Core Viewpoint - The article discusses the security risks associated with Computer-Use Agents (CUAs) and introduces RiOSWorld, a benchmark for evaluating these risks in real-world scenarios [1][8][29]. Group 1: Introduction to Computer-Use Agents - CUAs have advanced capabilities, allowing them to perform tasks such as coding, handling emails, and creating presentations with simple commands [1]. - However, there are significant security concerns regarding the delegation of computer control to these intelligent assistants, likening it to sharing sensitive information with strangers [1]. Group 2: RiOSWorld Benchmark - RiOSWorld is presented as a comprehensive testing benchmark designed to assess the security risks faced by CUAs in everyday computer usage [8]. - The benchmark includes 492 risk test cases that cover a wide range of scenarios, including web, social media, operating systems, multimedia, file operations, code IDE/GitHub, email, and Office applications [10][15]. Group 3: Risk Categories and Examples - The risks are categorized into two main types: environmental risks (254 cases) and user risks (238 cases) [11][13]. - Environmental risks include phishing websites, phishing emails, and pop-up ads, while user risks involve actions like executing high-risk commands or sharing sensitive information [19][20]. Group 4: Evaluation Methodology - RiOSWorld evaluates CUAs based on two dimensions: the intention to execute risky behavior and the successful completion of that behavior [16]. - The results indicate that most agents exhibit weak risk awareness, with an average intention to perform unsafe actions at 84.93% and a completion rate of 59.64% [25][28]. Group 5: Findings and Implications - The findings reveal that CUAs are prone to high failure rates in risky scenarios, with over 89% in phishing websites and 80% in web operations [26]. - The article emphasizes the need for safety measures in AI development, stating that without security, even powerful AI systems are unreliable [29].

Core Viewpoint - The article discusses the rapid evolution of AI technology, emphasizing the collaborative ecosystem that supports developers in accessing and utilizing AI models effectively. The ModelScope community is highlighted as a key platform facilitating this collaboration and innovation [1][2]. Group 1: ModelScope Community Development - ModelScope community has grown significantly since its establishment in November 2022, now hosting over 500 contributing organizations and more than 70,000 open-source models, representing a growth of over 200 times [1]. - User numbers have surged from 1 million in April 2023 to 16 million, marking an approximate 16-fold increase [1]. - The community provides comprehensive services for developers, including model experience, download, tuning, training, inference, and deployment across various AI fields [2]. Group 2: AI Trends and Innovations - The first ModelScope Developer Conference featured a main forum and six thematic forums covering 65 topics related to cutting-edge models and tools, with participation from renowned AI open-source teams [5][6]. - The rise of multi-modal AI allows for simultaneous understanding and generation of text, images, audio, and video, enhancing interaction with the world [11]. - The emergence of world models enables AI to understand physical world dynamics, facilitating applications in robotics and autonomous systems [13]. Group 3: Open Source and Ecosystem - By 2025, China is positioned as a critical driver of the global AI open-source movement, with companies like Alibaba and DeepSeek releasing competitive open-source models [8][10]. - The integration of open-source initiatives with national infrastructure, such as computing networks, is fostering deeper applications of AI in public services and industrial manufacturing [10]. Group 4: AI Efficiency and Edge Computing - The industry is increasingly focused on model efficiency and cost, leading to advancements in model compression, quantization, and distillation techniques [15]. - The development of edge AI models allows for operation on personal computers and IoT devices, reducing latency and enhancing user privacy [17]. Group 5: Embodied Intelligence - The combination of AI technologies with robotics is leading to breakthroughs in embodied intelligence, enabling robots to perform complex tasks in unstructured environments [20]. - The collaboration between hardware advancements and AI models is crucial for real-time interaction and learning from the physical world [21]. Group 6: Developer Incentives - The ModelScope community has launched a developer badge incentive program to reward contributors, providing free GPU computing resources and training vouchers [26]. - The initiative aims to foster a collaborative environment for developers to share ideas and innovate within the community [26].

