单词学习：![[9266253e-9aea-4af0-a839-ed8dd6c883b1.mp3]]

新闻播报：![[ea42bae2-c3cb-46d9-a01e-0929f55fef95.mp3]]

字节跳动推出120亿参数视频模型Vidi2，可自动剪辑长视频

简报：

• 字节跳动发布专用于视频理解的多模态大语言模型Vidi2，拥有120亿参数。
• 该模型新增“精细时空定位（STG）”功能，可同时识别视频中的时间戳和目标对象的边界框。
• Vidi2能够处理数小时长的原始素材，并根据简单提示生成完整的短视频或电影片段。
• 在超长视频（超过1小时）理解的基准测试中，其性能领先商业模型17.5个百分点。
• 相关技术已应用于TikTok的Smart Split功能，用于自动剪辑和内容重构。

相关链接：

• Vidi2论文地址

ByteDance Launches 12 Billion Parameter Video Model Vidi2, Capable of Automatic Long Video Editing

Brief:

• ByteDance has unveiled Vidi2, a multimodal large language model specifically designed for video understanding, boasting 12 billion parameters. 🎥
• The model introduces a new "Spatio-Temporal Grounding (STG)" feature, enabling simultaneous identification of timestamps and bounding boxes of target objects within videos. ⏱️
• Vidi2 can process raw footage lasting several hours and generate complete short videos or film clips based on simple prompts.
• In benchmark tests for ultra-long video (over 1 hour) understanding, its performance surpassed commercial models by 17.5 percentage points.
• The underlying technology has already been integrated into TikTok's Smart Split feature for automatic editing and content restructuring. ✨

Related Link:

• Vidi2 Paper

欧盟通过全球首部AI监管法，最高罚款可达全球营收7%

简报：

• 欧盟立法者已就全球首个全面人工智能法案达成协议，旨在为27个成员国建立一个基于风险等级的统一监管体系。
• 法案将全面禁止被认定为“不可接受风险”的AI应用，例如政府的社会评分系统以及在公共场所进行实时生物识别监控的执法行为。
• 对于基础设施和医疗设备中的高风险AI应用，法案规定了严格的义务，包括风险评估和高质量数据集要求。
• 法案对ChatGPT等通用AI模型采取双层监管方式：所有模型需遵守透明度要求，而被认定存在“系统性风险”的强大模型则需承担额外的评估和报告义务。
• 将设立一个新的欧洲人工智能办公室负责监督执行，违规企业将面临高达3500万欧元或全球年营收7%的罚款。

相关链接：

• 欧盟达成历史性协议：全球首个全面人工智能法案正式出台

EU Passes World's First AI Regulation Act, Fines Up to 7% of Global Revenue 🇪🇺🚨

Brief:

• EU lawmakers have reached an agreement on the world's first comprehensive Artificial Intelligence Act, aiming to establish a unified risk-based regulatory system across its 27 member states.
• The Act will outright ban AI applications deemed 'unacceptable risk,' such as government social scoring systems and real-time biometric identification surveillance by law enforcement in public spaces.
• For high-risk AI applications in infrastructure and medical devices, the Act mandates strict obligations, including risk assessments and requirements for high-quality datasets.
• The Act adopts a two-tiered regulatory approach for general-purpose AI models like ChatGPT: all models must comply with transparency requirements, while powerful models identified as posing 'systemic risks' will be subject to additional assessment and reporting obligations.
• A new European AI Office will be established to oversee enforcement, with non-compliant companies facing fines of up to €35 million or 7% of their global annual revenue. 💰

Related Links:

• EU Reaches Historic Agreement: World's First Comprehensive AI Act Officially Adopted

# SemiAnalysis报告：OpenAI陷入两年半前沿模型训练停滞

简报：

• 根据SemiAnalysis的报告，OpenAI在过去两年半内未能成功完成任何一次面向下一代前沿大模型的训练部署。
• 报告指出，多次扩大参数和数据规模的尝试均因收敛困难或性能倒挂而中途停止。
• 因此，备受期待的GPT-5系列实际上是GPT-4o的优化变体，并未实现架构上的突破。
• 这一停滞被行业视为AI进入“后Scaling时代”的信号，面临数据、算力与模型配方的三重瓶颈。
• 报告还提到，OpenAI利用评估谷歌TPU的可能性，迫使英伟达在GPU集群报价上让步约30%，从而节省了大量成本。

