Core Insights - The article emphasizes the importance of context engineering in developing AI agents, highlighting the need for rapid iteration and improvement in response to evolving models and technologies [1][2]. Group 1: KV Cache Design - KV cache hit rate is identified as the most critical metric for AI agents in production, directly impacting latency and cost [4]. - The average input to output token ratio in Manus is approximately 100:1, which significantly benefits from KV caching, reducing the cost of cached input tokens to $0.30 per MTok compared to $3 per MTok for uncached tokens [5]. - Key practices to improve KV cache hit rate include maintaining stable prompt prefixes, appending content only, and marking cache breakpoints explicitly [8][9][10]. Group 2: Tool Management - As agents develop more capabilities, the complexity of the action space increases, leading to potential inefficiencies if tools are dynamically added or removed during iterations [11][14]. - Manus employs a context-aware state machine to manage tool availability without removing tools, thus preventing confusion and maintaining KV cache integrity [14][15][16]. Group 3: Context as a File System - The article discusses the limitations of context windows in modern large language models, suggesting that a file system can serve as an infinite context, allowing agents to read and write files as structured external memory [21]. - Manus implements a recoverable compression strategy, retaining essential information like URLs while allowing for context length reduction [24]. Group 4: Attention Manipulation - Manus uses a "todo.md" file to keep track of tasks, which helps maintain focus and avoid losing sight of goals during complex tasks [26][30]. - Retaining errors in the context is proposed as a method to improve agent behavior, allowing the model to learn from mistakes and reduce the likelihood of repeating them [32][35]. Group 5: Sample Diversity - The article warns against the pitfalls of few-shot prompting in agent systems, which can lead to repetitive and suboptimal actions [36]. - Introducing structured variations in actions and observations can help break patterns and adjust the model's attention, enhancing overall performance [37][38]. Group 6: Conclusion - Context engineering is deemed essential for AI agents, influencing their speed, recovery capabilities, and scalability [39]. - The future of agents will focus on constructing context effectively, underscoring the importance of thoughtful design [40].

Core Viewpoint - Manus has withdrawn from the Chinese market to focus on international expansion, citing operational efficiency adjustments and a shift in strategy towards context engineering for product iteration [1]. Summary by Sections Technical Insights - Manus will emphasize context engineering, leveraging memory and processes for rapid product iteration, focusing on improving training efficiency rather than training new models [1][3]. - The importance of long context (Lossless Long Context) in AI-native products is highlighted, as it enhances personalized interactions and utilizes user interaction history effectively [2]. Lessons Learned - The founder reflects on past experiences with Peak Labs, where the decision to develop a proprietary model became irrelevant after the emergence of advanced models like OpenAI's GPT-3, underscoring the significance of context learning [3]. - Manus has opted to utilize open-source foundational models for training end-to-end agents, avoiding the pitfalls of developing a base model from scratch [3]. Market Challenges - Despite the strategic shift, Manus faces limitations compared to OpenAI's ChatGPT Agent, which benefits from proprietary model advantages and end-to-end training for complex tasks [4]. - The competitive landscape is challenging, with the agent market experiencing significant homogenization and unclear business models, necessitating continuous optimization and exploration of differentiated strategies for Manus [4].

Group 1 - Manus experienced rapid growth and controversy within four months, transitioning from a successful startup to facing significant public scrutiny [1][4][6] - The company raised $75 million in Series B funding led by Benchmark, achieving a valuation of $500 million, which generated high expectations from the market [5] - Controversies arose in late June, including unannounced layoffs, mass deletion of posts by the founding team, and the company's relocation to Singapore, leading to public outcry [6][7] Group 2 - Co-founder Ji Yichao addressed the controversies through a lengthy blog post, focusing on the product and technology rather than the company's issues [3][8] - Manus chose to focus on context engineering instead of developing an end-to-end model, learning from past experiences with large models like GPT-3 [8][12] - Key insights from the blog include the importance of KV cache hit rate, managing tool availability without dynamic changes, and treating the file system as an external memory [8][9][10][34] Group 3 - The company emphasizes the need to retain error information in the context to help the model learn from mistakes, which is crucial for improving agent behavior [11][50] - Manus aims to avoid being limited by few examples by introducing structured variations in actions and observations, which helps break patterns and adjust model attention [52][54] - The conclusion highlights that context engineering is vital for agent systems, influencing their speed, recovery ability, and scalability [56]

