单词学习：![[9a5c7031-e7ba-4450-b311-a0184a3a4a95.mp3]]

新闻播报：![[3c359c41-7b49-4424-8a79-cb0ee4630677.mp3]]

百度开源高效MoE模型，登顶Hugging Face榜单

简报：

• 百度近日开源了文心大模型新成员ERNIE-4.5-21B-A3B-Thinking，该模型已登顶Hugging Face文本生成模型榜单首位，并位居整体模型榜第三。
• 模型采用轻量级Mixture-of-Experts（MoE）架构，总参数为21B，但每个token仅激活3B参数，显著降低了计算开销，并支持128K长上下文窗口。
• ERNIE-4.5基于百度自研的飞桨（PaddlePaddle）深度学习框架进行训练，而非主流的PyTorch框架，全球范围内仅百度和谷歌采用自研框架训练大模型。
• 测试显示，该模型在逻辑推理、数学等任务上的部分指标接近或超越Gemini2.5Pro和GPT-5级别模型。
• 该模型已在Hugging Face平台发布，采用Apache2.0许可，支持商业用途。

相关链接：

• 百度文心新模型ERNIE-4.5-21B-A3B-Thinking强势登顶Hugging Face榜首

Baidu Open-Sources Efficient MoE Model, Tops Hugging Face Rankings

Briefing:

• Baidu recently open-sourced ERNIE-4.5-21B-A3B-Thinking, a new member of its Wenxin large model family. This model has already claimed the #1 spot on Hugging Face's text generation model leaderboard and ranks third overall on the platform. 🚀
• The model utilizes a lightweight Mixture-of-Experts (MoE) architecture, with a total of 21B parameters. However, only 3B parameters are activated per token, significantly reducing computational overhead and supporting a 128K long context window.
• ERNIE-4.5 was trained using Baidu's self-developed deep learning framework, PaddlePaddle, rather than the mainstream PyTorch framework. Globally, only Baidu and Google use their proprietary frameworks to train large models.
• Tests indicate that the model's performance in tasks like logical reasoning and mathematics approaches or even surpasses that of Gemini 2.5 Pro and GPT-5 level models in some metrics. 🧠✨
• The model has been released on the Hugging Face platform under the Apache 2.0 license, allowing for commercial use.

Related Links:

• Baidu Wenxin's new model ERNIE-4.5-21B-A3B-Thinking strongly tops Hugging Face leaderboard

苹果AI再受挫：前Siri主管离职，搜索项目添变数

简报：

• 苹果公司AI与搜索业务资深主管、前Siri负责人罗比·沃克将于下月离职，其此前正领导一个旨在与ChatGPT竞争的AI搜索系统项目，此次人事变动延续了苹果AI部门的人才流失潮，并加剧了外界对其AI转型的担忧。

相关链接：

• 苹果AI转型挑战加剧，Siri负责人离职，AI搜索前景不明

Apple AI Faces New Setback: Former Siri Head Departs, Search Project Hits Snag 🍎

Brief:

• Robby Walker, a veteran executive in Apple's AI and search division and former head of Siri, is set to depart next month. He was reportedly leading an AI search system project aimed at competing with ChatGPT. This personnel change continues the talent exodus from Apple's AI division and intensifies external concerns about its AI transformation. 📉

Related Link:

• 苹果AI转型挑战加剧，Siri负责人离职，AI搜索前景不明 🤔

Penske Media起诉谷歌，指控AI摘要功能侵犯版权并滥用垄断地位

简报：

• 美国知名媒体集团Penske Media Corporation（旗下拥有《滚石》、《综艺》等）已正式起诉谷歌，指控其“AI概览”（AI Overviews）功能未经授权使用其新闻内容生成摘要，构成版权侵犯。
• Penske称，该功能直接在搜索结果中提供信息，导致其网站流量、广告及联盟收入自2025年以来显著下降，其中联盟广告收入同比减少超过三分之一。
• 诉讼还指控谷歌利用其在搜索市场的垄断地位，强迫出版商接受其内容被用于AI训练，否则将面临被谷歌搜索排除在外的风险。
• 谷歌回应称，AI功能为网站带来了新的流量，并将为这些“毫无根据的指控”进行辩护。

