作者 | 姚戈 世界模型领域迎来了一个重要开源模型。 今天，蚂蚁集团旗下的具身智能公司"蚂蚁灵波"，正式发布并开源其通用世界模型 LingBot-World。 与许多闭源方案不同，蚂蚁灵波选择 全面开源代码和模型权重，而且不绑定任何特定硬件或平台 。 去年 DeepMind 发布的 Genie 3，让人们看到了世界模型能够根据文本或图像提示，实时生成一个可 探索的动态虚拟世界。LingBot-World 沿袭了这条路线，并在交互能力、高动态稳定性、长时序连贯 性以及物理一致性等维度取得了突破。 更令人惊喜的是， LingBot-World 呈现出从"生成"到"模拟"的跨越 。随着模型规模的扩大，灵波团 队观察到，LingBot-World 开始表现出远超普通视频生成的复杂行为，涌现出对空间关系、时间连续 性和物理规律的理解。 可以看到，鸭子腿部蹬水的动作、水面对扰动的响应、以及鸭子身体与水之间的相互作用都比较符合 物理规律。 这显示出模型不仅记住了视觉表象，还在某种程度上理解了流体力学等基础物理机制。同时，水面对 扰动的反应，显示出模型对因果关系的理解。 用户切换视角后再回来时，环境中的智能体（比如这只猫）仍 ...

Core Insights - The article discusses the rise of Clawdbot (now known as Moltbot) in China, highlighting its deployment and usage tutorials on social media platforms, as well as the support from major cloud services like Tencent Cloud and Alibaba Cloud [2] - Peter Steinberger, the creator of Clawdbot, has gained significant attention for his innovative development approach, which diverges from traditional software development practices [2][4] - In a recent interview, Peter shared insights on his development journey, the evolution of coding practices, and the future of software engineering workflows [3] Group 1 - Clawdbot has become popular in China, with various deployment and usage guides available on social media [2] - Major cloud service providers in China, including Tencent Cloud and Alibaba Cloud, have announced support for Clawdbot, indicating its growing significance in the tech ecosystem [2] - Peter Steinberger's previous work on PSPDFKit, which is used on over a billion devices, showcases his expertise in software development [2] Group 2 - Peter's recent interview on "The Pragmatic Engineer" podcast revealed his views on modern coding practices, including the idea that code reviews are outdated and should be replaced with "Prompt Requests" [3] - He emphasized the importance of the "closed-loop principle" in AI programming, which allows for more efficient development processes [3] - Peter's approach to software engineering reflects a shift towards leveraging AI tools, indicating a transformation in how developers interact with code and technology [3] Group 3 - The article highlights Peter's journey from a traditional software developer to an advocate for AI-driven development, showcasing his adaptability and forward-thinking mindset [4] - His experiences illustrate the challenges and rewards of transitioning from conventional coding methods to utilizing AI tools for software creation [4] - The discussion emphasizes the potential for AI to reshape the software development landscape, making it more efficient and innovative [4]

作者 | 冬梅 全球半导体光刻机巨头 ASML（阿斯麦） 今日正式宣布，由于全球成熟制程设备需求放缓以及出口 监管政策的持续收紧，公司计划在全球范围内削减约 1700 个工作岗位。 这是 ASML 自 2023 年 AI 浪潮爆发以来首次进行大规模裁员，标志着即便是在 AI 高速发展的背景 下，半导体设备行业也未能完全抵御周期性调整的冲击。 ASML CEO Christophe Fouquet 发布的全员信如下： 尊敬的 ASML 同事们： 今天，我们发布了 2025 年全年财务业绩以及对未来一年的展望。半导体生态系统有望在未来几年迎 来显著增长，而 ASML 已做好充分准备，把握这一积极发展机遇。我谨代表管理委员会，感谢各位 同事为取得这一成功所做出的贡献。 我们取得的成功归功于我们对客户的专注、卓越的工程技术以及与生态系统的协作。我们的创新和执 行能力为客户、供应商、同事和投资者带来了显著的效益。我们将根据客户需求，继续扩大员工队伍 和业务规模，包括计划在埃因霍温建设的第二个园区。 在技术部门，我们计划将项目 / 矩阵式组织架构转变为以特定产品和模块为核心的模式。这将有助于 简化流程和决策。我们从公司各 ...