Core Insights - Manus has withdrawn from the Chinese market and is focusing on international expansion, citing operational efficiency adjustments and internationalization strategies as the main reasons for this shift [2] - The co-founder of Manus, Ji Yichao, emphasized the importance of context engineering in their technology strategy, aiming to enhance product iteration speed by leveraging memory and process construction [2][4] - The company has learned from past experiences, particularly from their previous venture, Peak Labs, and has decided to avoid investing in foundational model development, instead opting to utilize open-source models for training [5] Context Engineering - Context in large models refers to the information set that models reference when processing tasks or generating outputs, which enhances understanding and performance [3] - The concept of Lossless Long Context is crucial for AI-native products, as it allows for personalized interactions by effectively utilizing user interaction history [3] - The Key-Value Cache (KV-Cache) hit rate is vital for improving inference efficiency and optimizing resource utilization, thereby reducing computational costs [3] Lessons Learned - Ji Yichao reflected on the lessons learned from Peak Labs, where the decision to develop a model from scratch became irrelevant after the emergence of advanced models like OpenAI's GPT-3 [4] - The Manus team has undergone multiple adjustments to their Agent framework to achieve a locally optimal solution, recognizing the challenges of relying on external models for task execution [5] - Despite the focus on efficiency, Manus faces limitations compared to competitors like OpenAI, which utilize proprietary models for better handling of complex tasks [5] Market Challenges - As Manus shifts to the international market, it faces competition from larger platforms that attract developers and users, posing a threat to market share for startups [5] - The current landscape for Agent products is characterized by significant homogenization, unclear business models, and high costs, making it challenging for startups to differentiate themselves [5] - Continuous optimization of technical strategies and exploration of differentiated development paths are essential for Manus to navigate these market challenges [5]

Core Viewpoint - The article discusses the choice between end-to-end training and context engineering in developing general AI agents, highlighting the latter as a more adaptable approach in a rapidly evolving landscape of large models [1][3]. Group 1: Context Engineering Insights - Manus AI's decision to adopt context engineering was influenced by past experiences where self-trained models quickly became obsolete after the release of GPT-3, emphasizing the need for flexibility in model development [4][5]. - The article outlines six core practices derived from Manus's experience, which significantly reduced product iteration cycles from weeks to hours, showcasing an effective technical path for startups [2][3]. Group 2: Key Practices for KV-Cache Optimization - The KV-cache hit rate is identified as the most critical metric for AI agents in production, directly affecting latency and cost, with a notable example showing a 10x cost difference between cached and uncached tokens [7][8]. - Strategies to enhance KV-cache hit rates include maintaining stable prompt prefixes, using only appended context, and employing file systems as external memory to overcome context limitations [8][19]. Group 3: Managing Tool Complexity - The article advises against dynamically adding or removing tools in the agent's action space, suggesting instead to manage tool availability through context-aware masking of token logits to maintain stability [12][13]. - This approach helps prevent confusion in the model when previous actions reference tools that are no longer defined, thereby reducing the risk of erroneous actions [12][17]. Group 4: Utilizing External Memory - Manus employs a file system as an externalized memory solution to address the limitations of context windows, allowing for persistent and unlimited storage that can be directly manipulated by the agent [18][22]. - This method mitigates the risks associated with irreversible context compression, ensuring that critical information is not lost [22]. Group 5: Attention Manipulation Techniques - The use of a todo.md file to continuously update task goals serves as a mechanism to keep the model focused on its objectives, preventing it from losing track during complex tasks [23][26]. - This technique helps maintain the model's attention on the task at hand, especially in lengthy interactions requiring multiple tool calls [26]. Group 6: Learning from Errors - Retaining failed attempts in the context is emphasized as a crucial learning mechanism, allowing the model to adapt and reduce the likelihood of repeating mistakes [30][31]. - The article argues that error recovery is a significant indicator of an agent's performance, yet it is often underrepresented in academic benchmarks [30]. Group 7: Avoiding Few-Shot Traps - The article warns against the pitfalls of few-shot learning in agent systems, where repetitive patterns in context can lead to suboptimal decision-making [32][34]. - Introducing structured variability in actions and observations can help break these patterns and enhance the model's adaptability [34]. Conclusion - Context engineering is presented as an essential and emerging science for agent systems, with the design of context playing a pivotal role in defining agent behavior, speed, recovery, and scalability [35].

