![[87b981ca-60e4-4693-8376-37cbeeaf95c4.mp3]]

快手12项计算机视觉研究成果获CVPR2025国际认可

简报：

• 快手科技共有12篇论文被2025年计算机视觉顶会CVPR接收，覆盖多模态视频理解、3D场景重建、高效模型压缩、生成式AI等前沿方向。
• 快手在视频内容理解领域论文《Hierarchical Attention for Long-form Video Understanding》获得最佳论文提名，实现国内企业在顶会评选重点突破。
• 相关研究将陆续应用于快手短视频推荐、特效生成、内容审核等业务，快手近三年AI研发经费年增长率超50%。
• CVPR为计算机视觉最高级别学术会议，今年全球投稿过万，接收率约25%，快手论文入选数在国内互联网公司中位居前列。

相关链接：

• 36kr新闻快报
• AIbase新闻1

Kuaishou's 12 Computer Vision Research Achievements Gain International Recognition at CVPR 2025 🚀

Brief:

• Kuaishou Technology has a total of 12 papers accepted by CVPR 2025, the top-tier conference in computer vision. These papers cover cutting-edge directions such as multi-modal video understanding, 3D scene reconstruction, efficient model compression, and generative AI.
• Kuaishou's paper in the field of video content understanding, titled "Hierarchical Attention for Long-form Video Understanding," received a Best Paper Nomination. This marks a significant breakthrough for a domestic enterprise in the selection process of a top-tier conference. ✨
• The related research will be progressively applied to Kuaishou's businesses, including short video recommendation, special effects generation, and content moderation. Kuaishou's AI R&D expenditure has seen an annual growth rate exceeding 50% in the past three years. 📈
• CVPR is the highest-level academic conference in computer vision. This year, global submissions exceeded 10,000, with an acceptance rate of approximately 25%. Kuaishou's number of accepted papers ranks among the top domestic internet companies.

Related Links:

• 36kr News Flash
• AIbase News 1

macOS Spotlight升级为智能操作中心，开发者AI工具大幅增强

简报：

• 在2025年WWDC大会上，苹果发布了macOS Tahoe，Spotlight升级为智能命令中心，可运行快捷方式、直接执行数百种操作、个性化推荐，并通过App Intents API加强第三方应用集成。
• 开发者获得Foundation Models框架，可将本地AI模型无缝集成至应用，保障隐私且免费使用，仅需三行代码接入Apple Intelligence功能。
• Xcode26推出AI助手，集成ChatGPT与本地AI模型，覆盖代码生成、测试、文档等开发全流程，提升开发效率。
• 这些升级体现苹果在AI时代提升Mac用户体验与开发者生态的战略布局。

相关链接：

• 苹果发布macOS Tahoe：Spotlight升级为智能命令中心
• WWDC25：苹果开放 Foundation Models，加速开发者AI集成
• Xcode 26 集成 AI 助手提升开发体验
• SETN：苹果AI功能、Spotlight大升级

macOS Spotlight Upgraded to Smart Command Center, Developer AI Tools Significantly Enhanced

Briefing:

• At WWDC 2025, Apple unveiled macOS Tahoe, featuring a transformed Spotlight into a smart command center capable of running shortcuts, directly executing hundreds of actions, offering personalized recommendations, and boasting enhanced third-party app integration via the App Intents API. 💡
• Developers gained access to the Foundation Models framework, enabling seamless integration of on-device AI models into apps, ensuring privacy and free usage. Apple Intelligence features can now be accessed with just three lines of code. 🚀
• Xcode 26 introduced an AI assistant, integrating ChatGPT and on-device AI models to cover the entire development workflow, including code generation, testing, and documentation, significantly boosting development efficiency.
• These upgrades underscore Apple's strategic commitment to enhancing the Mac user experience and developer ecosystem in the AI era. 🍎

Related Links:

• Apple releases macOS Tahoe: Spotlight upgraded to smart command center
• WWDC25: Apple opens Foundation Models, accelerating developer AI integration
• Xcode 26 integrates AI assistant to enhance development experience
• SETN: Apple's AI features, Spotlight greatly upgraded

OpenAI开源模型延期至夏末发布，奥尔特曼透露研发取得“惊人成果”

