智能体技术日益发展，但现有的许多通用智能体仍然高度依赖于人工预定义好的工具库和工作流，这极大限制了其创造力、可扩展性与泛化能力。 近期，普林斯顿大学 AI Lab 推出了 Alita ——一个秉持「 极简即是极致复杂 」哲学的通用智能体，通过「 最小化预定义 」与「 最大化自我进化 」的设 计范式，让智能体可以自主思考、搜索和创造其所需要的 MCP 工具。 Alita 目前已在 GAIA validation 基准测试中取得 75.15% pass@1 和 87.27% pass@3 的成绩，一举超越 OpenAI Deep Research 和 Manus 等知名智 能体，成为通用智能体新标杆。Alita 在 GAIA test 上也达到了 72.43% pass@1 的成绩。 极简架构设计，最大自我进化 「让智能体自主创造 MCP 工具而不靠人工预设」，是 Alita 的核心设计理念。 现有的主流智能体系统通常依赖大量人工预定义的工具和复杂的工作流，这种方法有三个关键缺陷： 覆盖范围有限 ： 通用智能体面临的现实任务种类繁多，预先定义好所有可能需要的工具既不可行亦不现实。而且预定义工具很容易过拟合 GAI ...

Core Viewpoint - The article discusses the significant upgrades in the Fellou browser, particularly the transition to version 2.0, which aims to create a more integrated and efficient AI assistant akin to Jarvis from the Marvel universe, enhancing user experience and task execution capabilities [4][5][6]. Group 1: Why Agentic Browser? - The Agentic Browser is designed to understand user needs and automate complex tasks, fundamentally changing how users interact with the internet and computers [8]. - Fellou's unique architecture combines Browser, Workflow, and Agent components, allowing it to function like an "autonomous surfing" browser [8]. - The goal is to free users from repetitive tasks, enabling them to focus on more fulfilling work while Fellou handles mundane tasks [9][11]. Group 2: Fellou 2.0 Features - The upgrade to Fellou 2.0 has resulted in a speed increase of 1.2 to 1.5 times compared to version 1.x, with significant improvements in task execution speed [13][14]. - The success rate for task completion has risen dramatically from 31% to 80%, showcasing enhanced reliability and performance [14][29]. - Fellou can now execute multiple tasks simultaneously, improving user productivity and efficiency [20][23]. Group 3: Key to Success - Eko 2.0 - Eko 2.0 is a crucial open-source infrastructure that has contributed to the improved task success rate, providing essential capabilities for browser and computer use [34][35]. - The framework supports multi-agent collaboration and task management, enhancing the overall functionality of Fellou [35]. Group 4: Future Plans for Fellou - Upcoming features include a Windows version, removal of the invitation system, and enhancements in model intelligence for richer deliverables [36]. - Continuous optimization of user experience is planned, focusing on speed, interaction quality, and additional functionalities [36].

一水 闻乐 发自 凹非寺 量子位 | 公众号 QbitAI 无邀请码，就可直接上手！ 北大校友官宣推出号称"最强通用Agent" Fairies （中译仙女） ，能执行Deep research、代码生成、发邮件等 1000个操作 。 编辑部的小伙伴一上手实测就发出了如下感叹~ 关键是无需邀请码，Mac和Windows用户只需下载APP就能立即上手试玩。(⊙ˍ⊙) 一打开Fairies，它支持自由选择GPT 4.1、Gemini 2.5 Pro、Claude 4在内的多款模型，还都是各家最新款的那种。 官方已经展现了一些很实用的玩法。比如帮我推荐一台最适合工作的Mac电脑，需要考虑到便携性、能连接多个显示器、能带动视频创建和编 辑等需求。 没过多久，Fairies不但在对话框用本文详细推荐了某款产品，还在右侧清晰展示了产品对比图。 选购哪个产品简直一目了然。 那么，号称"最强通用Agent"的Fairies真实能力究竟如何呢？ 量子位实测走起。 已经能看到未来智能体形态 再比如帮助团队安排一个合适的会议时间。 只需要给出成员的日程表和会议时长，Fairies就能自动做出最合理的安排，并且把会议通知发送给每位成 ...

