"AI 原生 100" 是虎嗅科技组推出针对 AI 原生创新栏目，这是本系列的第「 14 」篇文章。 共享算力，这件事放在今天，并不新鲜。早在20世纪80年代就有人尝试，甚至有人用它寻找外星文明，但却没人在共享算力上赚到钱。 一直到AI Agent的大爆发，这一模式第一次获得了商业上的成功。首次将共享算力这事儿，在全球范围内能跑通，并实现盈利，竟让一个00后团队做到了。 2025年7月，在世界人工智能大会上，付智十分忙碌，他是共绩科技的CEO。这一次，他在大会上对接了40多条线索、转化了20多条潜在商机。 2025年，被业界视为"AI Agent"爆发的元年，这直接导致了AI推理的算力需求暴增。根据第三方分析机构IDC的预测，推理服务器的工作负载占比从2020年 的51.5%，增加到了2026年的62.2%。 这些推理的需求，投射到终端，就是一系列AI视频生成公司、AI陪伴机器人公司、AI模型公司的爆发。但这些AI公司的痛点在于，每次都需要支付一大笔 长租服务器的费用给到云厂商，这造成的问题是，当需求较少时，他们需要为算力服务，承担闲置算力资源的成本，当需求增多时，他们又要让用户排队。 这对于成本敏感、精细化 ...

Core Insights - Manus AI's annual recurring revenue (RRR) has reached $90 million and is expected to surpass $100 million soon, with clarification that this figure is based on monthly revenue multiplied by 12 and does not equate to cash revenue [1] - The distinction between AI Agents and AGI (Artificial General Intelligence) is emphasized, with AI Agents being a subset of applied AI that interacts with the environment, while AGI possesses general capabilities to perform various tasks without specific design [3][4] Company Developments - Manus AI's team member highlighted that many AI products offer annual payment options, which can inflate revenue figures as they may represent prepayments rather than actual operating income [1] - The company aims to empower non-programmers by generalizing the use of AI tools like Cursor, which has gained traction among both engineers and non-engineers for tasks such as data visualization and writing [4] Industry Trends - The conversation addressed the challenges AI faces in interacting with the real world, such as the lack of APIs or standard interfaces and the prevalence of CAPTCHAs, which hinder AI's capabilities [4] - Despite current limitations, there is optimism about AI's future, with expectations for significant breakthroughs as the ecosystem evolves and infrastructure companies like Stripe contribute to advancements [4]

Group 1 - The core viewpoint of the article highlights various companies' recent developments and market activities, indicating potential investment opportunities and industry trends [3][5][8][10][12][19][30][32]. Group 2 - Pop Mart has seen a stock price increase of over 7% after founder Wang Ning announced the upcoming release of a mini version of Labubu, projecting an easy revenue target of 30 billion RMB for the year [3]. - China FAW Group is rumored to be planning to acquire approximately 10% of Leap Motor's shares, with both companies remaining non-committal on the matter [5]. - Manus, an AI platform, reported a revenue run rate of 90 million USD, indicating strong growth potential [5]. - Intel's stock has surged by 28% this month, adding approximately 24 billion USD to its market capitalization, reaching its highest valuation since the internet bubble [8]. - Xiaomi's automotive division reported a revenue of 21.3 billion RMB in Q2, a year-on-year increase of 234%, with expectations of profitability in the second half of the year [8]. - NIO has taken legal action against accounts spreading false information about the brand, indicating a proactive approach to brand protection [10]. - ByteDance has denied rumors of collaborating with Chipone on AI chip development, clarifying its strategic direction [10]. - Geely has reassured customers that promised basic data services will remain unchanged, addressing concerns over service modifications [15]. - The Beijing AIGC audiovisual industry innovation center has been launched, aiming to enhance the production capabilities of audiovisual content through AI technology [16]. - Databricks has announced a valuation exceeding 100 billion USD as it seeks funding, reflecting the growing interest in AI data analysis platforms [19].