机器之心报道 编辑：杜伟 知名 AI 技术博主、《Python 机器学习》作者 Sebastian Raschka 又来放福利了！ 今天，他宣布，正值夏季实习和技术面试之际，自己著作《机器学习 Q 与 AI：30 个必备问答》的全部 30 章内容免费开放。他希望能为大家带来帮助，并 祝面试的小伙伴好运。 这本书纸质版（+ 电子版）原价 49.99 美元（约合 358 元），电子版原价 39.9 美元（约合 286 元）。 如今，机器学习和人工智能领域正以前所未有的速度发展。研究人员和从业者常常疲于追赶层出不穷的概念与技术。 本书为你的成长旅途提供了碎片化的知识精华 —— 从机器学习新手到专家，涵盖多个领域的主题。即便是经验丰富的机器学习研究者和从业者，也能从中 发现可纳入自身技能库的新内容 。 评论区有人问，「这本书是用 AI 写的吗？」Sebastian 称当然不是，这样做违背他的个人伦理。有趣的是：这本书的大部分内容写于 2022 年 11 月第一 版 ChatGPT 发布前的几个月，最开始是在 LeanPub 上发布，后来在 2024 年由 No Starch 出版社出版。这本书可能曾是 ChatGPT ...

Core Viewpoint - Meta is aggressively recruiting top talent in the AI field, indicating a strong commitment to making significant advancements in artificial intelligence [1][3]. Group 1: Recruitment Strategy - Meta has successfully recruited 11 top researchers, including individuals from OpenAI and Google, showcasing its ambition to enhance its AI capabilities [2][4]. - The recruitment includes influential figures in their respective technical fields, which could significantly impact Meta's AI development [2]. Group 2: Key Talent Profiles - Alexandr Wang, founder and CEO of Scale AI, has joined Meta to lead the newly established Meta Superintelligence Labs. He is recognized for his contributions to AI and has been compared to Elon Musk [5][8][9]. - Shuchao Bi, a former OpenAI researcher, contributed to the development of GPT-4o and has a strong background in deep learning and optimization, previously generating over $100 million in incremental revenue for Google [10][14]. - Huiwen Chang, with a PhD from Princeton, was involved in creating the GPT-4o image generation system and has extensive experience from her time at Google and OpenAI [16][19]. - Ji Lin, who holds degrees from Tsinghua University and MIT, participated in the development of several key AI models at OpenAI [21][24]. - Shengjia Zhao, a Stanford PhD graduate, was a key contributor to multiple prominent projects at OpenAI, including ChatGPT and GPT-4 [25][28]. - Hongyu Ren, with a strong academic background and experience at major tech companies, contributed to the development of several AI models at OpenAI [31][34][36]. - Pei Sun, previously at Google DeepMind, led the development of perception models for Waymo and has a solid technical skill set [38][40]. - Jiahui Yu, who recently joined OpenAI, has a background in deep learning and high-performance computing, contributing to various AI projects [41][44][47].

本文基于 Yutong Bai、Danny Tran、Amir Bar、Yann LeCun、Trevor Darrell 和 Jitendra Malik 等人的研究工作。 几十年来，人工智能领域一直在思考一个看似简单但非常根本的问题： 如果一个智能体要在真实世界中行动、规划，并且和环境互动，它需要一个怎样的「世界模型」？ 在很多早期研究中，世界模型就是一个预测引擎：只要给它一个抽象的控制指令，比如「向前走一米」或者「向左转 30 度」，它就能模拟出未来的图像。这类方 式在实验室环境里已经发挥过很大作用，但一旦放到真正复杂的人类生活环境，就常常捉襟见肘。 毕竟，人并不是一个漂浮在空中的摄像头。人有四肢、有关节、有骨骼，也有着非常具体的物理限制： 这种「预演」能力让人类能及时修正动作并避免失误。也就是说，我们并不是光靠看到的画面做出决策，而是一直在用大脑里的「想象」，预测动作的结果。 如果未来的 AI 想在真实环境中做到和人一样自然地计划，就需要拥有同样的预测机制：「我如果这样动，接下来会看到什么？」 世界模型的老思路和新思路 这些物理约束决定了：并不是所有动作都能被执行，很多计划只能在可达、可平衡、可承受的 ...

