🔥 NEW: Ethereum launches "The Torch" NFT for its 10th anniversary – a commemorative token honoring the people and values that shaped Ethereum's first decade.The torch will be passed wallet-to-wallet over 10 days leading up to the anniversary celebration. https://t.co/0JkLvW71hl ...

Group 1 - OpenAI core scientist Jason Wei and Hyung Won Chung have left to join Meta, with Wei being the father of the thinking chain and Chung responsible for code models [1] - Meta has adopted an aggressive strategy in the AI field, investing $16 billion to recruit top talent, leveraging its own funds and decision-making autonomy to lead the competition [1] - Following its transformation into AI, Meta's stock price surged, reaching a new market capitalization high, with CEO Mark Zuckerberg transitioning from being mocked as a "metaverse dreamer" to a "strategic tech leader" [1] Group 2 - AI pioneers, including OpenAI, DeepMind, and Anthropic, have jointly called for in-depth research on monitoring thinking chains (CoT) to enhance AI safety [2] - Experts believe that CoT monitoring offers a unique opportunity for AI safety by observing the model's "thought process" to detect malicious intent, although its monitorability may decrease with different training methods [2] - The document proposes several research directions and recommendations for CoT monitoring, including assessing monitorability, publishing evaluation results, and incorporating monitorability into training decisions to prevent AI behavior from going out of control [2] Group 3 - Mistral AI has released its first open-source voice model, the Voxtral series, which includes 24B and 3B versions, licensed under Apache 2.0 [3] - Voxtral supports a 32k token context window, capable of processing 30 minutes of audio transcription or 40 minutes of semantic understanding, outperforming the open-source model Whisper in multiple tests [3] - The model supports eight major languages and inherits text understanding capabilities from Mistral Small 3.1, surpassing GPT-4o mini in some tests, but still lags behind top commercial models overall [3] Group 4 - MiniMax has launched an Agent full-stack development feature that allows users to build complete application systems with no-code, including backend hosting, payment integration, and scheduled tasks [4][5] - Users can create applications like concert seat selection systems, real-time financial dashboards, and e-commerce websites within 30 minutes, supporting real payment functions and data processing [5] - This feature employs a modular architecture, consisting of three core sub-Agents for research, development, and testing, and has released 12 updates in over a month, lowering the development barrier for enterprise applications [5] Group 5 - Kunlun Wanwei and Nanyang Technological University have introduced a new hierarchical multi-agent collaboration framework called AgentOrchestra, utilizing an "AI orchestra" collaboration model to tackle complex tasks [6] - The framework is coordinated by a top-level "conductor" Planning Agent, working alongside three types of specialized "musician" agents (Deep Researcher, Browser Use, Deep Analyzer) for collaborative tasks [6] - AgentOrchestra has performed excellently in authoritative evaluations such as SimpleQA and GAIA, achieving an 82.42% pass@1 score in the GAIA test, with complete open-source code and technical reports available [6] Group 6 - Google DeepMind has developed a software library named Concordia, creating an AI-hosted multi-AI character interaction environment similar to the AI virtual world in "Westworld" [7] - The system is designed based on a game engine's entity-component architecture, treating AI players and AI game masters (GMs) as configurable entities with different capabilities through pluggable components [7] - Concordia supports three main application scenarios: evaluative (testing AI capabilities), dramatic (creating interactive narratives), and simulation (building social science research environments), and has been open-sourced on GitHub [7] Group 7 - The ima platform offers note resources from top students at prestigious universities, including structured knowledge and thinking models across multiple subjects [8] - These notes not only compile knowledge but also include problem-solving strategies, key point breakdowns, and error analysis, such as high-scoring templates for Chinese and techniques for analyzing complex English sentences [8] - Users can directly ask "top student notes" on the ima platform for study methods, mindset adjustment advice, and can upload their own notes to build a personal knowledge base [8] Group 8 - NVIDIA CEO Jensen Huang praised the Chinese supply chain as a "miracle" during his first speech in Chinese at the China Supply Chain Expo, naming 11 Chinese companies [10] - He emphasized that Chinese open-source models are catalysts for global AI progress, providing opportunities for countries to join the AI revolution, and predicted that the next wave of AI will focus on understanding the physical world and robotic systems [10] - NVIDIA made its debut at the supply chain expo, showcasing humanoid robot products from four Chinese companies, including Galaxy General and Beijing Humanoid Robot Innovation Center, along with DIGITS mini supercomputers [10] Group 9 - The "verifier's law" states that the difficulty of AI solving tasks is proportional to the verifiability of the task rather than the complexity of the task itself [11] - Verifiability includes five key attributes: objective truth, rapid verification, scalable verification, low noise, and continuous rewards [11] - Any problem meeting these five attributes will be solved by AI in the future, creating an "intelligent serrated frontier" where AI will demonstrate higher intelligence on verifiable tasks [11] Group 10 - OpenAI's third podcast discusses the evolution of ChatGPT from an API "playground" to a flagship product and its profound impact on work and the economy [12] - COO Mira Murati and Chief Economist Dan Altman believe AI will significantly enhance productivity, especially in software engineering, scientific research, and small businesses, predicting that AI agents will become key partners in handling complex tasks [12] - They emphasize the need to focus on soft skills such as emotional intelligence, critical thinking, and adaptability in the AI era, advocating for educational reforms to cultivate collaboration skills with AI, and noting that AI is expected to create significant value in emerging markets and agriculture [12]