Important Market News - US stock indices closed mixed, with the Nasdaq down 0.67%, S&P 500 down 0.24%, and Dow Jones up 0.04%. The S&P 500 has declined for four consecutive trading days. Major tech stocks mostly fell, with Intel down about 7% and Google, Amazon, Apple, and Tesla down over 1% [1] - The Federal Reserve's FOMC meeting minutes revealed that the federal funds rate target range remains at 4.25%-4.5%. Committee members noted a slowdown in economic activity growth in the first half of the year and acknowledged persistent inflation and high uncertainty in the economic outlook [1] Industry Insights - The Guangdong Provincial Government has released policies to promote high-quality development in commercial aerospace from 2025 to 2028. This includes support for satellite constellation investments and encouraging the procurement of domestic satellite data and products. The commercial aerospace market in China is projected to grow from 9.2 billion yuan in 2020 to 310 billion yuan by 2024, with a compound annual growth rate exceeding 100% [3] - Zhiyuan Robotics announced its first partner conference in Shanghai on August 21, 2025, showcasing numerous robots and a "mystery new product." The humanoid robot market is expected to see significant advancements, with companies like Yushu Technology and Xiaopeng Motors pushing for performance upgrades and cost reductions [4] - On August 20, Zhipu AI launched AutoGLM 2.0, the world's first fully open "mobile Agent" for consumers. This product is based on the latest open-source models and is seen as a milestone in the AI Agent sector, with potential applications across various industries [5][6] Stock Movements - Nanjing New Hundred plans to reduce its repurchased shares by up to 6 million shares, representing 0.45% of its total share capital [7] - Xagong Co. announced a plan to reduce its shares by up to 17.74 million shares, or 1% of its total share capital [7] - Jin Hong Shun intends to reduce its holdings by up to 5.376 million shares, accounting for 3% of its total share capital [7] - Ying Shi Sheng plans to reduce its shares by up to 15.02 million shares, representing 2% of its total share capital [7] - Yi Jia He announced a plan to reduce its shares by up to 4.11 million shares, or 2% of its total share capital [7] - Puyuan Information's major shareholders plan to reduce their holdings by up to 2.9951% of the total share capital [8] - Dongfang Ocean's major shareholder plans to reduce its holdings by up to 58.74 million shares, or 3% of its total share capital [8]

my most contrarian trait is i find traders buying ai agent tokens instead of tracking almanak's $25m deployment into 40% pendle positions to always be missing the actual trade. ...

Core Viewpoint - The MEM1 framework, developed by MIT and the National University of Singapore, addresses the challenges faced by AI agents in managing complex tasks and memory efficiently, achieving significant improvements in inference speed and memory usage compared to traditional models [2][22]. Group 1: Framework Overview - MEM1 framework allows AI agents to autonomously manage their working memory and reasoning processes, akin to how humans organize thoughts after a period of work [4][10]. - The framework introduces a near constant memory usage model, significantly reducing the computational cost associated with increasing dialogue rounds [6][12]. Group 2: Performance Metrics - The MEM1-7B model demonstrates a 3.5 times faster inference speed compared to a traditional 14B model, while maintaining a peak token count that is approximately one-fourth of the latter [2][3]. - In a complex 16-target task, MEM1 outperformed larger models and those with external memory modules across accuracy, context length, and inference speed [17][18]. Group 3: Training Methodology - MEM1 employs an end-to-end reinforcement learning approach, utilizing an attention masking mechanism that allows the agent to focus on relevant historical information while compressing it efficiently [12][22]. - The training process involves three key operations: extracting key information, integrating it with internal memory, and pruning redundant content [14][20]. Group 4: Practical Applications - The MEM1 framework has been tested in various environments, including document retrieval QA, open-domain web QA, and multi-round online shopping scenarios, showcasing its adaptability and effectiveness in real-world applications [19][20]. Group 5: Industry Implications - The traditional approach in the industry has been to integrate external memory modules, which can be cumbersome and less effective; MEM1's approach suggests a shift towards self-managed memory systems through reinforcement learning [22].

Core Insights - Manus has announced a revenue run rate (RRR) of $90 million, indicating a significant financial milestone for the company [1] - The company has shifted its market positioning strategy, moving from Beijing to Singapore to enhance its global presence and compliance with international regulations [3][5] - The founder has expressed a commitment to establishing Manus as a world-class company, emphasizing the importance of global market integration while maintaining its Chinese roots [5] Financial Performance - The reported revenue run rate (RRR) of $90 million serves as a predictive indicator of Manus's annual revenue based on current monthly income [1] - The company has previously refrained from discussing annual recurring revenue (ARR) but is now more open about its financial metrics to establish a clearer market position [3] Strategic Decisions - The relocation of Manus's headquarters to Singapore is driven by the need for compliance with cross-border data regulations and access to necessary technology [5] - The decision reflects a broader strategy to address challenges in the Chinese market while aiming for a more robust international footprint [3][5] Product Development - Manus has shared insights on its development process for AI agents, including the choice of context engineering over developing proprietary models [4] - The company has faced criticism for not addressing core issues in its public communications, indicating a need for more transparency [4]

