单词学习：![[fcc3fbd2-297d-4839-afbd-6cee7c52bf5d.mp3]]

新闻播报：![[b5d191e3-79ce-4481-add8-0f170789b4b3.mp3]]

OpenAI斥资380亿美元与AWS达成合作，寻求算力多元化

简报：

• OpenAI宣布与亚马逊AWS达成一项价值380亿美元的长期合作协议，以拓展其AI云基础设施，满足日益增长的算力需求。
• 根据为期七年的协议，AWS将向OpenAI提供包括数千块NVIDIA GB200及GB300高性能GPU在内的计算资源，以支持其大规模AI模型训练与推理。
• 此前，OpenAI主要依赖微软Azure云服务，但微软已调整合作条款，允许OpenAI将ChatGPT和Sora等非API产品托管在其他云平台。
• 亚马逊同时指出，OpenAI的gpt-oss系列模型已是其Amazon Bedrock平台上最受欢迎的公有模型之一。

相关链接：

• OpenAI 与亚马逊 AWS 达成 380 亿美元合作，拓展 AI 云基础设施

OpenAI Partners with AWS in $38 Billion Deal, Seeks Compute Diversification 🚀

Brief:

• OpenAI announced a long-term partnership agreement with Amazon AWS, valued at $38 billion, to expand its AI cloud infrastructure and meet its growing computing power demands. 💡
• Under the seven-year agreement, AWS will provide OpenAI with computing resources, including thousands of high-performance NVIDIA GB200 and GB300 GPUs, to support its large-scale AI model training and inference. 💪
• Previously, OpenAI primarily relied on Microsoft Azure cloud services. However, Microsoft has adjusted its partnership terms, allowing OpenAI to host non-API products like ChatGPT and Sora on other cloud platforms.
• Amazon also noted that OpenAI's gpt-oss series models are already among the most popular public models on its Amazon Bedrock platform.

Related Link:

• OpenAI and Amazon AWS Reach $38 Billion Partnership to Expand AI Cloud Infrastructure

微软斥资超百亿美元，采购英伟达AI云算力

简报：

• 微软与AI云服务商IREN及Lambda签署两份多年期协议，合计金额超100亿美元，用于采购英伟达AI芯片算力。
• 与IREN的合同价值约97亿美元，微软将获得为期五年的英伟达GB300 GPU使用权。
• 与Lambda的协议价值数十亿美元，涉及为微软部署数万块英伟达GPU，其中包括GB300 NVL72机架级系统。
• 此项投资的背景是，微软采用英伟达GB300的Azure虚拟机在AI推理速度上创下行业新纪录，性能较上一代有显著提升。

相关链接：

• 加码英伟达 AI 芯片，微软再签两笔合计超百亿美元云算力订单
• 微软 Azure ND GB300 再创记录：每秒推理速度达 110 万token
• 掘金AI基建!Lambda与微软达成数十亿美元GPU部署协议

Microsoft Invests Over dJyH6kDkfC5RlM16.answerText0 Billion in NVIDIA AI Cloud Computing Power

Brief:

• Microsoft has signed two multi-year agreements with AI cloud service providers IREN and Lambda, totaling over dJyH6kDkfC5RlM16.answerText0 billion, to procure NVIDIA AI chip computing power. 💰
• The contract with IREN is valued at approximately $9.7 billion, granting Microsoft access to NVIDIA GB300 GPUs for five years.
• The agreement with Lambda is worth billions of dollars and involves deploying tens of thousands of NVIDIA GPUs for Microsoft, including GB300 NVL72 rack-scale systems.
• This investment comes as Microsoft's Azure virtual machines, powered by NVIDIA GB300, set new industry records for AI inference speed, showing significant performance improvements over the previous generation. 🚀📈

Related Links:

• Microsoft further invests in NVIDIA AI chips, signing two more cloud computing power orders totaling over dJyH6kDkfC5RlM16.answerText0 billion
• Microsoft Azure ND GB300 sets new record: 1.1 million tokens per second inference speed
• Investing in AI infrastructure! Lambda reaches multi-billion dollar GPU deployment agreement with Microsoft

