![[04dfa122-e2bb-47ae-8628-bd7a452a8eaa.mp3]]

英伟达Blackwell GPU创AI推理新纪录，Llama 4模型达每秒1000 token

简报：

• 英伟达采用单节点（8颗Blackwell GPU）的DGX B200服务器，在Llama 4 Maverick模型上实现每秒单用户生成1000个token（TPS/user）的世界纪录
• 单台GB200 NVL72服务器（配备72颗Blackwell GPU）整体吞吐量达到72,000 TPS
• 该记录由AI基准测试服务Artificial Analysis独立测量确认
• 性能提升得益于TensorRT-LLM优化框架、FP8数据格式应用、CUDA内核优化技术等一系列技术优化

相关链接：

• 英伟达再破世界纪录，每秒1000 token！刚刚，全球最快Llama 4诞生

• 破世界纪录，每秒1000 token！刚刚，全球最快Llama 4 推理速度诞生

NVIDIA Blackwell GPU Sets New AI Inference Record, Llama 4 Model Reaches 1000 Tokens Per Second

Newsletter:

• NVIDIA's DGX B200 server, utilizing a single node with 8 Blackwell GPUs, achieves a world record of 1000 tokens per second (TPS/user) for single-user generation on the Llama 4 Maverick model. 🚀
• A single GB200 NVL72 server (equipped with 72 Blackwell GPUs) delivers an overall throughput of 72,000 TPS. 💻
• This record has been independently measured and confirmed by the AI benchmarking service Artificial Analysis.
• The performance boost is attributed to a series of technological optimizations, including the TensorRT-LLM optimization framework, FP8 data format application, and CUDA kernel optimization techniques. 🌟

Related Links:

• NVIDIA Breaks World Record Again, 1000 Tokens Per Second! Just Now, the Fastest Llama 4 Emerges
• Breaking World Record, 1000 Tokens Per Second! Just Now, the Fastest Llama 4 Inference Speed is Born

Anthropic发布Claude 4系列大模型，连续编程能力达7小时

简报：

• Anthropic正式推出Claude 4系列大模型，包括旗舰型号Claude Opus 4和升级版Claude Sonnet 4
• Claude Opus 4在SWE-bench基准测试中达到72.5%准确率，能处理复杂长时间编程任务
• 在Rakuten测试中，Claude Opus 4实现了连续7小时的自主编码，创下AI编程时长新纪录
• 新模型具备工具辅助的延伸思考能力，支持GitHub Actions后台任务执行和IDE集成
• Claude Sonnet 4作为免费用户默认模型，在SWE-bench测试中达到72.7%准确率

相关链接：

• 刚刚！首个下一代大模型Claude4问世，连续编程7小时，智商震惊人类
• Claude 4发布：替代人类程序员所需的条件，现在它都有了
• Anthropic深度突破：Claude4系列实现连续7小时自动编程引领AI技术革新

Anthropic Releases Claude 4 Series Large Models, Achieves 7-Hour Continuous Coding Capability

Newsletter:

• Anthropic officially launches the Claude 4 series of large models, including the flagship model Claude Opus 4 and the upgraded Claude Sonnet 4. 🚀
• Claude Opus 4 achieves a 72.5% accuracy rate on the SWE-bench benchmark, capable of handling complex, long-duration programming tasks.
• In the Rakuten test, Claude Opus 4 accomplished autonomous coding for 7 consecutive hours, setting a new record for AI coding duration. ⏰
• The new models feature tool-assisted extended thinking capabilities, supporting GitHub Actions for background task execution and IDE integration.
• Claude Sonnet 4, as the default model for free users, achieves a 72.7% accuracy rate on the SWE-bench test. 💻

Related Links:

• Breaking News! The First Next-Generation Large Model Claude 4 Debuts, Coding for 7 Hours Straight, Intelligence Stuns Humanity
• Claude 4 Released: It Now Has Everything Needed to Replace Human Programmers
• Anthropic’s Major Breakthrough: Claude 4 Series Achieves 7-Hour Continuous Auto-Coding, Leading AI Technology Innovation

字节跳动与清华大学联合开源时序大模型ChatTS，入选VLDB 2025

简报：

• 字节跳动ByteBrain团队与清华大学合作开发的多模态时序大模型ChatTS正式开源，论文入选数据库顶级会议VLDB 2025
• ChatTS原生支持多变量时序问答与推理，解决了传统时序分析方法通用性和可解释性不足的问题
• 研究团队采用"纯合成驱动"方法构建端到端数据生成与训练框架，实现时序数据与自然语言的精确对应
• 该模型由清华大学博士生谢哲（一作）和字节跳动研究员李则言、何晓等共同完成，通讯作者为张铁赢（字节）和裴丹（清华）
• 项目已在GitHub开源，提供14B参数模型（HuggingFace）及完整数据集

