Core Insights - The article discusses the potential of AI startups in the context of globalization, emphasizing the need for these companies to target global markets from their inception [1][2]. Group 1: Event Overview - A salon event is organized featuring industry leaders from companies like Xiaoying Technology, FluxA, Google, JD, Agora, and Meshy, who will share reusable experiences in global expansion [4]. - The event aims to provide insights into the application, scenarios, and channels for AI startups looking to expand internationally [6][9]. Group 2: Guest Profiles - Lin Xiaodong, Vice President of Xiaoying Technology, has led the development of video editing tools and AI music applications that have achieved top rankings in various markets [12]. - Qiu Honglin, CTO of FluxA, is a co-founder with a background in architecture from Ant Group and Alibaba [13]. - Han Yuan, a Solutions Architect at Google Cloud, focuses on cloud solutions for AI applications [15]. - Li Jieyu, Product Head at JD, is working on AI-driven innovations in e-commerce [17]. - Yang Fan, AI Product Growth Lead at Agora, has extensive experience in the audio-visual and mobile internet sectors [18]. - Xu Shumu, responsible for 3D art at Meshy, has over 10 years of experience in the gaming and film industries, focusing on AI applications in these fields [19].

允中 发自 凹非寺 量子位 | 公众号 QbitAI 智元直接把一整套硬核装备塞进你的背包： 业界顶配机器人开发平台、海量真实数据集、以及通往机器人学术圣殿IEEE ICRA舞台的传送门 。 更有总价值超 百万美元 的奖金与资源掉落，等你来开箱！ 两大赛道，决战具身智能之巅 全球具身智能的极客玩家，是时候登场了！ 还在为缺乏 顶级硬件平台验证算法 而苦恼？还在为 仿真环境和真实产线差距太大 而挠头？还在苦恼自己的 顶尖算法只能在PPT上跑 ？ 现在，你的"天命副本"来了。 由 智元（AGIBOT） 主办的机器人领域顶级赛事—— AGIBOT WORLD CHALLENGE @ICRA 2026 ，正式向全球豪杰发出召集令！ 大赛共设置 推理-操作 和 世界模型 两大赛道，不管你擅长的是让机器人动手干活，还是让机器人动脑预判，这里都能找到你的主场。 1、Reasoning to Action（推理-操作）赛道 本赛道旨在评估模型的推理和动作执行能力，包括线上仿真和线下真机赛段。 参赛者将基于AGIBOT WORLD开源数据集，训练能够解决一系列复杂任务的模型。 本赛道聚焦于弥合Sim2Real Gap，实现从开 ...

Core Insights - The article discusses the launch of Deer-Flow2, a modular multi-agent management framework developed by ByteDance, which has quickly gained popularity on GitHub with 35.3k stars [2]. Group 1: Framework Features - Deer-Flow2 utilizes a modular multi-agent architecture, enabling agents to collaborate through LangGraph, and comes pre-equipped with various search engines and crawling tools [4]. - The framework supports extensibility, allowing users to customize APIs or models easily [5]. - Key capabilities include multi-agent collaboration, sandbox security execution, and one-click deployment, compatible with mainstream large models [6]. - The framework supports native integration with communication channels like Feishu, Telegram, and Slack, allowing operation without a public IP [7]. Group 2: Version Improvements - Version 1.0 featured a fixed 5-node multi-agent architecture focused on deep research scenarios [10]. - Version 2.0 has undergone a complete structural overhaul, adopting a new architecture with a single main agent, 11 middleware layers, and dynamic sub-agents, making the system lighter, more flexible, and easier to expand [11]. - The core capability of deep research has transitioned to a foundational ability within the framework, which now includes key modules like sub-agent scheduling and long-term memory [13]. Group 3: Skill System - Deer-Flow 2.0 features a pluggable skill system, pre-loaded with over ten common skills such as deep research, data analysis, and multimedia creation, which can be incrementally loaded based on task requirements [15]. - Users can create custom skills using the provided skill-creator tool, allowing for rapid expansion of agent capabilities [18]. Group 4: Execution Environment - The framework includes an independent isolated sandbox for each task, providing a complete file system and Bash execution permissions, supporting file read/write and script execution [20]. - It offers three operational modes: local, Docker, and Kubernetes, with Docker mode providing higher isolation and stability [21][22]. Group 5: Task Management - Deer-Flow 2.0 employs a scheduling mechanism and context engineering to manage complex long-duration tasks, allowing the main agent to structure tasks and dispatch up to three sub-agents for parallel execution [25]. - Each sub-agent operates in an independent context, preventing interference and pollution, while the framework addresses context window limitations through various design features [27]. Group 6: Deployment Instructions - The article provides detailed instructions for deploying Deer-Flow using Docker or locally, with Docker being the simpler option requiring minimal commands [33][39]. - Local deployment requires specific prerequisites and allows for source code modification and debugging [40][48]. Group 7: Communication Integration - Deer-Flow natively supports task reception from instant messaging applications, specifically Telegram, Slack, and Feishu/Lark, without needing a public IP [49].