相关链接：

• 预训练“卡壳”：SemiAnalysis 爆料 OpenAI 连续两年半未跑通新前沿大模型
• 谷歌 TPU 发布引发 AI 计算市场价格战，OpenAI 节省 30% 成本
• Win11 Copilot 直接送“满血”GPT-5.1，深度思考功能免费解锁！

SemiAnalysis报告：OpenAI陷入两年半前沿模型训练停滞

简报：

• 据SemiAnalysis报告，OpenAI在过去两年半内未能成功完成任何一次面向下一代前沿大模型的训练部署，多次尝试因收敛困难等问题而中途停止。
• 报告指出，目前被外界期待的GPT-5系列，实质上是GPT-4o的优化变体，未能实现架构级的突破。
• 这一停滞被行业视为AI发展遭遇数据、算力与模型配方三重瓶颈，进入“后Scaling时代”的标志性信号。
• 报告同时披露，OpenAI通过评估谷歌TPU的可能，成功迫使英伟达在GPU集群报价上做出约30%的让步，从而节省了成本。

相关链接：

• 预训练“卡壳”：SemiAnalysis 爆料 OpenAI 连续两年半未跑通新前沿大模型
• 谷歌 TPU 发布引发 AI 计算市场价格战，OpenAI 节省 30% 成本
• Win11 Copilot 直接送“满血”GPT-5.1，深度思考功能免费解锁！

# SemiAnalysis Report: OpenAI Stalled on Frontier Model Training for Two and a Half Years

Brief:

• According to a SemiAnalysis report, OpenAI has failed to successfully complete any training deployment for next-generation frontier large language models over the past two and a half years. 📉
• The report indicates that multiple attempts to scale up parameters and data size were halted midway due to convergence difficulties or performance degradation. 🛑
• Consequently, the highly anticipated GPT-5 series is effectively an optimized variant of GPT-4o, failing to achieve an architectural breakthrough.
• This stagnation is viewed by the industry as a signal that AI is entering the "Post-Scaling Era," facing a triple bottleneck of data, computational power, and model recipes. 💡
• The report also reveals that OpenAI leveraged the possibility of evaluating Google TPUs, successfully compelling Nvidia to concede approximately 30% on GPU cluster pricing, thereby saving significant costs.

Related Links:

• Pre-training "Stuck": SemiAnalysis Reveals OpenAI Hasn't Completed New Frontier LLM Training for Two and a Half Years
• Google TPU Launch Triggers AI Computing Price War, OpenAI Saves 30% on Costs
• Win11 Copilot Delivers "Full-Fledged" GPT-5.1 Directly, Deep Thinking Features Unlocked for Free!

SemiAnalysis Report: OpenAI Stuck in Frontier Model Training Stagnation for Two and a Half Years

Brief:

• As per the SemiAnalysis report, OpenAI has not successfully completed any training deployments for new frontier large language models over the past two and a half years, with multiple attempts halting midway due to issues like convergence difficulties.
• The report states that the GPT-5 series, currently awaited by the public, is essentially an optimized variant of GPT-4o and has not achieved an architectural breakthrough.
• This stagnation is seen by the industry as a landmark signal that AI development has encountered a triple bottleneck of data, computational power, and model recipes, entering the "Post-Scaling Era."
• The report further discloses that OpenAI, by evaluating the potential use of Google TPUs, successfully pressured Nvidia to concede approximately 30% on GPU cluster pricing, thereby reducing costs.

Related Links:

• Pre-training "Stuck": SemiAnalysis Reveals OpenAI Hasn't Completed New Frontier LLM Training for Two and a Half Years
• Google TPU Launch Triggers AI Computing Price War, OpenAI Saves 30% on Costs
• Win11 Copilot Delivers "Full-Fledged" GPT-5.1 Directly, Deep Thinking Features Unlocked for Free!