黑客利用OpenAI API打造新型后门

简报：

• 微软安全研究团队发现名为“SesameOp”的新型后门恶意软件，该软件被证实首次滥用OpenAI的Assistants API作为其隐蔽的指挥与控制（C2）通道。
• 恶意软件通过该API获取加密指令并回传窃取的数据，利用合法云服务实现长期潜伏，极大增加了检测难度。
• 攻击者利用的是API的内置功能，并非OpenAI平台的安全漏洞。
• 微软已与OpenAI合作，禁用了攻击者使用的账户及API密钥，并指出被滥用的Assistants API计划于2026年8月弃用。

相关链接：

• 微软示警：黑客武器化 OpenAI API，打造“无法检测”的后门

Hackers Weaponize OpenAI API to Create New Backdoor 🤖

Briefing:

• Microsoft's security research team has discovered a new backdoor malware named 'SesameOp' 🕵️, confirmed to be the first to abuse OpenAI's Assistants API as its covert Command and Control (C2) channel.
• The malware uses this API to retrieve encrypted instructions and exfiltrate stolen data, leveraging legitimate cloud services for long-term stealth, significantly increasing detection difficulty.
• The attackers exploited built-in functionalities of the API, not a security vulnerability in OpenAI's platform.
• Microsoft has collaborated with OpenAI to disable the accounts and API keys used by the attackers. They also noted that the abused Assistants API is scheduled for deprecation in August 2026. 🗓️

Related Link:

• Microsoft Warns: Hackers Weaponize OpenAI API, Creating an 'Undetectable' Backdoor

吉卜力、万代等日本公司联合要求OpenAI停止侵权训练

简报：

• 日本内容海外分发协会（CODA）已向OpenAI发函，要求其停止未经事先许可，使用协会成员的创意作品来训练其AI模型，特别是视频生成工具Sora 2。
• 该协会成员包括吉卜力工作室、万代南梦宫、Square Enix、东映动画及Cygames等众多日本知名游戏和动画公司。
• CODA指出，OpenAI默认使用版权作品进行训练的“选择退出”机制，违反了日本《版权法》中要求“事先许可”的规定。
• 协会强调，在日本法律体系中，不存在通过事后提出异议来规避侵权责任的机制，并要求OpenAI对其输出结果涉及的版权侵权质疑做出真诚回应。

相关链接：

• ​日本版权协会要求 OpenAI 停止使用其作品进行 AI 训练
• 万代南梦宫、SE、东映结盟，要求 OpenAI 停止使用其作品训练 Sora 2

Japanese Companies, Including Studio Ghibli and Bandai, Jointly Demand OpenAI Cease Infringement Training 🚨

Brief:

• The Content Overseas Distribution Association (CODA) of Japan has sent a letter to OpenAI, demanding that it stop using the creative works of its members to train its AI models, especially the video generation tool Sora 2, without prior permission.
• Members of the association include numerous renowned Japanese game and animation companies such as Studio Ghibli, Bandai Namco, Square Enix, Toei Animation, and Cygames. 🎮
• CODA points out that OpenAI's "opt-out" mechanism, which defaults to using copyrighted works for training, violates the "prior permission" requirement stipulated in Japan's Copyright Law.
• The association emphasizes that under the Japanese legal system, there is no mechanism to circumvent infringement liability by raising objections after the fact, and demands that OpenAI genuinely respond to queries regarding copyright infringement related to its output. ⚖️

Related Links:

• Japanese Copyright Association Demands OpenAI Stop Using Its Works for AI Training
• Bandai Namco, SE, Toei Form Alliance, Demand OpenAI Stop Using Their Works to Train Sora 2

Kimi Linear模型发布，月之暗面大幅提升长文本处理效率

简报：

• 月之暗面推出Kimi Linear模型，通过采用混合线性注意力架构，在处理长上下文时速度提高了2.9倍，解码速度提升6倍。
• 该模型的核心创新是Kimi Delta Attention（KDA），它通过引入细粒度的门控机制，改善了模型的记忆管理能力，能动态控制信息的遗忘与保留。
• Kimi Linear采用了Moonlight架构，将KDA与全注意力层以3:1的比例进行混合，以在计算效率与模型能力之间取得平衡。
• 新模型突破了传统Softmax注意力机制O(n²)的计算复杂度瓶颈，将处理长文本的效率大幅提升。