扩散模型（Diffusion Models, DMs）近年来展现出巨大的潜力，在计算机视觉和自然语言处理等诸多任务中取得了显著进展，而异常检测（Anomaly Detection, AD）作为人工智能领域的关键研究任务，在工业制造、金融风控、医疗诊断等众多实际场景中发挥着重要作用。近期，来自多伦多大学、 不列颠哥伦比亚大学 、麻省理工学院、悉尼大学、卡迪夫大学和复旦大学等知名机构的研究者合作完成题为 "Anomaly Detection and Generation with Diffusion Models: A Survey" 的长文 综述，首次聚焦于 DMs 在异常检测与生成领域的应用。该综述系统性地梳理了图像、视频、时间序列、表格和多模态异常检测任务的最新进展并从扩散模型视角 提供了全面的分类体系，结合生成式 AI 的研究动向展望了未来趋势和发展机遇，有望引导该领域的研究者和从业者。 论文标题： Anomaly Detection and Generation with Diffusion Models: A Survey 论文链接： https://arxiv.org/pdf/2506.09368 ...

机器之心报道 编辑：杜伟 该部门将由数据标注初创公司 Scale AI 前 CEO Alexandr Wang 领导，并担任公司首席人工智能官。同时， 扎克伯克还首次 曝光了 11 位从 OpenAI、Anthropic、谷歌 DeepMind 那里挖来的顶尖人才 。 扎克伯格表示，MSL 将吸纳公司的各个团队，致力于开发 Llama 开源系列大模型、相关产品和基础人工智 能研究项目等。 以下是扎克伯克完整的备忘录内容： 随着人工智能进步的加速，发展超级智能已指日可待。我相信这将是人类新纪元的开端，我本人将全力以 赴，确保 Meta 引领这一进程。今天，我想详细介绍一下我们如何调整组织架构，以实现我们的愿景：「为 每个人打造专属的超级智能」。 我们将把整个组织命名为「Meta 超级智能实验室」（Meta Superintelligence Labs，简称 MSL）。 这包括我 们所有的基础研究、产品和 FAIR 团队，以及一个新成立的专注于开发下一代模型的实验室 。 Alexandr Wang 已加入 Meta，担任我们的首席人工智能官 （Chief AI officer） 并领导 MSL。Alexandr ...

图 1 ：指代检测的应用场景实例 最近， IDEA 提出全新解决方案 Rex-Thinker ，首次将人类思维中的 "逻辑推理链" 引入视觉指代任务，让 AI 像人一样分步思考、验证证据，在权威测评中不仅准 确率显著提升，更展现出强大的 "知之为知之" 能力！ Caption ： Rex-Thinker 的思考过程 在日常生活中，我们常通过语言描述寻找特定物体："穿蓝衬衫的人""桌子左边的杯子"。如何让 AI 精准理解这类指令并定位目标，一直是计算机视觉的核心挑 战。现有方法常被两大问题困扰： 决策过程不透明 （"黑箱" 预测）和 拒识能力不足 （对不存在物体输出错误结果）。 Demo论文地址： https://arxiv.org/abs/2506.04034 突破在哪？让 AI 学会 "思考三步走" 传统模型直接输出目标检测框，而 Rex-Thinker 创新性地构建了可解释的推理框架： 1. 规划 （Planning） 拆解语言指令："找到坐在乌龟上的人" → 分解为 "第一步找到乌龟 → 第二步判断每个人是否坐在乌龟上" 2. 验证 （Action ）对每个候选目标（如 "Person 1""Perso ...