西风 发自 凹非寺 量子位 | 公众号 QbitAI 无需CUDA代码，给H100加速33%-50% ！ Flash Attention、Mamba作者之一 Tr i Da o 的新作火了。 他和两位普林斯顿CS博士生提出了 一个名叫 QuACK 的新SOL内存绑定内核库 ，借助CuTe-DSL，完全用Python写，一点CUDA C++代码 都没用到。 在带宽3TB/s的H100上，它的速度比像PyTorch的torch.compile、Liger这类已经过深度优化的库还要快33%-50%。 Tri Dao表示，让内存密集型的内核达到"光速"并非什么神秘技巧，只需把几个细节处理到位就行。 我很喜欢Phil Tillet对不同工具在生产力和性能方面各有取舍的观点，比如torch compile、triton、CUDA、PTX。 但CuTe-DSL以及类似的基于Python的DSL或许能改变这一局面，虽然目前还处于早期阶段。而且，说不定很快我们就能让大语言模型 来生成这些内核了！ 新作一经发出，吸引不少大佬关注。 英伟达CUTLASS团队资深架构师Vijay 转发，自夸他们团队做的CuTe-DSL把各种细节都打 ...

奕然 发自 凹非寺 量子位 | 公众号 QbitAI 能让非NVIDIA芯片跑CUDA的开源项目ZLUDA，起死回生了。 最新版增加了对大模型工作负载的支持，一举登上GitHub热榜！ 开发者@vosen本名Andrzej Janik，曾经在Intel工作。 该项目一度因AMD停止资助濒临破产，最终被一家神秘机构出手相救。 现在，创始人vosen带来好消息，表示ZLUDA团队新添一员猛将并稳定进行项目恢复中。 在2020年，Andrzej Janik想要尝试一下技术突破，让CUDA程序在非NVIDIA平台运行，一尝试，便有了可行性。 之后，ZLUDA被Intel接手，作为一个内部试验项目发展。 Intel分配资源给了ZLUDA，目的很明显了，让其 在Intel GPU上跑CUDA程序，作为Intel oneAPI生态的一种补充方式 。 无疑，这触碰到了NVIDIA的商业生态链。 没过多久，这个项目就被终止了。 2022年，ZLUDA得到了AMD的支持而重启，并支持AMD硬件。 好景不长，这次也仅仅维持2年，2024年2月宣布终止。 过往发展：起起伏伏又起起 一个月后，英伟达就发布CUDA 11.6版本，并明确 ...