相关链接：

• ChatTS论文（arXiv）
• GitHub开源代码
• HuggingFace模型

ByteDance and Tsinghua University Jointly Open-Source Time-Series Model ChatTS, Selected for VLDB 2025

Newsletter:

• ByteDance's ByteBrain team, in collaboration with Tsinghua University, has officially open-sourced the multimodal time-series large model ChatTS, with its paper accepted at the top database conference VLDB 2025. 📊✨
• ChatTS natively supports multivariate time-series Q&A and reasoning, addressing the lack of generality and interpretability in traditional time-series analysis methods.
• The research team adopted a "purely synthetic-driven" approach to build an end-to-end data generation and training framework, achieving precise alignment between time-series data and natural language.
• The model was co-developed by Tsinghua PhD student Xie Zhe (first author), ByteDance researchers Li Zeyan and He Xiao, among others, with corresponding authors Zhang Tieying (ByteDance) and Pei Dan (Tsinghua).
• The project is now open-sourced on GitHub, offering a 14B-parameter model (via HuggingFace) and a complete dataset. 🚀

Related Links:

• ChatTS Paper (arXiv)
• GitHub Open-Source Code
• HuggingFace Model

谷歌发布AI电影制作工具Flow，整合Veo 3、Imagen和Gemini技术

简报：

• 谷歌在I/O 2025开发者大会上发布影视级AI制作工具Flow，由Veo 3、Imagen和Gemini提供技术支持
• Flow具备文本指令遵循、多动一致性、色彩对比度优化等功能，物理模拟表现尤其出色
• 提供专业影视制作功能：相机控制（运动、角度、视角）、场景构建器（无缝编辑扩展镜头）、资产管理
• Flow TV平台展示Veo生成内容，用户可学习优秀片段的提示技巧
• 已向美国地区Pro和Ultra订阅用户开放，Pro用户每月100次生成，Ultra用户不限次数
• 知名电影制作人Dave Clark已使用Flow制作短电影《Freelancers》

相关链接：

• 谷歌重磅发布Flow，人人都能制作好莱坞级电影
• Google推出Flow 一款面向Gemini高级订阅用户的AI电影制作工具
• 独家对话谷歌Flow合作艺术家：3周完成9分钟短片

Google Unveils AI Filmmaking Tool Flow, Integrating Veo 3, Imagen, and Gemini Technologies

Brief:

• Google launched the cinematic-grade AI production tool Flow at the I/O 2025 Developer Conference, powered by Veo 3, Imagen, and Gemini technologies. 🎥
• Flow offers features like text instruction compliance, multi-motion consistency, color contrast optimization, and exceptional physical simulation performance.
• Provides professional filmmaking capabilities: camera control (motion, angle, perspective), scene builder (seamless editing for extended shots), and asset management.
• Flow TV platform showcases Veo-generated content, allowing users to learn prompt techniques from outstanding clips. 🌟
• Now available to Pro and Ultra subscribers in the US, with Pro users getting 100 generations per month and Ultra users enjoying unlimited access.
• Renowned filmmaker Dave Clark has already used Flow to create the short film Freelancers. 🎬

Related Links:

• Google's Major Release of Flow: Everyone Can Create Hollywood-Level Movies
• Google Launches Flow: An AI Filmmaking Tool for Gemini Premium Subscribers
• Exclusive Interview with Google Flow Collaborating Artist: 9-Minute Short Film Completed in 3 Weeks

OpenAI以64亿美元收购前苹果设计师Jony Ive的AI硬件公司io

简报：

• OpenAI宣布以64亿美元全资收购前苹果首席设计师Jony Ive创立的AI硬件初创公司io，这是OpenAI迄今为止最大规模的收购
• io由Jony Ive与OpenAI CEO Sam Altman于2024年共同创立，团队包括55名前苹果设计师和工程师，将整体并入OpenAI
• Jony Ive将担任OpenAI创意总监，负责产品设计工作，但其设计公司LoveFrom将保持独立运营
• 双方计划在2026年推出首款AI硬件产品，奥特曼称已试用原型机并评价为"世界上最酷的科技产品"
• 收购完成后OpenAI将支付50亿美元现金（已持有io 23%股份），几周前该公司刚以30亿美元收购AI编程工具Windsurf

相关链接：

• OpenAI以65亿美元收购前苹果首席设计师创业公司
• 苹果前设计师创业，公司卖出460亿元！买家是OpenAI
• 突发！461亿、OpenAI收购iPhone设计师Jony Ive的AI设备初创公司io

OpenAI Acquires Former Apple Designer Jony Ive’s AI Hardware Company io for $6.4 Billion

Newsletter:

• OpenAI has announced the full acquisition of io, an AI hardware startup founded by former Apple Chief Designer Jony Ive, for $6.4 billion, marking OpenAI's largest acquisition to date. 💰
• io was co-founded by Jony Ive and OpenAI CEO Sam Altman in 2024, with a team of 55 former Apple designers and engineers, who will now fully integrate into OpenAI.
• Jony Ive will take on the role of OpenAI's Creative Director, overseeing product design, though his design firm LoveFrom will remain independently operated.
• Both parties aim to launch their first AI hardware product in 2026, with Altman praising the prototype as "the coolest tech product in the world." 🚀
• Following the acquisition, OpenAI will pay $5 billion in cash (having already held a 23$3 billion. 💻

Related Links:

• OpenAI Acquires Former Apple Chief Designer’s Startup for $6.5 Billion
• Former Apple Designer’s Startup Sold for 46 Billion Yuan! Buyer is OpenAI
• Breaking! 46.1 Billion Yuan: OpenAI Acquires iPhone Designer Jony Ive’s AI Device Startup io

字节跳动开源Dolphin文档解析模型，性能超越GPT-4.1和Mistral-OCR

简报：

• 字节跳动开源轻量级文档解析模型Dolphin，采用创新的"先解析结构后解析内容"两阶段范式
• 测试显示Dolphin在文档解析准确率上超越GPT-4.1、Claude3.5-Sonnet、Gemini2.5-pro等通用多模态大模型
• 性能同样超过号称最强OCR大模型的Mistral-OCR，解析效率提升近2倍
• 论文已被ACL 2025收录，项目已在GitHub和Hugging Face开源
• 该模型解决了传统方案错误累积和通用大模型丢失版面结构信息的问题

相关链接：

• 字节开源高精度文档解析大模型Dolphin：轻量高效，性能超GPT4.1、Mistral-OCR！
• Dolphin GitHub项目
• Dolphin论文

ByteDance Open-Sources Dolphin Document Parsing Model, Outperforming GPT-4.1 and Mistral-OCR

Newsletter:

• ByteDance has open-sourced Dolphin, a lightweight document parsing model that adopts an innovative two-stage paradigm of "structure parsing first, then content parsing." 🐬
• Tests show that Dolphin surpasses general multimodal large models like GPT-4.1, Claude3.5-Sonnet, and Gemini2.5-pro in document parsing accuracy.
• Its performance also exceeds Mistral-OCR, claimed to be the strongest OCR model, with nearly 2x higher parsing efficiency. 🚀
• The research paper has been accepted by ACL 2025, and the project is now open-sourced on GitHub and Hugging Face.
• This model addresses issues of error accumulation in traditional approaches and the loss of layout structure information in general large models. 📑

Related Links:

• ByteDance Open-Sources High-Precision Document Parsing Model Dolphin: Lightweight, Efficient, Outperforms GPT4.1 and Mistral-OCR!
• Dolphin GitHub Project
• Dolphin Paper

AI生成代码中20%依赖"幽灵包"，苹果微软等公司受影响

简报：

• 最新研究分析57.6万个代码样本发现，超过20%的AI生成代码依赖不存在的软件包（"幽灵包"）
• 苹果、微软等大型科技公司曾因此类问题中招，可能面临供应链攻击风险
• Meta和微软高管预测未来AI将主导代码生成，Meta预计12-18个月内大部分代码将由AI生成
• 微软首席技术官Kevin Scott预测5年内95%代码将由AI生成，人类手动编写代码将几乎消失
• 研究人员建议在使用AI推荐代码前仔细检查软件包是否存在，避免安全漏洞

相关链接：

• 57.6万代码撕碎AI编程神话，20%"幽灵包"暗藏漏洞！苹果、微软已中招
• 57.6万代码撕碎AI编程神话，20%「幽灵包」暗藏漏洞！苹果、微软已中招

20% of AI-Generated Code Relies on "Ghost Packages," Affecting Companies Like Apple and Microsoft

Newsletter:

• A recent study analyzing 576,000 code samples found that over 20% of AI-generated code depends on non-existent software packages ("ghost packages") 👻
• Major tech giants like Apple and Microsoft have been impacted by such issues, potentially facing supply chain attack risks ⚠️
• Executives from Meta and Microsoft predict that AI will dominate code generation in the future, with Meta estimating that most code will be AI-generated within 12-18 months
• Microsoft CTO Kevin Scott forecasts that within 5 years, 95% of code will be AI-generated, and human manual coding will nearly disappear 💻
• Researchers advise thoroughly checking the existence of software packages before using AI-recommended code to avoid security vulnerabilities

Related Links:

• 576,000 Code Samples Shatter AI Programming Myths, 20% "Ghost Packages" Hide Vulnerabilities! Apple and Microsoft Already Affected
• 576,000 Code Samples Shatter AI Programming Myths, 20% "Ghost Packages" Hide Vulnerabilities! Apple and Microsoft Already Affected