Investment Dynamics - Manus' parent company, Butterfly Effect, is reportedly planning a new financing round of $100 million at a valuation of $1.5 billion, with state-owned enterprises participating. The funds will primarily be used to develop the Chinese market [3] - DTC snack brand Farmley successfully raised $40 million in Series C funding, led by L Catterton, with existing investors also participating. The funds will help expand its presence in the Indian health snack market [5] - AI and robotics service provider "Shouhua Technology" completed a Series A financing round of several tens of millions of RMB, led by a fund under Hangzhou Wen Guang Group. The funds will be used for AI model development, hardware upgrades, and global market expansion [6][7] Acquisition Dynamics - Spanish second-hand clothing platform Percentil was acquired by Israeli tech company MySize, avoiding bankruptcy. The acquisition includes the Percentil brand, central warehouse, AI pricing engine, quality assessment system, and over 120,000 items of inventory [11] - Dick's Sporting Goods is nearing a deal to acquire Foot Locker for an estimated price of $24 per share, totaling $2.3 billion. This news caused Foot Locker's stock to surge nearly 70% [14][15] - Borletti Group announced the acquisition of a minority stake in True Religion, which is known for its iconic "Super T" stitching. The financial details of the transaction were not disclosed [18] - Church & Dwight announced plans to acquire DTC hand sanitizer brand Touchland for $880 million, with $700 million in cash and stock, and an additional $180 million contingent on sales targets [22] - A consortium of investors has made a €60 million acquisition offer for French sportswear brand Le Coq Sportif, with Neopar set to hold 51% of the shares [25][26] Personnel Dynamics - Microsoft announced a company-wide layoff of 6,000 employees, representing less than 3% of its total workforce of 228,000, as part of a strategy to streamline management levels [29] - Balenciaga appointed Nathalie Raynaud as Vice CEO to strengthen its executive team in preparation for the arrival of a new creative director [31]

Core Viewpoint - Beijing has introduced the "Action Plan for Supporting Information Software Enterprises to Enhance AI Application Service Capabilities (2025)", which includes multiple financial incentive policies, with a maximum reward of 30 million yuan [1]. Group 1: Financial Incentive Policies - The "soft eight measures" include six main financial policies: computing power vouchers, model "first plan", software intelligent transformation projects, data vouchers, shared open-source project rewards, and small and medium-sized enterprise service vouchers, all of which are "achievable upon meeting standards" [1]. - The computing power voucher policy supports two main areas: subsidies for computing power deployment costs for MaaS platforms and support for general intelligent agent operations, with a maximum subsidy of 30 million yuan [2]. Group 2: Software Intelligent Transformation Projects - The software intelligent transformation project policy supports two aspects: enhancing software development efficiency through AI applications and upgrading software products' intelligence levels [3]. - For software development, it encourages the use of computing power, large model deployment, and data governance to transform development methods and improve efficiency, with a maximum reward of 30 million yuan for individual enterprises [3]. Group 3: Data Voucher Policy - The data voucher policy encourages enterprises to open AI model training datasets to the public, with support based on the scale, quality, update frequency, and application effectiveness of the datasets, providing up to 500,000 yuan for individual enterprises or institutions [3]. Group 4: Implementation and Future Plans - The application details for the "soft eight measures" are currently being refined, with plans to publicly release them in the second quarter as part of the Beijing High-Precision Industry Development Project Fund Implementation Guidelines (second batch) [3].

4 月 18-19 日，由 CSDN&Boolan 联合举办的 2025 全球机器学习技术大会（ML-Summit 2025）将在上海虹桥西郊庄园丽笙大酒店隆重举行。大会 云集院士、顶尖学者、IEEE Fellow、一线科技企业技术实战专家组成的超 50 位重磅嘉宾。他们将以独特的视角，解读智能体、联邦学习、多模态大 模型、强化学习等覆盖 AI 当下热门的技术实践专题。 在大会首日下午的「AI 智能体」专题论坛上，来自 DeepWisdom 算法研究 员、OpenManus 项目一作，MetaGPT 开源核心贡献者梁新兵将带来 《通用 Agent 的构建与赋能：OpenManus 的实践与探索》分享。 梁新兵是 DeepWisdom 算法研究员，华东师范大学硕士。他不仅是 OpenManus 项目一作，同时也是论文 Data Interpreter /Self-Supervised Prompt Optimization 作者之一。如今，他正以其在智能体领域的丰富经验和 对开源的满腔热情，不断探索通用 Agent 的构建与赋能。 3 小时复刻传奇：OpenManus 背后的极速行动 作为 MetaGPT 开源 ...

