单词学习：![[a7c0b12e-6b0f-4258-bde9-33f03df7c042.mp3]]

新闻播报：![[0868b91d-d72d-47af-95c6-5c5e91822da7.mp3]]

雷军：AI将重塑所有产业，人形机器人五年内进厂

简报：

• 小米创始人雷军提出“所有产业都值得用AI再做一遍”的论断，认为人工智能与传统产业的深度融合将开启新的万亿级市场。
• 他以小米汽车工厂为例，通过AI视觉大模型检测大压铸件，检测时间缩至2秒内，效率是人工的10倍，精度提升5倍以上。
• 雷军预测，未来5年内人形机器人将大面积进入工厂，并指出工厂应用只是第一步，家庭对人形机器人的需求将更大。

相关链接：

• 雷军:所有产业都值得用AI重做一遍，人形机器人五年内将大面积进厂
• 快手旗舰 Keye-VL-671B-A37B 重磅发布，多模态推理能力实现新突破
• 深圳上线全国首个 AI 智能劳动仲裁系统 覆盖 111 个劳动争议类型
• 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
• 天风证券：全球算力供需缺口或将持续至2026年

Lei Jun: AI Will Reshape All Industries, Humanoid Robots to Enter Factories Within Five Years 🤖

Brief:

• Xiaomi founder Lei Jun proposed the assertion that "all industries are worth redoing with AI," believing that the deep integration of artificial intelligence with traditional industries will unlock a new trillion-dollar market. 💰
• He cited Xiaomi's auto factory as an example, where large die-cast parts are inspected using an AI vision large model. This has reduced detection time to under 2 seconds, achieving 10 times the efficiency of manual inspection and improving accuracy by over 5 times.
• Lei Jun predicted that humanoid robots will extensively enter factories within the next five years, noting that factory applications are just the first step, and the demand for humanoid robots in homes will be even greater. 🏡

Related links:

• Lei Jun: All Industries Are Worth Redoing with AI; Humanoid Robots Will Extensively Enter Factories Within Five Years
• Kuaishou Flagship Keye-VL-671B-A37B Grandly Launched, Achieving New Breakthroughs in Multimodal Reasoning Capabilities
• Shenzhen Launches China's First AI Intelligent Labor Arbitration System, Covering 111 Types of Labor Disputes
• NeurIPS 2025 Best Paper Revealed: Alibaba's Tongyi Qianwen Wins China's Only Award with 'Gated Attention'
• Tianfeng Securities: Global Computing Power Supply-Demand Gap May Persist Until 2026

通义千问“门控注意力”获NeurIPS最佳论文，成唯一中国获奖者

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AI顶会NeurIPS 2025最佳论文奖，是4篇获奖论文中唯一的中国团队。
• 该研究提出的“门控注意力”机制，在标准注意力层后增加可学习门控以过滤无效信息，实验证明在仅增加1%参数的情况下，模型MMLU得分可提升2分，困惑度降低0.2。
• 该技术将被集成到即将发布的Qwen3-Next模型中，相关代码与1.7B实验模型已在GitHub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
• 快手旗舰 Keye-VL-671B-A37B 重磅发布，多模态推理能力实现新突破
• Meta 发布 CoT 验证模型：基于 Llama 3.1 的白盒推理纠错工具
• 国产“数学金牌”横空出世：DeepSeek-Math-V2 开源文件已上传，性能对标 GPT-4o

Tongyi Qianwen's "Gated Attention" Wins NeurIPS Best Paper, Becoming the Only Chinese Awardee 🏆

Briefing:

• The Alibaba Tongyi Qianwen team won the Best Paper Award at NeurIPS 2025, a premier global AI conference, for their paper "Attention Gating Makes Better Foundation Models." They are the only Chinese team among the four award-winning papers.
• The research introduces a "Gated Attention" mechanism, which adds learnable gates after standard attention layers to filter out irrelevant information. Experiments show that this mechanism improves the model's MMLU score by 2 points and reduces perplexity by 0.2, with only a 1% increase in parameters. 📈
• This technology will be integrated into the upcoming Qwen3-Next model. The related code and a 1.7B experimental model have already been open-sourced on GitHub. 💻

Related Links:

• NeurIPS 2025 Best Papers Announced: Alibaba Tongyi Qianwen Wins China's Sole Grand Prize with "Gated Attention"
• Alibaba Tongyi Launches Z-Image Model, Surpassing 500,000 Downloads on Day One
• Kuaishou's Flagship Keye-VL-671B-A37B Grandly Released, Achieving New Breakthroughs in Multimodal Reasoning Capabilities
• Meta Releases CoT Verification Model: A White-box Inference Error Correction Tool Based on Llama 3.1
• Domestic "Math Gold Medalist" Emerges: DeepSeek-Math-V2 Open-source Files Uploaded, Performance Rivals GPT-4o