相关链接：

• 谷歌AI摘要功能惹大祸！知名媒体集团起诉垄断巨头"偷内容毁生意"
• 彭斯克集团起诉 Google，指控 AI 摘要侵权并威胁数字媒体未来
• 《滚石》杂志母公司就 AI Overviews 起诉谷歌，称其进退两难

Penske Media Sues Google, Alleging Copyright Infringement and Monopoly Abuse by AI Overviews

Brief:

• Renowned U.S. media conglomerate Penske Media Corporation (owner of publications like Rolling Stone and Variety) has officially sued Google, alleging its "AI Overviews" feature constitutes copyright infringement by using its news content to generate summaries without authorization. 🚨
• Penske claims that this feature, by directly providing information within search results, has led to a significant decline in its website traffic, advertising, and affiliate revenue since 2025, with affiliate ad revenue decreasing by over one-third year-over-year. 📉
• The lawsuit further accuses Google of leveraging its monopoly in the search market to coerce publishers into allowing their content to be used for AI training, threatening exclusion from Google Search if they refuse. 😠
• Google has responded, stating that the AI feature brings new traffic to websites and that it will defend itself against these "baseless accusations."

Related Links:

• Google's AI Summary Feature Causes Big Trouble! Renowned Media Group Sues Monopoly Giant for "Stealing Content and Ruining Business"
• Penske Group Sues Google, Alleging AI Summary Infringement and Threat to Digital Media Future
• Rolling Stone Magazine's Parent Company Sues Google Over AI Overviews, Claiming a Dilemma

谷歌Gemini应用凭图像功能Nano Banana登顶App Store榜首

简报：

• 谷歌推出的Gemini应用在美国、印度、加拿大等多地登顶苹果App Store免费应用榜，超越了OpenAI的ChatGPT。
• 其内置的图像编辑工具“Nano Banana”被认为是此次用户增长的关键驱动力，因其能精准保持人物特征一致性而广受欢迎。
• 自8月26日发布以来，Gemini在短时间内新增2300万用户，Nano Banana功能被使用了超过5亿次。
• 免费用户每天可使用Nano Banana生成或编辑100张图片，付费订阅用户额度可提升至1000张。

相关链接：

• 谷歌Gemini火速登顶：iPhone用户的免费应用新宠
• 谷歌靠 Nano Banana 超越 ChatGPT，登顶苹果 App Store 第一

Google Gemini App Soars to Top of App Store, Driven by Image Feature Nano Banana

Brief:

• Google's Gemini app has topped the Apple App Store's free app charts in multiple regions, including the US, India, and Canada, surpassing OpenAI's ChatGPT. 🚀
• Its built-in image editing tool, "Nano Banana," is considered a key driver of this user growth, gaining widespread popularity for its ability to precisely maintain consistent character features. ✨
• Since its release on August 26, Gemini has gained 23 million new users in a short period, with the Nano Banana feature being used over 500 million times. 🍌
• Free users can generate or edit 100 images daily using Nano Banana, while paid subscribers can increase their allowance to 1,000 images.

Related Links:

• Google Gemini's Rapid Ascent: iPhone Users' New Favorite Free App
• Google Surpasses ChatGPT with Nano Banana, Reaching #1 on Apple App Store

字节跳动与清华大学联手，开源人体视频生成框架HuMo

简报：

• 字节跳动智能创作团队与清华大学联合推出并开源了一款名为HuMo的多模态框架，专注于人体视频生成（HCVG）领域。
• HuMo框架支持同时利用文本、图像和音频三种信息作为输入，通过渐进式训练方法，可生成长达97帧、分辨率高达720P的可控人物视频。

相关链接：

• 字节跳动与清华大学携手开源前沿多模态框架HuMo
• 相关论文

ByteDance and Tsinghua University Collaborate to Open-Source Human Video Generation Framework HuMo

Brief:

• ByteDance's Smart Creation Team and Tsinghua University have jointly launched and open-sourced HuMo, a multi-modal framework specializing in Human-Centric Video Generation (HCVG). 🚀
• The HuMo framework supports simultaneously using text, image, and audio as input. Through a progressive training method, it can generate controllable human videos up to 97 frames long with a resolution of up to 720P. ✨