Core Insights - Worthington Enterprises Inc. reported strong fourth quarter results for fiscal 2025, showing year-over-year and sequential growth in adjusted EBITDA, adjusted EPS, and free cash flow, driven by effective cost management and execution in its Consumer and Building Products segments [3][4][6]. Financial Performance - Net sales for Q4 2025 were $317.9 million, a slight decrease of 0.3% compared to Q4 2024, primarily due to the deconsolidation of the Sustainable Energy Solutions segment [5][6][8]. - The operating loss improved to $30.4 million from $56.1 million in the prior year, with adjusted operating income rising to $21.8 million, an increase of $16.0 million [5][9]. - Net earnings from continuing operations increased by 111% to $3.6 million, with adjusted EBITDA growing 35% to $85.1 million [6][9]. - Earnings per share from continuing operations improved from a loss of $(0.64) to a profit of $0.08, while adjusted EPS rose from $0.74 to $1.06 [6][28]. Cash Flow and Capital Management - Operating cash flow increased by 38% to $62.4 million, and free cash flow rose by 46% to $49.3 million [6][12]. - The company repurchased 200,000 shares for $9.8 million and declared a quarterly dividend of $0.19 per share, a 12% increase from the previous quarter [6][12]. Segment Performance - Consumer Products segment generated net sales of $125.6 million, remaining flat year-over-year, while adjusted EBITDA increased by $3.7 million to $20.8 million [14]. - Building Products segment saw net sales rise by 25.2% to $192.3 million, with adjusted EBITDA increasing by $19.6 million to $71.3 million, driven by higher volumes and contributions from the Ragasco acquisition [15][18]. Strategic Developments - The acquisition of Elgen Manufacturing for approximately $93 million was completed on June 19, 2025, aligning with the company's growth strategy in niche markets [6][16]. - The company expressed confidence in its ability to drive sustainable growth and long-term value heading into fiscal 2026 [16].

点击下方 卡片 ，关注" 自动驾驶之心 "公众号 戳我-> 领取 自动驾驶近15个 方向 学习 路线 从事大模型方向的小伙伴有福利了！斯坦福大学 2025 年春季的 CS336 课程「从头开始创造语言模型（Language Models from Scratch）」相关课程和材料现已在网上全面发布！ 该课程教职工团队，阵容十分豪华~ 课程视频：https://www.youtube.com/watch? v=SQ3fZ1sAqXI&list=PLoROMvodv4rOY23Y0BoGoBGgQ1zmU_MT_ 课程主页：https://stanford-cs336.github.io/spring2025/ 讲师Tatsunori Hashimoto：现为斯坦福大学计算机科学系助理教授。其为斯坦福大学 John C. Duchi 和 Percy Liang 的博士后，研究机器学习模型平均性能和最差性能之间的权衡。此前在麻省理工学院攻读研究生，导师是 Tommi Jaakkola 和 David Gifford。本科就读于哈佛大学学习统计学和数学，导师是 Edoardo Airoldi。并且该讲 师的研究成果已 ...

Core Viewpoint - Worthington Enterprises has acquired Elgen Manufacturing for approximately $93 million, enhancing its position in the HVAC market and aligning with its strategy to build leadership in niche markets [1][4]. Company Overview - Worthington Enterprises is a designer and manufacturer of brands that improve everyday life, operating primarily in two segments: Building Products and Consumer Products [5][6]. - The Building Products segment includes solutions for heating, cooling, construction, and water applications, while the Consumer Products segment covers tools and outdoor living [5]. Acquisition Details - Elgen Manufacturing, based in Closter, New Jersey, specializes in HVAC parts and components, generating net sales of $114.9 million and EBITDA of $13.3 million for the trailing 12 months ended April 30, 2025 [4]. - The acquisition is expected to create synergies and growth opportunities by leveraging Worthington's manufacturing expertise and distribution model [2][3]. Strategic Fit - The acquisition aligns with Worthington's strategy to acquire businesses with strong market positions, as Elgen's manufacturing processes and sales strategies complement those of Worthington [2][3]. - Elgen's products are used in commercial buildings, and its sales strategy focuses on direct sales to contractors and partnerships with distributors, enhancing customer service and lead times [2]. Leadership and Integration - Elgen's leadership team, including CEO David Young, will remain with the company, ensuring continuity and commitment to customer service and innovation [3].

