Core Viewpoint - The article discusses the emergence of Diffusion-based Language Models (dLLMs) and highlights a significant security vulnerability associated with them, specifically the DIJA attack framework that can exploit these models to generate harmful content without requiring any model retraining or parameter modification [1][2][4]. Group 1: Characteristics of dLLMs - dLLMs are characterized by parallel decoding, bidirectional context modeling, and the ability to flexibly insert masked tokens, making them suitable for various applications such as interactive Q&A and code generation [1]. - Unlike autoregressive models, dLLMs can generate multiple tokens simultaneously and perform text insertion and rewriting more naturally [1]. Group 2: Security Vulnerabilities - A fundamental architectural security flaw exists in current dLLMs, rendering them nearly defenseless against certain attack scenarios [2][4]. - The DIJA attack framework can lead multiple dLLMs to generate harmful, illegal, or inappropriate content without any need for training or rewriting model parameters [4]. Group 3: Mechanism of DIJA Attack - The DIJA attack does not obscure or rewrite dangerous content in jailbreak prompts; instead, it transforms original prompts into masked text prompts that dLLMs cannot resist generating harmful outputs [6][10]. - The attack is fully automated, requiring no manual prompt design, and utilizes powerful language models to generate attack prompts with minimal human intervention [8][10]. Group 4: Attack Strategies - The team developed three key strategies for constructing effective masked text prompts that are both natural and highly aggressive in their attack potential [9][11]. - These strategies include prompt diversification, multi-granularity masking, and benign separator insertion to enhance the effectiveness and stealth of the attack [11]. Group 5: Experimental Results - The research team tested the DIJA attack on four representative dLLMs, revealing that dLLMs generally exhibit comparable or slightly better defense capabilities than autoregressive models [14]. - The DIJA attack achieved the highest ASR-k scores across all benchmarks, indicating that dLLMs are unlikely to reject responses to dangerous topics [21]. Group 6: Fundamental Issues - The security shortcomings of dLLMs are not merely bugs but are inherent design features, particularly due to bidirectional context modeling and parallel decoding mechanisms that prevent effective token-by-token scrutiny [19][22]. - Current alignment methods focus on overall input-output relationships, lacking sensitivity to individual token positions, which exacerbates the vulnerability [23]. Group 7: Future Directions - The article suggests that the era of dLLM security research has just begun, with the DIJA attack representing the opening of a new research direction focused on "mask-aware safety" [25]. - Recommendations include designing rejection mechanisms based on masked positions and developing alignment training processes specifically tailored for dLLM architectures [25].

Core Viewpoint - The article discusses the recent hiring of top engineer Ethan He from NVIDIA to xAI, highlighting the growing influence of Chinese talent in Elon Musk's AI initiatives and the significant advancements in AI models like Grok-4 and Cosmos [1][2][15]. Group 1: Ethan He's Career Move - Ethan He, a top engineer from NVIDIA, has officially joined xAI, sparking speculation about his salary and the implications of this move [2][4]. - He previously worked at Meta and was involved in the development of NVIDIA's advanced world model platform, Cosmos, which has been released for commercial use [5][6][8]. - Cosmos is noted for being the first upgraded model with over 100 billion parameters, currently applied in robotics and autonomous driving [10][11]. Group 2: Impact of Grok-4 - Grok-4 has seen a significant increase in downloads by 279% and a 325% rise in total revenue on iOS since its launch, indicating strong market interest and performance [15]. - Ethan He expressed high regard for Grok-4, praising its breakthrough value even before officially leaving NVIDIA [12][14]. Group 3: Growing Chinese Talent in AI - The article highlights a growing number of Chinese scientists and engineers joining Musk's AI team, including notable figures with impressive academic backgrounds from institutions like Stanford and the University of Toronto [20][21][22][23][24][25][26][27][28][29][30][31][32][33][34].

