攀升至1740亿美元（约5.33万亿新台币），位居全球富豪第九名，今年累计暴增约1.84万亿新台币
• 黄仁勋宣布新一代Grace Blackwell系统累计订单已达5000亿美元，预计2026年底前出货2000万颗GPU，增速是前一代Hopper的5倍
• 英伟达股价推升至203.15美元创新高，今年涨幅超45%，远高于费城半导体指数27%涨幅
• 黄仁勋今年已卖股40次，套现约5.797亿美元（约185.5亿新台币）

相关链接：

• 富可敵國！輝達股價飆漲創新高　AI教父黃仁勳榮升「五兆男」
• 英伟达市值突破5万亿美元，推动 AI 行业蓬勃发展

Jensen Huang's Net Worth Soars to NT$5.33 Trillion as NVIDIA Market Cap Approaches$5 Trillion 💰🚀

Brief:

• Following major announcements at the GTC conference in Washington D.C., NVIDIA's market capitalization surged 5% to nearly $5 trillion, reaching an all-time high and surpassing the UK stock market's$4.3 trillion valuation 📈
• Jensen Huang's net worth climbed to dJyH6kDkfC5RlM16.answerText74 billion (approximately NT$5.33 trillion), ranking him ninth among global billionaires, with a cumulative increase of about NTdJyH6kDkfC5RlM16.answerText.84 trillion this year
• Huang announced that the new generation Grace Blackwell system has accumulated orders totaling $500 billion, with an expected shipment of 20 million GPUs by the end of 2026—five times faster than the previous Hopper generation
• NVIDIA's stock price reached a new high of $203.15, up over 45% this year, significantly outpacing the Philadelphia Semiconductor Index's 27% gain
• Huang has sold shares 40 times this year, cashing out approximately $579.7 million (about NTdJyH6kDkfC5RlM16.answerText8.55 billion)

Related Links:

• 富可敵國！輝達股價飆漲創新高　AI教父黃仁勳榮升「五兆男」
• 英伟达市值突破5万亿美元，推动 AI 行业蓬勃发展

微软发布Agent Lightning开源框架，通过强化学习优化多代理系统

简报：

• 微软于2025年10月30日发布开源框架Agent Lightning，旨在通过强化学习优化多代理系统
• 该框架可在不改变现有代理架构的情况下，将真实代理行为转化为RL过渡，提升策略大规模语言模型性能
• Agent Lightning将代理建模为部分可观测的马尔可夫决策过程，采用"训练代理解耦"方法，由Lightning Server进行训练和服务
• 实验评估涵盖文本转SQL、检索增强生成和数学问答三项任务，均显示出稳定的奖励提升
• 框架支持OpenTelemetry和轻量级嵌入式追踪器两种数据收集路径，并提供与OpenAI兼容的API接口

相关链接：

• 微软推出Agent Lightning:全新AI框架助力强化学习训练大规模语言模型
• 论文地址

Microsoft Releases Agent Lightning Open-Source Framework, Optimizing Multi-Agent Systems Through Reinforcement Learning 🚀

Brief:

• Microsoft released the open-source framework Agent Lightning on October 30, 2025, designed to optimize multi-agent systems through reinforcement learning 🤖
• The framework can transform real agent behavior into RL transitions without modifying existing agent architectures, enhancing policy large language model performance
• Agent Lightning models agents as Partially Observable Markov Decision Processes (POMDPs), adopting a "training-agent decoupling" approach with training and serving handled by the Lightning Server
• Experimental evaluations covering three tasks—text-to-SQL, retrieval-augmented generation, and mathematical question answering—all demonstrated consistent reward improvements ✨
• The framework supports two data collection paths: OpenTelemetry and a lightweight embedded tracer, and provides an OpenAI-compatible API interface

Related Links:

• Microsoft Launches Agent Lightning: New AI Framework Empowers Reinforcement Learning Training for Large Language Models
• Paper

Meta 申请驳回 3.5 亿美元色情内容侵权诉讼

简报：

• Strike3Holdings 起诉 Meta 通过 2500 个隐藏 IP 地址组成的秘密网络种子下载 2400 部成人影片，索赔超 3.5 亿美元
• Meta 向美国地区法院申请驳回诉讼，称下载行为发生在 2018 年，远早于其 AI 项目启动时间（约四年后）
• Meta 辩称下载行为属个人使用且与 AI 训练无关，每年仅 22 次下载无法可靠关联到公司员工，可能由承包商或访客所为
• Meta 强调其服务条款明确禁止生成成人内容，与使用成人素材进行 AI 训练的指控相矛盾
• Meta 表示 Strike3 未提供证据证明公司指示下载行为或知晓此类活动，指控完全虚假

相关链接：

• Meta Accused of Torrenting Porn to Advance Its Goal of AI 'Superintelligence'

Meta Seeks Dismissal of $350 Million Pornographic Content Infringement Lawsuit

Brief:

• Strike3Holdings sued Meta for allegedly downloading 2,400 adult films via a secret network of 2,500 hidden IP addresses using torrent technology, seeking over $350 million in damages 📋
• Meta filed a motion with the U.S. District Court to dismiss the lawsuit, arguing the downloads occurred in 2018—approximately four years before the company's AI projects were launched ⚖️
• Meta contends the downloads were for personal use and unrelated to AI training, stating that only 22 downloads per year cannot be reliably linked to company employees and may have been conducted by contractors or visitors
• Meta emphasized that its terms of service explicitly prohibit the generation of adult content, which contradicts allegations that it used adult material for AI training 🤖
• Meta states that Strike3 has provided no evidence proving the company directed the downloads or had knowledge of such activities, calling the allegations completely false

Related Links:

• Meta Accused of Torrenting Porn to Advance Its Goal of AI 'Superintelligence'