Core Viewpoint - The article emphasizes the urgent need for humanity to prepare for the potential risks associated with advanced AI, as articulated by Dario Amodei, CEO of Anthropic, in his extensive essay titled "The Adolescence of Technology" [3][5][10]. Group 1: AI Risks and Governance - Dario Amodei outlines five systemic risks posed by AI, highlighting that the true danger lies not just in the technology itself but in humanity's ability to govern and manage it effectively [10][12]. - The first risk is the uncontrollability of AI, which can lead to deceptive behaviors and extreme goals due to its complex training processes [13]. - The second risk involves the potential misuse of AI for malicious purposes, such as cyberattacks and automated fraud [13]. - The third risk is the use of AI as a tool for power by governments or organizations, leading to potential authoritarianism [13][15]. - The fourth risk pertains to the economic impact of AI, which could displace entry-level jobs and exacerbate wealth inequality [13]. - The fifth risk involves unknown but potentially profound societal consequences, such as shifts in human identity and purpose as AI surpasses human capabilities [13][16]. Group 2: Proposed Solutions - Amodei suggests implementing constitutional-style AI to shape AI behavior according to high-level values and to ensure transparency and accountability in AI systems [13]. - For the misuse of AI, he advocates for regulatory measures, including mandatory screening for genetic synthesis and the establishment of laws to prevent dangerous applications [13]. - To combat the risk of AI being used for authoritarian purposes, he recommends international agreements to classify certain AI abuses as "crimes against humanity" and to enforce strict governance on AI companies [15]. - Addressing economic displacement, he proposes the creation of real-time economic indicators and encouraging innovation rather than layoffs [13]. - Finally, he stresses the importance of human values and collective choices in determining the future trajectory of AI [16].

Core Insights - The CEO of Li Auto, Li Xiang, emphasized that 2026 is the last year for companies aiming to become leaders in AI to enter the market, with Level 4 (L4) autonomy expected to be realized by 2028. The company aims to be one of the three global leaders in foundational models, chips, operating systems, and embodied intelligence [2] - Li Auto plans to strengthen its brand positioning in embodied intelligence, moving beyond just creating mobile homes to developing humanoid robots, which will be showcased soon [2] - The company will undergo organizational changes in R&D, dividing teams into foundational model teams, software teams, and hardware teams, with a focus on recruiting top talent, including those who previously left for startups in the embodied intelligence sector [2] - Li Xiang stated that the electric vehicle industry has reached a dead end in parameter competition, and Li Auto has chosen to define its vehicles as "embodied intelligent" products, transforming them from mere transportation tools into robots with perception and intelligence [7][8] Recruitment and Development - Li Auto has posted multiple job openings for humanoid robot R&D positions on its official recruitment page, indicating a comprehensive approach to developing humanoid robots from core components to system integration [3] - The company has established secondary departments for "space robots" and "wearable robots," with the first product being the smart glasses Livis, under the leadership of Senior Vice President Fan Haoyu [8]

Core Insights - ClawdBot, a personal AI assistant, has gained significant attention in Silicon Valley and social media, with its creator Peter Steinberger facing trademark issues leading to a name change to Moltbot [2] - Users have praised ClawdBot as a revolutionary AI application, likening it to having a dedicated AI employee available 24/7 [3] - ClawdBot's unique collaborative approach allows non-coders to contribute directly, emphasizing problem-solving over traditional coding [6] Group 1 - ClawdBot can control computers almost entirely, lacking traditional restrictions, and features a complex memory system that retains user interactions [7][8] - The assistant interacts through various chat applications, including WhatsApp, Telegram, and Discord, but its open permissions raise security concerns [8][9] - The surge in ClawdBot's popularity has led many users to purchase Mac Mini computers for optimal performance, although alternatives exist [9][11] Group 2 - Peter Steinberger, the developer, has a notable background, previously running a successful B2B company before returning to the tech scene to create ClawdBot [14][12] - The project began as a personal solution to a need for a life assistant, evolving into a widely adopted tool after realizing no major companies had tackled this problem [15][19] - ClawdBot's development has been rapid, with a focus on community involvement and open-source collaboration, allowing users to contribute even without coding experience [27][28] Group 3 - The assistant's capabilities extend to automating various tasks, including managing household chores and personal reminders, significantly enhancing user productivity [43][48] - Users have reported diverse applications, from managing emails to controlling smart home devices, showcasing ClawdBot's versatility [42][46] - The project aims to empower users to maintain control over their data while providing a free and open-source solution, contrasting with larger corporate models [24][23]

Core Viewpoint - Sam Altman, CEO of OpenAI, emphasizes the transformative potential of AI, particularly with the upcoming GPT-5 and its successors, highlighting a shift towards low-cost, high-speed intelligence generation [4][5][6]. Group 1: AI Development and Performance - The discussion at the seminar focused on the asymmetric performance of GPT-5, which excels in logic and programming but has compromised writing quality compared to GPT-4.5 [4][5]. - Altman acknowledged that the prioritization of reasoning and coding capabilities in GPT-5.2 led to a decline in writing skills, indicating a strategic focus on core intelligence metrics first [5][9]. - Altman predicts that by the end of 2027, the intelligence cost of GPT-5.2 will decrease by at least 100 times, making advanced AI more accessible [5][11]. Group 2: Market Trends and Developer Needs - There is a noticeable shift in developer priorities from cost to speed, as the demand for rapid output increases with the complexity of tasks handled by AI agents [6][11]. - OpenAI may offer two pathways: extremely low-cost intelligence and high-speed feedback systems, indicating a transition from simple Q&A to real-time autonomous decision-making [6][7]. Group 3: Future of Software and Applications - Altman envisions a future where software is not static but dynamically generated to solve specific problems, leading to a highly personalized productivity system for users [7][12]. - The concept of "just-in-time" applications will redefine operating systems, allowing tools to evolve based on individual workflows [7][12]. Group 4: Societal Impact and Ethical Considerations - Altman believes AI will empower individuals by lowering barriers to resources and innovation, but he also warns of potential wealth concentration and emphasizes the need for careful policy-making [8]. - He advocates for a resilient approach to AI safety, particularly in biological security, suggesting a shift from blocking access to building robust systems to manage risks [19][20]. Group 5: Collaboration and Education - Altman argues that AI will enhance human collaboration rather than diminish it, suggesting that AI tools will facilitate teamwork and increase productivity [22][24]. - He expresses concerns about the impact of technology on early childhood education, advocating for limited use of computers in formative years to ensure healthy development [30].