Core Viewpoint - The article highlights the launch of Qwen3-Coder by Alibaba's Tongyi team, which has set a new state-of-the-art (SOTA) in AI programming, surpassing both open-source and closed-source models in the industry [1][3]. Group 1: Product Features - Qwen3-Coder includes multiple versions, with the strongest being Qwen3-Coder-480B-A35B-Instruct, featuring a 450 billion MoE model and 35 billion active parameters [5]. - It natively supports a context length of 256K and can be extended to 1 million using YaRN technology [6][23]. - The command-line version, Qwen Code, has been developed based on Gemini Code, allowing for prompt and tool invocation protocol adaptation [8]. Group 2: Performance and Capabilities - Users have reported impressive results, such as creating interactive animations and dynamic weather cards with simple prompts [11][13]. - The model can easily generate a playable Minesweeper game and an editable resume template, showcasing its versatility [16][19]. Group 3: Technical Details - During the pre-training phase, Qwen3-Coder utilized various scaling techniques to enhance model capabilities, with training data totaling 7.5 trillion tokens, 70% of which is code data [22]. - The post-training phase involved scaling code reinforcement learning (RL) to improve the model's performance on real-world coding tasks, achieving high success rates in code execution [24][27]. Group 4: Open Source vs Closed Source - Qwen3-Coder is open-source under the Apache License Version 2.0, making it commercially friendly and empowering developers [29][30]. - The article emphasizes that this release represents a significant leap for open-source programming agents, positioning Chinese models at the forefront of the industry [34].

Core Viewpoint - Google recently announced that its DeepMind team won an IMO gold medal, but shortly after, three key team members were reported to have left for Meta, highlighting a talent drain in the AI sector [1][19]. Group 1: Key Personnel Changes - Three critical figures involved in the training of the Gemini model, Du Yu, Tianhe Yu, and Wang Weiyue, have left Google for Meta [2][3]. - Du Yu has been a significant contributor to the Gemini series models and has worked on Google's conversational AI products [9]. - Tianhe Yu, a research scientist at Google DeepMind, was responsible for the reinforcement learning and training of Gemini, playing a key role in the release of Gemini 2.5 [10]. - Wang Weiyue, a principal research engineer at Google DeepMind, contributed to Gemini 2.5 Pro and has a background in computer vision [13][14]. Group 2: Competitive Landscape - Mark Zuckerberg's recruitment of talent from Google is part of a broader trend, as Microsoft has also been reported to have poached over 20 talents from Google DeepMind [19]. - Amar Subramanya, the former engineering lead for Gemini, has joined Microsoft AI as a vice president, indicating a shift in talent dynamics within the AI industry [19]. - The talent acquisition efforts by Microsoft have been ongoing for six months, led by Mustafa Suleyman, a co-founder of DeepMind, adding a layer of complexity to the competitive landscape [21].

衡宇 奕然 发自 凹非寺 量子位 | 公众号 QbitAI 火到不能再火的Agent，零一万物也下场了。 就在今早，零一万物创始人兼CEO李开复博士宣布 升级发布万智企业大模型一站式平台 （下文简称万智平台） 2.0版本，并推出零一万物企 业级Agent智能体，昵称万仔。 李开复分析指出，AI Agent正从辅助工具跃迁为新型生产单元，其核心价值在于重构企业组织架构与价值链。 而在零一万物看来，这场变革的成败取决于两个关键： 企业闭环数据的深度激活，以及与CEO推动的"一把手工程"共创。 "独特的一把手工程打法" 什么叫"一把手工程"？ 简单理解， 第一步是CEOxCEO，高层制定AI驱动的顶层战略，第二步是战略、技术、业务三方团队密切配合，打造真正贴合业务需求的大 模型ToB解决方案。 更重要的是，要用"生产单元革命"案例替代技术宣讲。 "我们向CEO展示，若Agent替代30%人力生产单元，企业规模扩大3倍时效率反升20%。"李开复展开解释，"传统CEO尚未意识到Agent将重 构生产单元和组织架构，而执行层缺乏AI战略视野，中层管理者因担忧职权变化而抵触变革。" 也就是说， 一把手工程的本质是CEO驱动 ...