Related Links:

• ByteDance and Tsinghua University Collaborate to Open-Source Cutting-Edge Multi-modal Framework HuMo
• Related Paper 📝

Nvidia市值登頂，黃仁勳因低持股比例財富排名未進前列

简报：

• Nvidia已成為全球市值最高的公司，但其創辦人暨執行長黃仁勳在《彭博億萬富豪榜》上僅排名第十。
• 主要原因是黃仁勳的持股比例因長期「股權稀釋」已降至不足4%，遠低於馬斯克（Tesla持股約13%）、阿爾諾（LVMH持股約50%）等其他頂級富豪。
• 股權稀釋源於公司為吸引和留住人才，長期發放股票作為薪酬獎勵，導致流通股數大幅增加，創辦人持股比例下降。
• 黃仁勳在公司1999年上市前曾持有12.8%的股份，若該比例維持至今，其財富將足以成為全球首富。

相关链接：

• 新聞幕後／Nvidia全球市值第一！黃仁勳卻只排富豪榜第十？原因曝光

Nvidia Climbs to Top Market Cap, Jensen Huang's Wealth Ranking Not Among the Elite Due to Low Stake 🚀

Briefing:

• Nvidia has become the world's most valuable company, yet its founder and CEO, Jensen Huang, ranks only tenth on the Bloomberg Billionaires Index. 🤔
• The primary reason is that Jensen Huang's ownership stake has fallen to less than 4% due to long-term 'equity dilution,' significantly lower than other top billionaires like Elon Musk (approx. 13% stake in Tesla) and Bernard Arnault (approx. 50% stake in LVMH).
• Equity dilution stems from the company's long-term strategy of issuing shares as compensation to attract and retain talent, leading to a significant increase in outstanding shares and a decrease in the founder's ownership percentage.
• Before the company went public in 1999, Jensen Huang held a 12.8% stake. Had that percentage been maintained to this day, his wealth would be enough to make him the richest person in the world. 💰

Related Link:

• Behind the News / Nvidia is the World's Number One by Market Cap! But Jensen Huang Only Ranks Tenth on the Rich List? The Reason is Revealed

鸿海聚焦化合物半导体，应对AI服务器能耗挑战

简报：

• 鸿海于“功率暨光电半导体论坛”指出，为应对生成式AI带来的庞大运算需求与能源挑战，化合物半导体是提升服务器能源效率的关键。
• 鸿海S事业群总经理陈伟铭强调，碳化矽（SiC）与氮化镓（GaN）是AI服务器与高效能运算的核心基础，其市场规模预估将从170亿美元增长至2030年的300亿美元。
• 鸿海研究院公布了与阳明交通大学、伊利诺大学在第四代半导体氧化镓（β-Ga₂O₃）上的研发突破，该技术未来可望应用于AI电源架构，提供高功率解决方案。

相关链接：

• AI 用電需求暴增！鴻海談伺服器怎麼省電　化合物半導體是關鍵

Foxconn Focuses on Compound Semiconductors to Address AI Server Energy Consumption Challenges

Brief:

• At the "Power and Optoelectronics Semiconductor Forum," Foxconn highlighted that compound semiconductors are crucial for enhancing server energy efficiency, especially in response to the massive computational demands and energy challenges posed by generative AI. ⚡️
• Chen Wei-ming, General Manager of Foxconn's S Business Group, emphasized that Silicon Carbide (SiC) and Gallium Nitride (GaN) are core foundations for AI servers and high-performance computing, with their market size projected to grow from $17 billion to$30 billion by 2030. 📈
• The Foxconn Research Institute announced R&D breakthroughs in fourth-generation semiconductor Gallium Oxide (β-Ga₂O₃) in collaboration with National Yang Ming Chiao Tung University and the University of Illinois. This technology is anticipated to be integrated into future AI power architectures, providing high-power solutions. 💪

Related Links:

• AI Power Demand Soars! Foxconn Discusses Server Energy Saving: Compound Semiconductors Are Key