在爆火仅四个月后，Manus AI 突然几乎全面撤出中国市场，不仅清空全部社交账号内容，而且国行版本的 Manus 也疑似暂停推进。 早在上个月，Manus 联合创始人张涛便曾宣布，公司已将全球总部迁至新加坡，并在东京和加州设有办公室。尽管官方未正面回应，只称是「基于经营效 率的调整」，但出海所引发裁员等一连串争议问题，也让外界普遍猜测其是否正在「跑路」。 风波之中，今天凌晨，Manus 联合创始人季逸超发布了一篇技术博客，试图将外界关注点重新拉回产品技术本身。 经过四次重构和数百万真实交互，他在文中坦诚地总结了团队在构建 Manus 过程中积累的经验教训。内容既有实操干货，也不乏反思，对业内同行与普 通用户来说，都不失为一份值得一读的参考材料。 1. 押注上下文，不再训练模型 与其耗时训练，不如围绕大模型构造「记忆」和流程。上下文工程让你在几小时而不是几周内发布产品更新。 2. KV-Cache 命中率至关重要 输入越稳定，缓存命中率越高，成本和延迟越低。三条实战建议： - 避免提示中使用时间戳； - 只追加上下文，避免修改历 史记录； - 手动标记缓存断点，保障前缀一致性。 3. 工具不要动态添加，而是用 ...

Manus 官网昨天更新了一篇文章，分享了他们为 Manus 搭建合适的上下文工程的经验教训。 作者季逸超 (Peak)，Manus 公司联合创始人、首席科学家。 文章基于 Kimi K2 翻译，我们进行了一些调整。 在 Manus 项目伊始，我和团队就面临一个关键抉择：是利用开源基础模型训练一个端到端的智能体，还是依托前沿模型的上下文学习能力，在其之上 构建智能体？ 在我投身 NLP 的第一个十年里，我们并没有这种奢侈的选择。遥想当年 BERT 问世（没错，那已是七年前），模型必须先经过微调——还要评估—— 才能迁移到新任务。每次迭代往往耗时数周，尽管那时的模型体积与今日的 LLMs 相比微不足道。对于快速迭代的应用，尤其是 PMF 之前的阶段，如 此缓慢的反馈循环几乎是致命的。这是我上一家初创公司留下的惨痛教训：当时我从零开始训练模型，用于开放信息抽取和语义搜索。随后 GPT-3 与 Flan-T5 横空出世，我那些自研模型一夜之间便失去了意义。颇具讽刺意味的是，正是这些新模型开启了上下文学习的大门——也为我们指明了一条全 新的道路。 这个来之不易的教训让选择变得清晰：Manus 将押注于上下文工程。这让 ...

Core Insights - The article discusses the limitations of large language models (LLMs) in handling complex enterprise-level programming tasks, highlighting the "hallucination" problem where AI generates inaccurate or irrelevant code outputs [1] - A study by METR revealed that using AI programming assistants did not improve efficiency but instead increased development time by an average of 19%, due to high costs associated with reviewing and debugging AI-generated content [1] - ByteDance has introduced ABCoder, a tool designed to address these challenges by providing a clear and unambiguous code "worldview" through deep parsing of abstract syntax trees (AST), enhancing the model's contextual understanding [2] Group 1 - The hallucination problem in LLMs leads to inaccurate code generation, particularly in complex systems [1] - The METR study involved 16 experienced engineers completing 246 programming tasks, showing a 19% increase in development time when using AI tools [1] - ABCoder aims to improve the reliability of AI programming by enriching the model's context acquisition capabilities, thus reducing hallucinations and enabling more accurate code generation [2] Group 2 - ABCoder's implementation will be explained in a live session, showcasing its real-world applications in backend development [3] - The live session will feature a case study on the CloudWeGo project, demonstrating how ABCoder enhances code development efficiency and optimizes the programming experience [3] - ABCoder functions as a powerful toolbox for developers, offering tools for code understanding and conversion to tackle complex programming challenges [3]