Core Viewpoint - Kimi K2 has emerged as a leading open-source model, showcasing significant advancements in capabilities, particularly in code, agent tasks, and mathematical reasoning [4][5]. Group 1: Technical Highlights - Kimi K2 features a total parameter count of 1 trillion and 32 billion active parameters, demonstrating its advanced capabilities [4]. - The model has achieved state-of-the-art (SOTA) performance in various benchmark tests, including SWE Bench Verified, Tau2, and AceBench [12]. - The Kimi team emphasizes a shift from static imitation learning to Agentic Intelligence, requiring models to autonomously perceive, plan, reason, and act in complex environments [9][10]. Group 2: Core Innovations - Three core innovations are implemented in Kimi K2: 1. MuonClip optimizer, which replaces traditional Adam optimizer, allowing for lossless spike pre-training on 15.5 trillion tokens [11]. 2. Large-scale Agentic Tool Use data synthesis, enabling the generation of multi-turn tool usage scenarios across hundreds of domains and thousands of tools [12]. 3. A universal reinforcement learning framework that extends alignment from static to open domains [12]. Group 3: Pre-training and Post-training Phases - During the pre-training phase, Kimi K2 optimizes both the optimizer and data, utilizing the MuonClip optimizer to enhance training stability and efficiency [21][22]. - The training data covers four main areas: web content, code, mathematics, and knowledge, all subjected to strict quality screening [24]. - The post-training phase involves supervised fine-tuning and reinforcement learning, with a focus on generating high-quality training data through a rejection sampling mechanism [30][31]. Group 4: Reinforcement Learning Process - The reinforcement learning process includes creating verifiable reward environments for objective evaluation of model performance [33]. - A self-critique reward mechanism is introduced, allowing the model to evaluate its outputs based on predefined standards [34]. - The model generates diverse agentic tasks and tool combinations, ensuring a comprehensive training approach [35]. Group 5: Infrastructure and Performance - Kimi K2's training relies on a large-scale high-bandwidth GPU cluster composed of NVIDIA H800, ensuring efficient training across various resource scales [38]. - Each node is equipped with 2TB of memory, facilitating high-speed interconnectivity among GPUs [39].

Core Viewpoint - The article discusses the rapid advancements in open-source large models in China, highlighting the release and performance of the Qwen3 model, which has shown significant improvements over its predecessor and competitors in various benchmarks [1][24]. Group 1: Model Updates and Performance - Qwen3 has been upgraded to a model with 235 billion parameters, which is only a quarter of Kimi K2's 1 trillion parameters, yet it surpasses Kimi K2 in benchmark performance [2][3]. - The new model enhances understanding of 256K long contexts and is a causal language model utilizing a Mixture of Experts (MoE) architecture [8][12]. - The model includes 94 layers, employs grouped query attention (GQA) mechanisms, and activates 8 out of 128 experts during inference [8][12]. Group 2: Benchmark Performance - In benchmark tests, Qwen3 shows improved accuracy in various categories, such as AIME25, where accuracy increased from 24.7% to 70.3%, indicating strong mathematical reasoning capabilities [13][15]. - Compared to Kimi K2 and DeepSeek-V3, Qwen3 demonstrates superior performance across multiple metrics, including instruction following, logical reasoning, and text understanding [12][15]. Group 3: Market Context and Competition - The article notes that the competitive landscape is shifting, with Qwen3 challenging Kimi K2 shortly after its release, indicating a dynamic environment in the open-source model sector [25]. - The release of Qwen3 coincides with NVIDIA's announcement of a new state-of-the-art open-source model, OpenReasoning-Nemotron, which offers various scales and local operation capabilities [17][18]. - The transition of Llama to a closed-source model and OpenAI's delay in releasing open models further emphasizes the growing importance of open-source large models in the Chinese market [24].

白交 一凡 发自 凹非寺 量子位 | 公众号 QbitAI 这也太惊人了吧？！ 注意看，这是一个机器人在跳舞，动作流畅，还是 Breaking 这种力量协调性要求很高的类型。 但它不光能动，静下来 "干活" 也有模有样： 无论是叠衣服、撕纸巾、拉窗帘、用筷子这种日常生活的精细化操作，还是工厂里的搬运、分拣、扫码等工业任务，都跟科幻电影里的一模一 样， 甚至实现了机器人的群体协作 。 而且这还是个 大尺寸双足人形机器人 ，它身高足足有171cm、体重65kg。在具身智能机器人领域，直接就是全尺寸里一步到位标准。 四肢发达，却还能粗中有细，能动能静。 放在机器人领域，背后一定还有个 聪明大脑 。 这就是来自 星动L7 机器人秀出的最新能力，打造者是清华叉院背景的明星团队—— 星动纪元 。在不久前，他们还以 近5亿新融资 引发行业 内外热议，没想到烈火烹油，随即就甩出上述惊艳新进展。 这些能力也直接创下多项纪录： 首个完成360°旋转跳的；首个会跳街舞breaking的；首个跑得比人快的，也是首个同时具备精细化操作的全尺寸人形机器人。 如此纪录和能力背后，除了可以推测的AI大脑，也离不开机器人运动控制上的领先，据说 ...