谷歌TPU芯片拟对外销售，剑指英伟达市场

简报：

• 根据摩根士丹利研报，谷歌计划将其TPU AI芯片商业模式从“自用”转向“对外创收”，旨在打破英伟达的市场垄断。
• 计划目标是到2027年出货100万颗TPU，首批重点客户包括Meta，后者正在评估一笔数十亿美元的采购订单。
• 报告预测，仅销售50万颗TPU即可为谷歌云业务带来约130亿美元的增量收入，占其云总收入的11%。
• 分析师认为，与英伟达GPU相比，TPU在特定大模型训练和推理场景中具有功耗与成本优势。

相关链接：

• 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

Google's TPU Chips Go Commercial, Eyeing Nvidia's Dominance 🚀

Briefing:

• According to a Morgan Stanley research report, Google plans to pivot its TPU AI chip business model from "internal use" to "external revenue generation," aiming to disrupt Nvidia's market monopoly. This marks a significant strategic shift! 💡
• The ambitious plan targets shipping 1 million TPUs by 2027. Initial key customers reportedly include Meta, which is currently evaluating a multi-billion-dollar procurement order.
• The report forecasts that selling just 500,000 TPUs could generate approximately dJyH6kDkfC5RlM16.answerText3 billion in incremental revenue for Google Cloud, accounting for 11% of its total cloud revenue. 💰
• Analysts believe that compared to Nvidia's GPUs, TPUs offer power efficiency and cost advantages in specific large model training and inference scenarios.

Related Links:

• 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

无问芯穹获近5亿A+轮融资，加码Agentic AI基础设施

简报：

• AI基础设施公司无问芯穹于近日宣布完成近5亿元A+轮融资。
• 本轮融资由珠海科技集团和孚腾资本(元创未来基金)联合领投，惠远资本、尚颀资本、弘晖基金跟投，老股东洪泰基金、达晨财智、联想创投等也参与了追投。
• 募集资金将重点用于巩固软硬协同与异构计算技术、推动AI云产品与终端解决方案的规模化应用，并大力投入Agentic AI（智能体AI）基础设施的研发。
• 公司表示，未来将以构建新一代可学习、可进化的Agentic Infra为战略核心，加速智能体在数字世界与物理世界的规模化落地。

相关链接：

• 无问芯穹获近5亿元A+轮融资：加速构建Agentic AI基础设施

WENWENXINQIONG Secures Nearly 500 Million RMB in Series A+ Funding, Boosting Agentic AI Infrastructure 💰

Brief:

• WENWENXINQIONG, an AI infrastructure company, recently announced the completion of its Series A+ funding round, securing nearly 500 million RMB.
• This round was co-led by Zhuhai Technology Group and Futeng Capital (Yuanchuang Future Fund), with follow-on investments from Huiyuan Capital, Shangqi Capital, and Honghui Fund. Existing shareholders, including Hongtai Fund, Dalun Capital, and Lenovo Capital & Incubator Group, also participated in the additional investments.
• The raised capital will primarily be used to strengthen software-hardware synergy and heterogeneous computing technologies, promote the scaled application of AI cloud products and edge solutions, and significantly invest in the research and development of Agentic AI infrastructure.
• The company stated that its strategic core for the future will be to build next-generation learnable and evolvable Agentic Infra, accelerating the widespread deployment of agents in both the digital and physical worlds. 🚀🌐

Related Links:

• WENWENXINQIONG Secures Nearly 500 Million RMB in Series A+ Funding: Accelerating the Construction of Agentic AI Infrastructure

AI算力需求迈向千万卡级别，供需缺口预计延续至2026年

简报：

• 天风证券在2026年度策略会上表示，AI科技产业的核心驱动力源于大规模算力，全球算力供需缺口或将持续至2026年。
• 报告指出，算力需求的指数级增长已成为产业发展的“地基逻辑”，计算集群规模正从万卡向十万卡、百万卡乃至千万卡升级。
• 需求的旺盛体现在产业层面，例如OpenAI为满足算力需求，其合作伙伴已举债近千亿美元建设AI基础设施。
• 同时，AI数据中心的扩张也导致美国2025年电力需求激增，尽管光伏发电的爆发式增长抵消了超80%的新增需求。

相关链接：

• 天风证券：全球算力供需缺口或将持续至2026年

AI Compute Power Demand Reaching Tens of Millions of Cards, Supply-Demand Gap Expected to Persist Until 2026 📈

Brief:

• Tianfeng Securities stated at its 2026 annual strategy conference that the core driving force of the AI tech industry stems from large-scale computing power, and the global supply-demand gap for compute power may continue until 2026.
• The report points out that the exponential growth in compute power demand has become the "foundational logic" for industrial development. Computing cluster scales are upgrading from tens of thousands of cards to hundreds of thousands, millions, and even tens of millions of cards. 🚀
• The robust demand is reflected at the industry level. For example, to meet its compute power needs, OpenAI's partners have borrowed nearly dJyH6kDkfC5RlM16.answerText00 billion to build AI infrastructure.
• Concurrently, the expansion of AI data centers has led to a surge in U.S. electricity demand for 2025, although the explosive growth in photovoltaic power generation has offset over 80% of the new demand. ⚡

Related Links:

• Tianfeng Securities: Global Compute Power Supply-Demand Gap Expected to Persist Until 2026