Core Trends - The "Chaos AI Business Practical National Tour" has successfully commenced, aiming to ignite practical applications of AI across 20 innovative cities in China, with events already held in Changsha and Nanchang [1][2] - The AI application landscape is evolving with lower entry barriers due to open-source models and contextual engineering, enabling disruptive innovations that empower ordinary individuals [2] - AI penetration in vertical industries is increasing, particularly in pharmaceuticals, digital healthcare, and live service sectors, indicating potential transformative changes [2] AI Applications - Feishu has launched a comprehensive upgrade of its AI product matrix, including knowledge Q&A and AI meetings, along with the industry's first AI application maturity standard to facilitate enterprise AI adoption [3][4] - Google DeepMind's spinoff, Isomorphic Labs, is set to begin human trials for its AI-assisted cancer drug, marking a significant milestone in the pharmaceutical industry [12][13] Investment and Financing - Star Sea Map has raised over $100 million in its A4/A5 financing rounds, with a total pre-A and A round financing of nearly 1.5 billion yuan, reflecting strong capital interest in the embodied intelligence sector [6][7] - TARS, founded by former Huawei employees, completed a record $122 million angel round financing, showcasing investor confidence in embodied intelligence technologies [13] - Cloud Deep has secured nearly 500 million yuan in financing, positioning itself as a leader in the quadruped robot field with over 600 industry projects [14] - Star Motion Era has raised nearly 500 million yuan in its A round financing, emphasizing breakthroughs in humanoid robot technology and significant global demand [16] Business Cases - Wix's acquisition of AI startup Base44 for $80 million highlights the trend of AI enabling entrepreneurship, with Base44 allowing users to generate full-stack application code through natural language [7][8] - The AI personal finance assistant, Kapi Accounting, has gained over one million users in six months, indicating a shift in personal finance management through AI [21][22] Market Insights - The digital human market in China is projected to reach 30 billion yuan by 2025, with significant cost reductions in enterprise live streaming [19][20] - The rise of "contextual engineering" in Silicon Valley is reshaping AI model development, enhancing efficiency and application quality [18][20] Technology Developments - Baidu has open-sourced ten major models, significantly lowering the barriers for AI development and enhancing multi-modal capabilities [21] - The introduction of the Star Stream Agent, designed for Chinese designers, aims to revolutionize the design industry with automated processes and multi-modal content creation [24]

Core Insights - The AI field continues to be an exciting area in 2025, with numerous research releases from major tech companies and institutions [1] - The rapid pace of technological advancements in AI is overwhelming, with new models and paradigms emerging almost weekly [3][4] - Developers and researchers are increasingly engaging in conferences and academic sharing to stay updated on cutting-edge research [5] Event Overview - The ACL conference, a significant event in the NLP field, received over 8,000 submissions this year, marking a historical high [6] - The ACL 2025 conference will take place from July 27 to August 1 in Vienna, Austria, featuring various activities such as keynote speeches, paper presentations, roundtable discussions, and poster sessions [6][7] - The event aims to provide a platform for domestic AI talent, with a full schedule of presentations and discussions announced [6] Keynote Speakers and Topics - The keynote address on "Trends and Outlook for ACL 2025" will be delivered by Che Wanxiang, a prominent professor from Harbin Institute of Technology [9][17] - Liu Pengfei from Shanghai Jiao Tong University will present on "Reinforcement Learning and Complex Reasoning in Large Models" [11][19] Paper Presentations - Various papers will be presented, covering topics such as the intrinsic self-correction of large language models and the acceleration of inference in large language models [9][12] - The event will also feature poster sessions and opportunities for industry engagement [21]