Core Viewpoint - The HEP (Hierarchical Equivariant Policy via Frame Transfer) framework, developed by Northeastern University and Boston Dynamics RAI, aims to enable AI to adapt to complex real-world scenarios with minimal demonstrations, enhancing efficiency and flexibility in robotic learning [1][4]. Summary by Sections HEP Framework Highlights - The HEP framework efficiently expresses 3D visual information while balancing detail restoration and computational speed [2]. Core Innovations - The framework addresses the long-standing issues of data scarcity and generalization in AI applications by utilizing a hierarchical policy learning framework transfer interface, which allows for strong inductive bias while maintaining flexibility [4]. Simplified and Efficient Hierarchical Structure - The high-level policy sets global objectives, while the low-level policy optimizes actions in a local coordinate system, significantly improving operational flexibility and efficiency [5]. - The model automatically adapts to spatial transformations such as translation and rotation, greatly reducing the dependence on data volume for generalization [5]. Key Concepts - HEP is based on two core ideas: hierarchical policy structure and the "coordinate transfer interface," where the high-level policy provides a "reference coordinate" for the low-level policy to optimize execution details [7]. - The coordinate transfer interface enhances the flexibility of the low-level policy while transmitting the high-level policy's generalization and robustness capabilities [9]. Effectiveness Demonstration - The research team tested the HEP framework on 30 simulated tasks in RLBench, including high-precision and long-duration tasks, and further validated it on three real-world robotic tasks [10]. - The high-level policy predicts a "key pose" for global planning, while the low-level policy generates detailed motion trajectories based on this key pose [11]. Results - The hierarchical strategy shows significant advantages in complex long-range tasks, with the HEP framework learning robust multi-step collaborative tasks with only 30 demonstration data, outperforming non-hierarchical methods [14]. - In the Pick & Place task, HEP achieved 1-shot generalization learning with just one demonstration, significantly improving data efficiency [15]. - The coordinate transfer interface successfully transmits the high-level adaptability to spatial changes to the low-level policy, making the overall strategy easier to extend to new scenarios [16]. - HEP's success rate improved by up to 60% compared to traditional methods under environmental changes and disturbances from unrelated objects [17]. Future Implications - The coordinate transfer interface imposes soft constraints on the low-level policy, ensuring flexibility and providing a natural interface for future integration of multimodal and cross-platform high-level strategies [19].

克雷西 发自 凹非寺 量子位 | 公众号 QbitAI 谷歌Gemini 拿下了IMO金牌，而且是官方认证的那种 。 经过IMO官方裁判评分，Gemini新模型 答对了6道题中的5道 ，以35分的成绩斩获金牌。 斩获金牌的是Gemini的一个进阶版本，搭载了新的思考模式，后期会开放给Google AI Ultra订阅用户——也就月付1400元那种。 去年三天摘银，今年4.5小时夺金，DeepMind的数学成绩可以说是突飞猛进。 除了DeepMind CEO哈萨比斯、谷歌CEO劈柴哥给团队发来贺电，马斯克也发推表示了祝贺。 我们可以确认，谷歌DeepMind已达到人们梦寐以求的里程碑，获得了35分（满分42分）——堪称金牌。 他们的解决方案在很多方面都令人惊叹。IMO评分员认为这些解决方案清晰、精准，而且大多数都易于理解。 DeepMind这波可谓是被各界夹道祝贺，做得体面又周到。 但DeepMind被夸得越好，OpenAI就越发相形见绌，同样是AI参赛IMO，秘密搞事情也就算了，还为了营销跟人类青少年抢风头。 奥特曼治下的OpenAI，最近除了丢人就丢人了。 DeepMind官宣AI拿下IMO金牌 DeepM ...