克雷西 发自 凹非寺 量子位 | 公众号 QbitAI 版本号没变，API供应商却悄悄偷换模型？现在这种小伎俩可以轻松被戳穿了。 来自法国的研究人员，开发出了新的检测技术，为识别云端模型的隐秘变动提供了"照妖镜"。 仅需极短的提示词，通过1个输出token，就能实现持续监控，成本只有传统方式的千分之一。 而且这种高度灵敏的方法，能捕捉到仅经过一个训练步数微调的模型差异。 这样一来，供应商出于成本考量而私自替换量化版本或降级模型的行为，将无所遁形。 而用户最常面对的是完全的黑盒环境，此时只能提交提示词并接收生成的最终文本，无法窥视任何中间运算逻辑。 回到灰盒环境，对数概率揭示了模型在生成Token时对整个词汇表空间的信心分布，对数概率追踪技术正是利用这些数据作为识别模型身份的 数字指纹。 由于权重的微调、量化或架构变动都会导致这一分布发生偏移，对数概率成为了监测后台变动的最敏感指标。 灰盒环境下的对数概率追踪 第一种检测手段名为对数概率追踪，它主要针对灰盒访问环境进行设计。 所谓灰盒环境，是指供应商虽然不公开核心模型权重，但允许用户通过API获取每个输出Token对应的对数概率，这种权限介于白盒与黑盒之 间。 ...

UniCalli团队 投稿 量子位 | 公众号 QbitAI 苦于AI单字拼凑没行气，或是排版秒变"鬼画符"？ 这个痛点，终于被终结了。 现在，只需输入一段文字，就能让AI立刻化身王羲之、颜真卿或是米芾，全自动挥毫泼墨。 UniCalli ，这个由香港科技大学 （广州） 等团队推出的全新统一扩散框架，不仅能完美拿捏书法的整列排 版 （Column-level） ，甚至连相邻字符之间大小错落的缩放、自然流畅的游丝连笔 （Ligatures） 都能精 准生成。 更重磅的是，它首次将"书法生成"和"古籍识别"两大任务统一在了同一个模型里。目前，该工作已被 ICLR2026 正式接收。 代码、超大规模数据集全部开源，还同步上线了可一键在线试玩的Demo！ huggingface模型蒸馏后效果，十秒能同步生成数张列级书法作品： △ 从左到右是：文征明/行，乾隆/楷，米芾/行，王羲之/草，王羲之/行，怀素/草，宋徽宗/楷瘦金体 △ 从左到右是：宋克/草，苏轼/行，文天祥/草，颜真卿/楷，赵孟頫/楷，黄庭坚/行 降维打击：连笔、排版都能搞定 技术方案：生成与识别的"双向奔赴" UniCalli的核心思路在于"统一" （Un ...

Core Viewpoint - The article emphasizes the ongoing AI boom and invites individuals to join the company "Quantum Bit," which focuses on tracking AI advancements and has established itself as a leading content platform in the industry [1]. Group 1: Job Opportunities - The company is hiring for three main directions: AI Industry, AI Finance, and AI Product, with positions available for both experienced professionals and fresh graduates [3][4]. - Positions are full-time and based in Beijing, with roles ranging from editors to chief editors, tailored to match individual capabilities [4][5]. Group 2: AI Industry Direction - Responsibilities include tracking innovations in foundational layers such as chips, AI infrastructure, and cloud computing, as well as interpreting cutting-edge research and technical reports [6][8]. - Candidates should have a basic understanding of chips, GPUs, NPUs, servers, and cloud computing, with a preference for those with technical backgrounds in engineering or computer science [8]. Group 3: AI Finance Direction - This role focuses on venture capital, AI startups, public companies, and capital movements within the industry, producing analyses of financing, IPOs, and company strategies [7][8]. - Candidates should be data-sensitive, interested in financial reports and strategic planning, and possess strong logical structuring skills [8]. Group 4: AI Product Direction - The position involves monitoring the implementation of AI in software and hardware products, conducting in-depth evaluations, and engaging with entrepreneurs and product experts [9][11]. - Candidates should be keen on smart hardware trends and possess strong logical and structured communication skills [11]. Group 5: Company Overview - By 2025, Quantum Bit aims to have over 2.4 million subscribers on WeChat and more than 7 million users across platforms, with a daily reading volume exceeding 2 million [10]. - The company is recognized as the top new media outlet in the AI and frontier technology sector according to third-party data platforms [10].