简报：

• OpenAI CEO 山姆・奥尔特曼在 X 平台宣布，原定于今年初夏发布的开源模型将推迟至夏末，理由是公司希望对开源权重模型进行进一步打磨，以呈现更完善的产品。
• 奥尔特曼表示，OpenAI 的研究团队在模型开发过程中取得了一些“意想不到且相当惊人”的成果，最终产品将“非常值得等待”。
• 新的开源模型计划具备与 OpenAI O 系列模型相似的推理能力，并力争超越现有其他开源推理模型。
• OpenAI 目前尚未公布推迟后具体的发布时间表。

相关链接：

• OpenAI 开源模型发布推迟至夏末，CEO 奥尔特曼称“非常值得等待”

OpenAI Open-Source Model Release Delayed Until Late Summer, Altman Reveals "Stunning Results" in Development 🚀

Briefing:

• OpenAI CEO Sam Altman announced on X that the open-source model originally slated for early summer this year will be postponed until late summer, citing the company's desire to further refine the open-source weight model to deliver a more polished product.
• Altman stated that OpenAI's research team has achieved "unexpected and quite stunning" results during the model's development, and that the final product will be "well worth the wait." ✨
• The new open-source model is planned to possess reasoning capabilities similar to OpenAI's O-series models and aims to surpass other existing open-source reasoning models.
• OpenAI has not yet announced a specific release timeline following the postponement.

Related Links:

• OpenAI Open-Source Model Release Postponed Until Late Summer, CEO Altman Says "Well Worth the Wait" 🗓️

英伟达黄仁勋在伦敦科技周聚焦AI变革与未来基础设施

简报：

• 英伟达CEO黄仁勋近期在“伦敦科技周”开幕式上发表主题演讲，并与英国首相基尔·斯塔默会面，共同探讨AI未来及其对产业的深刻影响。
• 黄仁勋强调，过去十年AI硬件性能提升了100万倍，产业变革速度惊人；英伟达将协助英国推进AI基础设施与人才培养，并设立专注机器人、环境建模等领域的AI研究中心。
• 演讲中还涉及推动AI纳入国家经济规划、建设高性能计算平台以及在英国开展全国性开发者教育计划。

相关链接：

• 36氪新闻快讯
• IT之家报道
• AIBase新闻

NVIDIA CEO Jensen Huang at London Tech Week: Focusing on AI Transformation and Future Infrastructure

Brief:

• NVIDIA CEO Jensen Huang recently delivered a keynote speech at the opening ceremony of "London Tech Week" and met with UK Prime Minister Keir Starmer to discuss the future of AI and its profound impact on industries. 🤝
• Huang emphasized that AI hardware performance has increased a millionfold over the past decade, highlighting the astonishing pace of industry transformation. NVIDIA will assist the UK in advancing AI infrastructure and talent development, and establish AI research centers focused on areas such as robotics and environmental modeling. 🚀
• The speech also covered integrating AI into national economic planning, building high-performance computing platforms, and launching a nationwide developer education program in the UK. 💡

Related Links:

• 36Kr News Flash
• ITHome Report
• AIBase News

小红书发布1420亿参数开源大模型dots.llm1，强调真实数据训练

简报：

• 小红书hi lab团队正式推出其首个开源大模型dots.llm1，采用混合专家（MoE）架构，参数总量高达1420亿，激活参数为140亿。
• dots.llm1经过11.2TB高质量真实场景文本数据训练，性能可与阿里Qwen2.5-72B媲美，未使用任何合成数据。
• 小红书此次开源旨在推动AI技术创新，并鼓励开发者和社区参与，共同探索模型在内容创作、智能客服、对话系统等领域的应用。

相关链接：

• 小红书开源大模型dots.llm1发布
• 混合专家，1420亿参数！小红书开源超大语言模型dots.llm1

Xiaohongshu Releases 142 Billion Parameter Open-Source Large Model dots.llm1, Emphasizing Real-World Data Training

Brief:

• Xiaohongshu's hi lab team has officially launched its first open-source large language model, dots.llm1. It adopts a Mixture-of-Experts (MoE) architecture, boasting a total of 142 billion parameters with 14 billion activated parameters. 🚀
• Trained on 11.2TB of high-quality real-world text data, dots.llm1's performance is comparable to Alibaba's Qwen2.5-72B, notably without using any synthetic data. ✨
• Xiaohongshu's aim with this open-sourcing initiative is to foster AI technology innovation and encourage developers and the community to participate in exploring the model's applications in areas such as content creation, intelligent customer service, and conversational systems. 💡

Related Links:

• Xiaohongshu Open-Source Large Model dots.llm1 Released
• Mixture of Experts, 142 Billion Parameters! Xiaohongshu Open-Sources Super Large Language Model dots.llm1