插播 ： 6月10日的"2025年全球用户侧储能产业价值峰会暨应用示范展"圆满收官，演讲嘉宾PPT，点击 "阅读原文 " 了解详情 11日，行家说储能报道了"储能定义权之争升级！TOP30集结上海"（ 点这里 ），今日，行家说储能将继续关注SNEC展。 与往年相比，此次展会由光伏展变成名副其实的储能展，储能企业数量、储能产品含量、热度都有极大提升。而且所发布的产品除了围绕第三代电 芯的定义权外，还因应今年136号文落地和市场价值化变革，推出136政策响应方案、光储电站经济性评估以及AI+工商业储能系统等。 在产品迭代加速的同时，多家企业在展会期间密集签署重磅合作协议以及GWh级采购订单，如瑞浦兰钧总签约订单量超20GWh，国轩高科新品拿 下3GWh订单，蜂巢能源则 一举斩获2.1GWh储能订单…… | 参展企业 | 主要储能展品 | | --- | --- | | 采日能源 | Serlattice G3 10MWh智储系统等 | | 中车株洲所 | 构网型储能系统、中车"云枢"储能变流器等 | | 华为数字能源 | 构网型光储解决方案 FusionSolar9.0 | | 比亚迪储能 | 全场景储能新品 | ...

1. 迁得快：让三方框架模型 "零成本"迁移，避免重复造轮子，同时模型精度完全对齐。 2 . 部署快：训转推全流程自动化，让大模型部署像执行一行命令一般敏捷高效 。 Figure 1 MindSpore生态快速迁移解决方案的技术架构 接下来，我们将揭开昇思 MindSpore的破局之道。 一、支持训练Day0迁移，构建跨框架的"无感智能翻译"能力 当大模型架构日新月异，开发者最怕被生态绑定。昇思 MindSpore通过三重兼容术打通主流 技术栈 ， 支持 主流 加速库 模型0代码迁移 ，通过精度自动对比工具 实现跨框架、跨版本、 跨策略快速调优， 精度对齐原模型， 实现 在分布式并行策略保持不变的情况下，训练性能提 升5 %+ 。 在训练生态方面，通过 MindSpeed /Megatron桥接层实现 PyTorch 模型零代码迁移，训练 脚本可直接运行； 通过动态图能力重构，昇思让 PyTorch 开发者获得"原生体验"，同时借力 MSAdapter 工具自动转换9 5 %以上接口， 主流模型如 DeepSeek 、 Pan g u 等 迁移损耗 逼近于零。 大模型发展日新月异，新的大模型层出不穷，参数规模 ...

作者 | 王启隆 出品丨AI 科技大本营（ID：rgznai100） 席卷全球的 AI 淘金热中，一个词正被悄悄地掏空——那就是 "开放" 。 近日，PyTorch 基金会执行董事、Linux 基金会 AI 总经理 Matt White 在北京智源大会 揭示了一个 充满张力的现实：一方面，开源吞噬世界，AI 的开源更形成了一个自我加速的"良性循环"；但另一 方面，一场围绕"开放"定义权的无声战争已经打响。 这是我们这个时代的十字路口：是任由"开放"沦为一个漂亮的营销词汇，还是为它注入坚实的灵魂？ 在演讲中，Matt White 带来了两件精心铸造的"武器"：一张名为 "模型开放框架"（MOF） 的地 图，用清晰的等级标准终结含糊，让真正的开放者得以彰显；以及一本名为 "OpenMDW 许可证" 的护照 ，专为 AI 模型打造，给予使用者最大限度的自由。 他的演讲，与其说是一次技术分享，不如说是一份宣言，一份行动指南。它为我们接下来这场更深入 的对话，精准地校对了焦距。 在演讲结束后，我们与 Matt White 坐下来，继续探寻这场"为开放而战"的深层动机与未来图景。 《新程序员》 ：嗨 Matt，我们其实和 ...