相关链接：

• 月之暗面推Kimi Linear模型：处理长上下文速度提高了2.9倍

Moonshot AI Unveils Kimi Linear Model, Dramatically Boosting Long-Text Processing Efficiency 🚀

Briefing:

• Moonshot AI has launched the Kimi Linear model, which leverages a hybrid linear attention architecture to achieve a 2.9x speed increase in long-context processing and a 6x boost in decoding speed. ✨
• The model's core innovation is Kimi Delta Attention (KDA), which improves the model's memory management capabilities through a fine-grained gating mechanism, dynamically controlling information forgetting and retention.
• Kimi Linear employs the Moonlight architecture, blending KDA with full attention layers at a 3:1 ratio to strike a balance between computational efficiency and model capability.
• The new model breaks through the O(n²) computational complexity bottleneck of traditional Softmax attention mechanisms, significantly boosting long-text processing efficiency. 💡

Related Links:

• Moonshot AI Unveils Kimi Linear Model: 2.9x Faster Long-Context Processing

AI实盘投资赛落幕：阿里通义千问夺冠，GPT-5巨亏超62%

简报：

• 在美国研究机构Nof1举办的首届AI大模型真实投资比赛Alpha Arena中，阿里通义千问Qwen3-Max以22.32%的收益率获得冠军。
• 比赛为Qwen3-Max、DeepSeek v3.1、GPT-5等六大顶尖AI模型各提供1万美元起始资金，在真实市场中自主交易加密货币。
• 赛果显示，除通义千问和DeepSeek v3.1盈利外，其余四个模型均亏损，其中GPT-5亏损率超过62%。
• 该比赛旨在动态、竞争的真实环境中检验AI模型在量化交易领域的零样本能力，模型决策仅基于数值市场数据，不能查阅新闻。
• 研究团队表示，此举意在推动AI研究从静态基准测试转向对“真实世界”、“实时决策”能力的考察。

相关链接：

• 首届 AI 大模型真实投资比赛 Alpha Arena 落幕：阿里通义千问 Qwen3-Max 以 22.32% 收益率夺冠，GPT-5 亏损超 62%

AI Live Trading Competition Concludes: Alibaba's Qwen3-Max Crowned Champion 🏆, GPT-5 Suffers Over 62% Loss

Brief:

• In the inaugural AI large model live investment competition "Alpha Arena" organized by U.S. research institution Nof1, Alibaba's Qwen3-Max secured the championship with a 22.32% return on investment 🎉.
• The competition provided six top AI models, including Qwen3-Max, DeepSeek v3.1, and GPT-5, with dJyH6kDkfC5RlM16.answerText0,000 in starting capital each, to autonomously trade cryptocurrencies in a real market.
• The results showed that, apart from Qwen3-Max and DeepSeek v3.1 which generated profits, the other four models incurred losses, with GPT-5's loss exceeding 62% 💸.
• The competition aimed to test AI models' zero-shot capabilities in quantitative trading within a dynamic, competitive real-world environment. Model decisions were based solely on numerical market data, without access to news.
• The research team stated that this initiative intends to shift AI research from static benchmarks to an evaluation of "real-world," "real-time decision-making" capabilities.

Related Link:

• Inaugural AI Large Model Live Investment Competition Alpha Arena Concludes: Alibaba's Qwen3-Max Wins with 22.32% Return, GPT-5 Loses Over 62%

高德布局Robotaxi，拓展全球出行服务

简报：地图导航巨头高德宣布正式入局Robotaxi（无人驾驶出租车）领域，并将其定位为公司战略级新业务，计划利用其在高精地图和出行服务方面的现有优势拓展全球市场，以寻求新的增长点。

相关链接：

• 战略级新增长点!地图巨头高德启动无人驾驶出租车业务

Amap Enters Robotaxi Sector, Expands Global Mobility Services

Brief: Map and navigation giant Amap has officially announced its entry into the Robotaxi (driverless taxi) sector, positioning it as a new strategic business for the company. Amap plans to leverage its existing strengths in high-precision mapping and mobility services to expand into global markets, seeking new growth opportunities. 🌍🚗✨

Related Link:

• Strategic New Growth Point! Map Giant Amap Launches Driverless Taxi Business