衡宇 发自 凹非寺 量子位 | 公众号 QbitAI "对具身智能来说，力觉比视觉更重要。" 听起来，这句话好像有那么点非主流。放眼当下，大多数具身智能的叙事，视觉几乎是机器人认知世界的第一扇窗。 但源自斯坦福机器人和人工智能实验室的通用机器人独角兽 非夕科技 ，却十年如一日地坚持另一条路径：按照真实操作行为的重要性排序， 力觉能力优先于视觉。 联合创始人兼CEO王世全的解释很直白："人类在做大多数操作时，并不会一直盯着目标，只用'瞟一眼'，剩下的动作主要依赖手感完成，核 心在于实时控制施加的力度，以及与物体的接触关系。" 朋友们，咱天天把具身智能挂在嘴边，但 别光顾着智能，忽视了"具身"啊 。 本质上，具身智能的难点在于如何"在真实世界完成操作"，而 力觉正是实现机器人具身交互的核心能力 。 过去十年，非夕将大量精力倾注于自研机械臂本体，尤其专注于提升力控能力，尝试让它具备接近人类手臂的"操作感"。 从最初的Rizon系列7轴机械臂到Moonlight系列力控并联机器人，非夕的自适应机器人已经深入汽车、消费电子、一般工业、食品加工、实验 室、医疗理疗、商用服务等各大领域，源源不断形成富有创新、独特差异化的行 ...

组委会 发自 凹非寺 量子位｜公众号 QbitAI 中国生成式AI正在进入产业深水区。 这两年，AI从"新技术"变成了"新工具"，又从"新工具"慢慢变成企业必须面对的现实。它不只在改变内容生产，也在影响研发效率、营销方 式、团队协作，甚至决策流程。 时值第四届中国AIGC产业峰会， 量子位将根据过去一年里生成式AI企业、产品的表现与反馈，结合对2026年技术与场景的观察与预判，评 选出： 量子位将结合对公司的深入调研及数十位行业知名专家的意见，评选结果将于2026年5月中国AIGC产业峰会上公布。 届时，量子位也将邀请数百万行业从业者，共同见证这些优秀企业的荣誉。 2026年度值得关注的AIGC企业 将评选出拥有最创新、最前瞻或最有规模落地潜力的AI企业。 【参选条件】 2026年度值得关注的AIGC企业 2026年度值得关注的AIGC产品 2026年度值得关注的AIGC产品 将评选出拥有最创新、最实用、最热门或最有应用潜力的AI产品。 【参选条件】 【评选维度】 1. 技术维度 ｜关注公司的技术实力、研发能力和创新性，包含技术成果、研发投入、人才储备等角度； 2. 产品维度 ｜关注核心产品的创新性、市场适配 ...

Core Viewpoint - The departure of John Nguyen from Meta to join AMI, a company founded by Yann LeCun, highlights the ongoing challenges and internal turmoil at Meta, particularly within its FAIR team, as it struggles with technological advancements and employee retention [1][5][30]. Group 1: John Nguyen's Background and Contributions - John Nguyen, a key figure at Meta's FAIR, has a strong academic background with dual degrees in statistics and computer science from the University of California, Davis, and has been with Meta for over six years [12][15]. - His research trajectory at Meta included significant contributions to federated learning, large-scale deep learning training, and multi-modal systems, aligning with Meta's technological evolution [16][18][20]. - Nguyen's expertise in both foundational training and practical system implementation positions him as a valuable asset in the AI industry, particularly as the focus shifts from language modeling to real-world modeling [20][28]. Group 2: Meta's Current Challenges - Meta is experiencing significant internal challenges, including rumors of leadership changes and difficulties in model development, particularly with the delayed release of its new model "Avocado," originally expected by late last year [30][34]. - The company has faced public relations issues, including a recent incident involving unauthorized data leaks, contributing to a negative perception of its operational stability [36][37]. - The contrast between Meta's struggles and the rapid growth of AMI, which secured $1.03 billion in seed funding, suggests a potential trend of further departures from Meta's FAIR team to join LeCun's new venture [28][38].

Core Viewpoint - The article discusses the departure of Guo Dayan, a key engineer at DeepSeek, who has significantly contributed to various models including V2, V3, and R1, raising concerns about the potential impact on DeepSeek's future developments [1][6][7]. Group 1: Guo Dayan's Background and Achievements - Guo Dayan is recognized as a technical prodigy with a remarkable academic and competitive history, often referred to as the "Lei Jun of Sun Yat-sen University" [2][42]. - He completed his doctoral thesis requirements just three days after starting his postdoctoral studies, showcasing exceptional research efficiency [3][35]. - Guo has won multiple championships in competitions such as the Tencent Advertising Algorithm Competition and the ATEC Technology Elite Competition, earning substantial monetary rewards [4][44][46]. Group 2: Contributions to DeepSeek - Guo Dayan joined DeepSeek after completing his PhD in 2023, focusing on code intelligence and large language model inference [8][10]. - He was a core contributor to several models, including DeepSeek-Coder, DeepSeek-Math, and DeepSeek-Prover, which have shown significant advancements in mathematical reasoning and formal proof generation [13][18][21]. - The training cost for the DeepSeek-R1 model was approximately $294,000, indicating a relatively low investment for the capabilities achieved [25]. Group 3: Future Implications - Guo's departure raises questions about the continuity of DeepSeek's innovative projects, particularly the development of the upcoming DeepSeek-V4 model [6][10]. - His contributions have been pivotal in demonstrating that large models can achieve reasoning capabilities without relying on human annotations, which could influence future AI model development strategies [24].