ChatGPT测试版代码泄露，广告功能或将上线

简报：

• 开发者在ChatGPT的安卓测试版代码中发现了与广告相关的功能代码，包含“商业内容”和“搜索广告”等关键词。
• 这一发现表明，OpenAI可能正在测试在AI的回复中嵌入广告，以提升其商业盈利能力。
• 尽管OpenAI首席执行官山姆·奥特曼曾对在AI中加入广告表示担忧，但面对市场竞争，这可能成为其新的盈利途径。

相关链接：

• 曝光的代码显示​ChatGPT 即将加入广告功能

ChatGPT Beta Code Leak Hints at Upcoming Ad Features

Brief:

• Developers have discovered ad-related feature code in ChatGPT's Android beta version, containing keywords such as "commercial content" and "search ads." 🔍
• This discovery suggests that OpenAI might be testing embedding advertisements within the AI's responses to enhance its commercial profitability. 💰
• Although OpenAI CEO Sam Altman has previously expressed concerns about integrating ads into AI, this could become a new revenue stream amidst intense market competition. 🚀

Related Links:

• Leaked Code Reveals ChatGPT Is About to Get Ad Features

黄仁勋力推公司内部AI化，承诺员工岗位无忧

简报：

• 英伟达CEO黄仁勋在一次全体员工会议上，要求员工尽可能利用AI工具完成任务，目标是“所有能够通过AI完成的任务都能实现自动化”。
• 他对部分不鼓励员工使用AI的管理者表示不满，并鼓励工程师使用公司开发的AI辅助编码平台Cursor以提高效率。
• 针对AI可能导致岗位流失的担忧，黄仁勋回应称公司正在快速扩张，员工人数已增至36000人，另有约10000个岗位空缺，并向员工保证“你们会有工作可做”。

相关链接：

• 英伟达 CEO 黄仁勋动员员工尽可能利用 AI 工具完成各项任务

Jensen Huang Pushes Internal AI Adoption, Assures Employees of Job Security 🚀

Brief:

• NVIDIA CEO Jensen Huang, during an all-hands meeting, urged employees to leverage AI tools for tasks as much as possible, with the goal that "all tasks amenable to AI should be automated." 🤖
• He expressed dissatisfaction with some managers who discourage AI use by employees and encouraged engineers to utilize Cursor, the company's AI-assisted coding platform, to boost efficiency.
• Addressing concerns about potential job displacement due to AI, Huang responded that the company is rapidly expanding, with its workforce growing to 36,000, and approximately 10,000 job openings. He assured employees, "you will have work to do." 📈

Related Link:

• NVIDIA CEO Jensen Huang Mobilizes Employees to Maximize Use of AI Tools for Tasks

快手发布Kling Omni模型，主打2分钟音画同步与导演级控制

简报：

• 快手科技于本周开启“Kling Omni Launch Week”，将发布其最新的多模态视频基础模型Kling O1（Omni One）。
• 该模型可生成长达2分钟的连续视频，并集成了原生音频同步技术，实现声音与画面的帧级精确匹配。
• Kling Omni提供“导演级”的精准控制能力，支持文本、图像和视频的多输入参考，旨在将AI视频创作从“随机生成”转向“程序化导演”模式。
• Kling系列模型目前已服务全球4500万用户和2万家企业。

相关链接：

• 快手 Kling Omni 本周即将发布：实现“导演级”精准控制，可生成2分钟带原生音频的长视频

Kuaishou Launches Kling Omni Model, Featuring 2-Minute Audio-Visual Synchronization and Director-Level Control

Brief:

• Kuaishou Technology initiated "Kling Omni Launch Week" this week, announcing the release of its latest multimodal video foundation model, Kling O1 (Omni One). 🚀
• The model can generate continuous videos up to 2 minutes in length, integrating native audio synchronization technology for frame-level precise matching of sound and visuals. 🎬
• Kling Omni offers "director-level" precise control capabilities, supporting multiple input references including text, images, and videos, aiming to shift AI video creation from "random generation" to a "programmatic directing" mode. ✨
• The Kling series models currently serve 45 million users and 20,000 enterprises globally.

Related Links:

• Kuaishou Kling Omni to Be Released This Week: Achieving "Director-Level" Precise Control, Capable of Generating 2-Minute Long Videos with Native Audio