Core Viewpoint - The article discusses the rapid evolution and potential explosion of AI applications, particularly focusing on the emergence of general AI agents and the competitive landscape among major players in the AI industry, with a specific emphasis on the anticipated developments in 2025 [2][3][5]. Group 1: AI Application Landscape - The AI application market is expected to explode in 2025, with significant opportunities in software for large companies and in hardware for startups [3]. - The introduction of DeepSeek and other AI models has sparked enthusiasm for AI applications, with many industry insiders believing that 2025 will be a pivotal year for AI application deployment [2][3]. - The launch of products like Manus and the upgrade of Alibaba's Quark to a flagship AI application signify a shift towards general AI agents that can meet diverse user needs [2][3][17]. Group 2: Competitive Dynamics - The competition among large AI models will continue, as companies must keep pace with advancements to retain users [23]. - The emergence of DeepSeek has challenged the dominance of established players like OpenAI, indicating a shift in the competitive landscape [12][23]. - The article highlights that future AI companies will likely integrate both model development and application deployment, blurring the lines between traditional model providers and application developers [24]. Group 3: User Experience and Market Trends - User experience remains crucial in the AI landscape, with companies needing to create compelling applications to attract and retain users [15][19]. - The article suggests that the AI market will evolve similarly to the mobile internet era, where user engagement and feedback will drive product development and innovation [25][27]. - The potential for AI to revolutionize productivity tools is emphasized, with predictions that AI will significantly enhance coding and other tasks in the near future [21][22]. Group 4: Hardware and Robotics - The article discusses the challenges of human-like robots, suggesting that while software may advance rapidly, hardware improvements will be more gradual [4][29]. - There is a belief that hardware startups may have a better chance against larger companies due to longer development cycles for hardware products [29]. - The potential for smart hardware, such as Meta's smart glasses, is highlighted as a significant opportunity in the AI landscape [28].

Summary of Manus Conference Call Company Overview - Manus is a general-purpose AI agent designed to execute tasks via cloud virtual machines, optimized for specific domains, particularly in text processing, with continuous learning and memory capabilities [2][3]. Core Industry Insights - Manus utilizes a multi-agent framework for task planning and execution, covering seven areas including travel planning, stock analysis, and financial report analysis, although it experienced a 16-hour system outage [2][3][8]. - The product's performance in generating a seven-day travel guide for Japan took 49 seconds, compared to ChatGPT's 5 seconds, highlighting differences in execution speed and user experience [2][11]. - The concept of "Model is a product" gained traction with the launch of DeepSeek, indicating a shift towards single models with simple UIs being competitive in the market [2][21]. Key Features and Innovations - Manus can adapt to user behavior for cognitive adjustments, such as automatically generating reports in specific formats, enhancing user experience [5]. - Compared to earlier platforms, Manus simplifies user interaction by eliminating complex operations, making it more user-friendly [6]. - The system's architecture is based on HuggingFace and IOM large models, which provide good support for specific scenarios but may lack flexibility [12]. Performance and Limitations - Despite its innovative features, Manus has faced challenges, including system outages and incomplete functionalities, which have not dampened user enthusiasm [8][9]. - In negative testing, Manus struggled with general scenarios, such as modifying documents or understanding vague user intentions, indicating limitations in its capabilities [10]. Competitive Landscape - Currently, there are no direct competitors to Manus, but platforms from OpenAI, OpenDNA, and Microsoft are exploring similar functionalities [25]. - The ease of replicating Manus by large internet companies is noted, as they may integrate similar functionalities into their models rather than relying on external modules [31]. Market Trends and Future Outlook - The debate over "true" versus "false" AI agents continues, with concerns about cost efficiency and token usage in automated workflows [4][24]. - The potential for AI agents in specialized fields like healthcare and law is limited due to high barriers and reliance on expert knowledge [19]. - The future of AI agents in consumer markets is uncertain, as achieving profitability remains a challenge due to the need for generalization and the inability to meet specific user needs effectively [33][34]. Conclusion - Manus represents a significant step in the evolution of AI agents, with its innovative features and multi-agent framework. However, it faces challenges in execution speed, system reliability, and competition from established players in the market. The ongoing discussions about the nature of AI agents and their applications in various industries will shape the future landscape of this technology.