单词学习：![[0948ea86-3b92-4687-aca5-37f2e38ab365.mp3]]

新闻播报：![[71b50401-5b04-4359-960b-22a782d43514.mp3]]

西藏AI迈入新阶段：首个千亿级藏语大模型“阳光清言”发布

简报：随着千亿级参数藏语基座大模型“阳光清言”V1.0的发布，西藏人工智能发展正从场景应用转向系统性研发阶段，该模型依托约288亿Token覆盖多个领域的高质量藏语数据完成训练。

相关链接：

• 原文链接

Tibet's AI Enters New Stage: First 100-Billion-Parameter Tibetan Large Model "Sunlight Qingyan" Released 🚀

Brief: With the release of the 100-billion-parameter Tibetan foundation large model "Sunlight Qingyan" V1.0, Tibet's artificial intelligence development is transitioning from scenario-based applications to a stage of systematic research and development. The model was trained using approximately 28.8 billion tokens of high-quality Tibetan data covering multiple domains. ✨💡

Related Links:

• Original Article Link

谷歌TPU芯片外销计划曝光，目标2027年打破英伟达垄断

简报：

• 根据摩根士丹利研报，谷歌计划将其TPU芯片商业模式从自用转向对外销售，目标到2027年出货100万颗。
• 预计50万颗芯片即可为谷歌云业务带来约130亿美元的增量收入，并显著提升每股收益。
• Meta是首批重点潜在客户之一，正在评估一项价值数十亿美元的采购订单，用于新建AI数据中心。
• 谷歌希望借此打破英伟达在AI芯片市场的垄断地位，若计划顺利，将成为继英伟达后第二家靠AI芯片直接创收的大型云厂商。
• 分析师提醒，该计划能否成功取决于谷歌能否解决定价、软件栈兼容性及渠道建设等挑战。

相关链接：

• 相关报道1
• 相关报道2
• 相关报道3
• 相关报道4
• 相关报道5

Google's TPU Chip Sales Plan Revealed, Aims to Break Nvidia's Monopoly by 2027

Briefing:

• According to a Morgan Stanley research report, Google plans to shift its TPU chip business model from internal use to external sales, targeting 1 million units shipped by 2027. 📈
• An estimated 500,000 chips could generate approximately dJyH6kDkfC5RlM16.answerText3 billion in incremental revenue for Google Cloud and significantly boost earnings per share.
• Meta is among the first key potential customers, currently evaluating a multi-billion dollar procurement order for new AI data centers. 💰
• Google hopes to leverage this move to break Nvidia's monopoly in the AI chip market. If successful, it would become the second major cloud vendor, after Nvidia, to directly generate revenue from AI chips.
• Analysts caution that the plan's success hinges on Google's ability to address challenges such as pricing, software stack compatibility, and channel development. 🤔

Related Links:

• Related Report 1
• Related Report 2
• Related Report 3
• Related Report 4
• Related Report 5

Sora2被曝监管漏洞，低龄账号可生成暴力视频

简报：

• 消费者监督组织Ekō发布报告揭示，OpenAI的视频生成模型Sora2存在监管漏洞，允许以13岁青少年身份注册的账号通过简单提示词生成描绘枪支暴力与吸毒等违反其政策的敏感视频。

相关链接：

• 研究揭示 OpenAI Sora2 监管漏洞：13 岁账号即可轻易生成“校园枪击”视频
• 收到反 AI 极端分子暴力威胁后，OpenAI 封锁旧金山办公室
• 需求塞爆服务器，OpenAI、谷歌对 Sora 和 Nano Banana Pro 采取“限流”措施
• 最新测试版 App 代码显示，ChatGPT 未来或引入广告

Sora2 Exposed for Regulatory Loopholes, Underage Accounts Can Generate Violent Videos 😨

Brief:

• Consumer watchdog organization Ekō has released a report revealing regulatory loopholes in OpenAI's video generation model, Sora2. The report indicates that accounts registered as 13-year-olds can generate sensitive videos depicting gun violence and drug use, violating Sora2's policies, through simple text prompts. 🚨

Related Links:

• Research Reveals OpenAI Sora2 Regulatory Loophole: 13-Year-Old Accounts Can Easily Generate "School Shooting" Videos
• OpenAI Locks Down San Francisco Office After Receiving Violent Threats from Anti-AI Extremists
• Servers Overwhelmed by Demand, OpenAI and Google Implement "Throttling" Measures for Sora and Nano Banana Pro
• [Latest Test Version App Code Shows ChatGPT May Introduce Ads in the Future] 😮

AI生成服务需求火爆，OpenAI与谷歌开始限制免费使用次数

简报：

• 由于用户需求激增导致GPU资源紧张，OpenAI与谷歌已分别对旗下的Sora和Nano Banana Pro服务采取更严格的生成次数限制。
• OpenAI将Sora免费用户的每日视频生成上限设为六段，并表示“GPU快被撑爆了”，同时用户可以按需购买额外生成次数。
• 谷歌将Nano Banana Pro免费用户的每日图片生成额度从三张下调至两张，并提醒用户额度可能会随时调整且不提前通知。

相关链接：

• 需求塞爆服务器，OpenAI、谷歌对 Sora 和 Nano Banana Pro 采取“限流”措施

Demand for AI Generative Services Soars, OpenAI and Google Start Limiting Free Usage 🔥

Brief:

• Due to surging user demand leading to strained GPU resources, OpenAI and Google have independently implemented stricter generation limits for their respective Sora and Nano Banana Pro services.
• OpenAI has set a daily video generation limit of six segments for free Sora users, stating that "GPUs are nearly bursting at the seams" 🤯, while also allowing users to purchase additional generations as needed.
• Google has reduced the daily image generation quota for free Nano Banana Pro users from three to two, reminding users that quotas may be adjusted at any time without prior notice.

Related Link:

• Servers Overloaded by Demand, OpenAI and Google Implement "Traffic Control" for Sora and Nano Banana Pro 🚀

应对需求激增，谷歌调整Gemini 3 Pro免费额度

简报：

• 因Gemini 3 Pro及图像生成工具Nano Banana Pro上线后用户需求激增，谷歌面临容量压力，决定收紧免费用户的使用权限。
• Gemini 3 Pro的免费用户被调整为“基础访问权限”，“思考”功能的使用限额可能频繁变化。
• 图像生成工具Nano Banana Pro的每日免费使用上限从3张收紧至2张，并可能根据需求随时调整。
• NotebookLM平台中由Nano Banana Pro支持的“信息图”和“幻灯片套组”功能已对免费用户暂时关闭。
• 此次使用限制的调整不影响Google AI Pro或Ultra的付费订阅用户。

相关链接：

• 谷歌因用户激增收紧 Gemini 3 Pro 免费使用限制，用户体验或将受影响！
• 百万 TPU 外销计划曝光！谷歌 2027 年欲把 AI 芯片卖给 Meta，云收入或暴增 130 亿美元

Amid Surging Demand, Google Adjusts Free Tier for Gemini 3 Pro

Brief:

• Due to a surge in user demand following the launch of Gemini 3 Pro and its image generation tool Nano Banana Pro, Google is facing capacity pressure and has decided to tighten usage restrictions for free users. 🚀
• Free users of Gemini 3 Pro have been reclassified to "basic access," and usage limits for the "Think" feature may change frequently.
• The daily free usage limit for the image generation tool Nano Banana Pro has been tightened from 3 images to 2 images and may be adjusted at any time based on demand. 🍌
• The "Infographics" and "Slide Decks" features in the NotebookLM platform, powered by Nano Banana Pro, have been temporarily disabled for free users.
• These adjustments to usage restrictions do not affect paid subscribers of Google AI Pro or Ultra. 💎

Related Links:

• Google Tightens Gemini 3 Pro Free Usage Limits Due to User Surge, User Experience May Be Affected!
• Million-Unit TPU Export Plan Revealed! Google Aims to Sell AI Chips to Meta by 2027, Cloud Revenue Could Soar by dJyH6kDkfC5RlM16.answerText3 Billion

阿里通义千问凭“门控注意力”研究获NeurIPS 2025最佳论文奖

简报：

• 阿里巴巴通义千问团队凭借论文《Attention Gating Makes Better Foundation Models》获得全球人工智能顶会NeurIPS 2025的最佳论文奖，是四篇获奖论文中唯一的中国团队。
• 该研究提出了一种“门控注意力”机制，在标准注意力层后增加一个可学习的门控，以决定哪些信息可以进入下游计算，从而提高效率与鲁棒性。
• 实验数据显示，该机制在仅增加1%参数的情况下，能使模型MMLU得分提升2分，困惑度降低0.2。
• 阿里方面表示，该技术将被整合到即将发布的Qwen3-Next模型中，并已将相关代码与实验模型开源。