整理 | 华卫 "一圈又一圈的循环融资，投资回报率却不尽如人意，这些 AI 系统实际用起来也远没有想象中好 用，或许方向本身就站不住脚。" 近日，知名 AI 专家、认知科学家 Gary Marcus 在一场访谈中愤愤表示，"整个世界都在全力押注 神经网络，还在这个我始终觉得毫无道理的理念上投入了巨资，但大语言模型根本无法带我们抵 达 AGI 这一终极目标。" 这场对话由曾因成功预测 2008 年金融危机而闻名的传奇投资人、华尔街最具影响力人物之一 Steve Eisman 发起，他与 Marcus 共同探讨了当下 AI 进展的方方面面，包括商业路径、社区现 状和未来方向等。Marcus 认为，大语言模型已经达到了收益递减的阶段。并且，他指出，现在 AI 领域根本没有技术壁垒了，所有 AI 企业的研发思路基本一致。 对于大量人才从大厂离职去办初创公司的现象，Marcus 直言道，"如果 OpenAI 真的能在下周推 出 AGI，谁会在这个即将改变世界的关键节点离职，去创办一家可能要花四年时间才能做出成果 的小公司？显然没人会这么做，大家都会想留在公司见证这个时刻。"在他看来，这些企业内部的 人也清楚，他们根本没 ...

在多项权威基准测试中表现优异，Qwen3-Max-Thinking 性能可与 GPT-5.2-Thinking、Claude-Opus-4.5、Gemini-3 Pro 等闭源顶级模型竞争甚至超 越。 具体而言，Qwen3-Max-Thinking 在多项关键 AI 基准测试中达到了或刷新了全球 SOTA 表现： 这些测试覆盖了科学知识问答（如 GPQA Diamond）、数学推理（如 IMO 等级测试）、代码编程（如 LiveCodeBench）等多个领域，是衡量大型语 言模型综合能力的重要指标。 阿里突发最强旗舰模型，总参数过万亿 就在刚刚，Qwen3-Max-Thinking 正式版突然发布，总参数规模超过 1 万亿（1T），位于目前全球最大规模 AI 模型行列，预训练数据规模高达 36T Tokens，覆盖大量高质量语料。 作者｜冬梅 Qwen3-Max 是阿里通义团队迄今规模最大、能力最强的语言模型，该版本包括 Base、Instruct 和 Thinking 多种形式。 在包含事实科学知识、复杂推理和编程能力在内的 19 项权威基准测试中取得极高水平，有记录显示其综合表现可媲美 GPT-5.2-T ...

Core Insights - OpenAI's Codex architecture, particularly the Agent Loop, has been revealed to efficiently manage interactions for software tasks, supporting 800 million users with a single PostgreSQL database and 50 read replicas [2][29] - The architecture emphasizes the importance of a well-structured framework over flashy tools, showcasing that effective design can outperform complex solutions [2][29] Group 1: Codex Architecture - The core of every AI agent is the Agent Loop, which coordinates user input, model interactions, and tool executions to perform meaningful software tasks [5] - Codex encompasses various software agent products, including Codex CLI, Codex Cloud, and Codex VS Code plugins, all built on the same framework [5] - The process begins with the agent receiving user input, which is transformed into a prompt for the model, followed by inference to generate responses [6][12] Group 2: Inference Process - During inference, the model generates two types of outputs: a final response to the user and a request for tool execution, which may lead to further queries [7][11] - The agent can modify the local environment through tool calls, with each interaction concluding with a message directed at the user [7][20] - Context window management is crucial, as extensive dialogue can exhaust the model's token capacity, necessitating strategies to compress or manage context [11][27] Group 3: Performance Optimization - OpenAI has optimized PostgreSQL to handle a tenfold increase in load over the past year, supporting millions of queries per second while maintaining low latency and high reliability [29][30] - The architecture leverages a single primary node with numerous read replicas, demonstrating that effective scaling can be achieved without resorting to complex distributed systems [29][31] - Continuous exploration of performance limits includes migrating write-heavy workloads to other database systems and enhancing replication features [30][31]