Core Insights - The article highlights the unprecedented influence and demand for top talent in the AI era, with companies competing aggressively to attract these individuals [1][2][3]. Group 1: Talent Competition - The supply of top talent is struggling to keep pace with the rapid expansion of internet giants and startups, leading to strong bargaining power for these individuals [2]. - Major internet companies are implementing various high-profile talent acquisition programs, offering top salaries and even unlimited compensation to attract elite talent [4][42]. - The competition for talent is characterized by high intensity, systematic approaches, and a global perspective [3]. Group 2: Talent Development Initiatives - Companies like JD.com are actively engaging with young technical talents through initiatives such as the "JD Technology Salon," which fosters discussions on cutting-edge technologies and talent development [6][7]. - JD.com has launched the "TGT Top Young Technical Talent Program," aimed at graduates from global universities, offering unlimited salaries across eight research areas [42][43]. Group 3: Case Studies of Young Engineers - The article shares the experiences of young engineers like Luochuan, who transitioned from academia to industry, emphasizing the importance of overcoming challenges and adapting to real-world applications [10][21]. - Luochuan successfully identified and addressed technical bottlenecks in AI infrastructure, demonstrating the potential for academic research to be applied in practical scenarios [20]. - Other engineers, such as Qianyi and Tianye, faced challenges in shifting from academic to industrial mindsets, ultimately thriving by focusing on real business pain points and contributing to innovative projects [26][32]. Group 4: Future Outlook - The article concludes that the ongoing investment in talent development and the integration of cutting-edge technology will enable companies like JD.com to maintain a competitive edge in AI, big data, and cloud computing [44].

Core Insights - The article discusses the need for transformation in the architecture of large language models (LLMs), particularly focusing on the limitations of current chain-of-thought (CoT) techniques, which face challenges such as task complexity, high data requirements, and latency issues [2][4]. Group 1: Hierarchical Reasoning Model (HRM) - The Hierarchical Reasoning Model (HRM) is introduced as a novel cyclic architecture inspired by the human brain's layered and multi-timescale processing mechanisms, achieving high computational depth while maintaining training stability and efficiency [3][6]. - HRM operates through two interdependent cyclic modules: a high-level module for slow, abstract planning and a low-level module for fast, detailed computations, achieving remarkable performance on complex reasoning tasks with only 27 million parameters and 1,000 training samples [4][5]. - HRM does not require pre-training or CoT data, yet it performs nearly perfectly on challenging tasks such as complex Sudoku puzzles and optimal pathfinding in large mazes, outperforming larger models with longer context windows [5][6]. Group 2: Design and Mechanisms - The core design of HRM is based on hierarchical processing and time-scale separation, where high-level brain regions integrate information over longer time scales while low-level regions handle immediate sensory information [12][13]. - HRM incorporates feedback loops similar to the brain's dense recurrent neural network connections, enhancing representation accuracy and contextual adaptability while avoiding issues related to backpropagation through time (BPTT) [14][19]. - The model introduces approximate gradients and deep supervision, allowing for efficient memory usage and improved training dynamics, which contrasts with traditional methods that require extensive memory and time [20][23]. Group 3: Performance and Adaptability - HRM demonstrates hierarchical convergence, with the high-level module stabilizing while the low-level module converges repeatedly, leading to rapid convergence and minimal residuals compared to deep neural networks [17][36]. - The model features adaptive computation time (ACT), enabling it to dynamically adjust computational resources based on task complexity, thus optimizing performance without significant resource expenditure [25][27]. - HRM can seamlessly extend inference computation by adjusting parameters without the need for retraining or architectural changes, showcasing its flexibility in handling complex reasoning tasks [28][36]. Group 4: Experimental Results - Experimental results indicate that HRM excels in complex reasoning tasks, raising questions about the underlying reasoning algorithms it employs, which is crucial for enhancing model interpretability [31][39]. - Visualizations of HRM's reasoning processes reveal its strategies in maze and Sudoku tasks, demonstrating a combination of exploration and optimization techniques that resemble depth-first search methods [31][38]. - The hierarchical structure of HRM emerges as a natural characteristic during the learning of complex reasoning tasks, rather than being an inherent property of the model architecture [34].