相关链接：

• NeurIPS 2025 最佳论文揭晓：阿里通义千问靠“门控注意力”摘下中国唯一大奖
• 阿里通义推出 Z-Image 模型，首日下载量突破 50 万

Alibaba Tongyi Qianwen Wins NeurIPS 2025 Best Paper Award for "Gated Attention" Research

Briefing:

• The Alibaba Tongyi Qianwen team secured the Best Paper Award at NeurIPS 2025, a top global AI conference, for their paper "Attention Gating Makes Better Foundation Models," making them the only Chinese team among the four award-winning papers. 🏆
• The research proposes a "Gated Attention" mechanism, which adds a learnable gate after the standard attention layer to determine which information proceeds to downstream computations, thereby enhancing efficiency and robustness.
• Experimental data shows that this mechanism, with only a 1% increase in parameters, can improve the model's MMLU score by 2 points and reduce perplexity by 0.2. 📈
• Alibaba stated that this technology will be integrated into the upcoming Qwen3-Next model, and that the related code and experimental models have already been open-sourced. 💻

Related Links:

• NeurIPS 2025 最佳论文揭晓：阿里通义千问靠“门控注意力”摘下中国唯一大奖
• 阿里通义推出 Z-Image 模型，首日下载量突破 50 万

Meta开源CoT-Verifier，可剖析并修正AI推理错误

简报：Meta推出基于Llama 3.1的开源工具CoT-Verifier，该工具通过将AI的链式思维（CoT）推理过程转化为归因图，分析其结构来诊断、定位并干预推理中的错误步骤，从而在无需重训主干模型的情况下提升了模型准确率。

相关链接：

• Meta 甩出“白盒手术刀”：CoT-Verifier 把 AI 推理失误钉在归因图上

Meta Open-Sources CoT-Verifier to Dissect and Correct AI Reasoning Errors 🛠️

Brief: Meta has launched CoT-Verifier, an open-source tool built on Llama 3.1. This innovative tool transforms an AI's Chain-of-Thought (CoT) reasoning process into an attribution graph, analyzing its structure to diagnose, locate, and intervene in erroneous steps within the reasoning. This method effectively enhances model accuracy without the need to retrain the core backbone model. 💡✅

Related Links:

• Meta Unveils a "White-Box Scalpel": CoT-Verifier Pins AI Reasoning Errors on Attribution Graphs

雷军谈AI赋能产业：未来五年人形机器人将进入工厂

简报：

• 小米集团创始人雷军表示，所有产业都值得用AI再做一遍，预测未来五年人工智能将深度融合传统产业，开启一个新的万亿级大市场。
• 雷军以小米汽车工厂为例，称通过AI视觉大模型检测大压铸件，检测时间缩短至2秒内，效率是人工的10倍，精度提升5倍以上。
• 他预测，未来5年内人形机器人将大面积在工厂上岗，并指出家庭对人形机器人的需求将创造更大的市场空间。
• 雷军强调，产业升级需要与最强的伙伴合作，没有任何一家公司能独立掌握所有环节。

相关链接：

• 雷军：所有产业都值得用AI重做一遍

Lei Jun on AI Empowering Industries: Humanoid Robots to Enter Factories in the Next Five Years

Brief:

• Lei Jun, founder of Xiaomi Group, stated that all industries are worth redoing with AI. He predicted that in the next five years, artificial intelligence will deeply integrate with traditional industries, opening up a new trillion-dollar market. 🚀
• Taking Xiaomi's car factory as an example, Lei Jun mentioned that by using large AI vision models to inspect large die-cast parts, detection time has been reduced to within 2 seconds. This is 10 times more efficient than manual inspection and improves accuracy by over 5 times. ✨
• He predicted that humanoid robots will be widely deployed in factories within the next 5 years, adding that household demand for humanoid robots will create an even larger market space.
• Lei Jun emphasized that industrial upgrading requires collaboration with the strongest partners, as no single company can independently master all aspects.

Related Links:

• Lei Jun: All Industries Are Worth Redoing with AI 🤖