Core Viewpoint - The article discusses the career transition of Duolong (Cai Jingxian), a legendary programmer from Alibaba, who has joined the AI startup Beibeilianzhuan to revolutionize operations and maintenance services using AI Agents [1][19]. Group 1: Duolong's Background and Achievements - Duolong, known as "the first programmer of Taobao," has a remarkable history at Alibaba, where he contributed significantly to the development of the Taobao platform and its search engine [3][5]. - Despite not having a formal computer science background, Duolong's technical prowess and problem-solving abilities earned him a reputation as a "god" among his peers at Alibaba [7][8]. - He reached the highest technical position (P11) at Alibaba and was recognized as a partner due to his substantial contributions to Taobao's success [9][11]. Group 2: Transition to Beibeilianzhuan - After leaving Alibaba, Duolong joined his old friend Bi Xuanyuan (also known as "Bi Dashi") at Beibeilianzhuan, a startup focused on AI-driven cloud resource management [13][15]. - Beibeilianzhuan aims to leverage AI Agents to transform the operations and maintenance service sector, addressing the challenges of scaling professional services [17][18]. - The company has secured significant funding, including a 50 million yuan angel round and additional investments for its Pre-A round, indicating strong investor confidence in its vision [14][15]. Group 3: Future Vision and Impact - The collaboration between Duolong and Bi Dashi is seen as a pivotal moment in the AI era, with the potential to enhance service quality and efficiency through AI technology [17][18]. - Beibeilianzhuan's development of the SREAgent aims to provide clients with access to expertise across various fields, effectively creating multiple "Duolong" agents for operational support [18]. - The article concludes with a hopeful outlook on Duolong's future contributions to the tech industry, emphasizing his enduring passion for coding and innovation [19][20].

Core Viewpoint - The article discusses the recent criticisms of the DiT (Diffusion Transformers) model, which is considered a cornerstone in the diffusion model field, highlighting the importance of scientific scrutiny and empirical validation in research [3][10]. Group 1: Criticism of DiT - A user has raised multiple concerns about DiT, claiming it is flawed both mathematically and in its structure, even questioning the presence of Transformer elements in DiT [4][12]. - The criticisms are based on a paper titled "TREAD: Token Routing for Efficient Architecture-agnostic Diffusion Training," which introduces a strategy that allows early-layer tokens to be passed to deeper layers without modifying the architecture or adding parameters [12][14]. - The critic argues that the rapid decrease in FID (Fréchet Inception Distance) during training indicates that DiT's architecture has inherent properties that allow it to easily learn the dataset [15]. - The Tread model reportedly trains 14 times faster than DiT after 400,000 iterations and 37 times faster at its best performance after 7 million iterations, suggesting that significant performance improvements may undermine previous methods [16][17]. - The critic also suggests that if parts of the network are disabled during training, it could render the network ineffective [19]. - It is noted that the more network units in DiT that are replaced with identity mappings during training, the better the model evaluation results [20]. - The architecture of DiT is said to require logarithmic scaling to represent the signal-to-noise ratio differences during the diffusion process, indicating potential issues with output dynamics [23]. - Concerns are raised regarding the Adaptive Layer Normalization method, suggesting that DiT processes conditional inputs through a standard MLP (Multi-Layer Perceptron) without clear Transformer characteristics [25][26]. Group 2: Response from Xie Saining - Xie Saining, the author of DiT, responded to the criticisms, asserting that the Tread model's findings do not invalidate DiT [27]. - He acknowledges the Tread model's contributions but emphasizes that its effectiveness is due to regularization enhancing feature robustness, not because DiT is incorrect [28]. - Saining highlights that Lightning DiT, an upgraded version of DiT, remains a powerful option and should be prioritized when conditions allow [29]. - He also states that there is no evidence to suggest that the post-layer normalization in DiT causes issues [30]. - Saining summarizes improvements made over the past year, focusing on internal representation learning and various methods for enhancing model training [32]. - He mentions that the sd-vae (stochastic depth variational autoencoder) is a significant concern for DiT, particularly regarding its high computational cost for processing images at 256×256 resolution [34].

本文来自微信公众号：APPSO （ID：appsolution），作者：APPSO，题图来自：AI生成 每天睁眼后的第一件事是什么？刷手机。 睡前的最后一件事是什么？还是刷手机。 但你有没有算过，每天要在不同App之间切换多少次？淘宝比价、美团点外卖、小红书找攻略——我们的手机里装着几十个App，却要靠十个手指在它们 之间来回奔波。 这些碎片化的时间往往一天下来，足以让我们开始怀疑人生——时间都去哪儿了。 所以当AI Agent浪潮席卷而来时，我们的第一反应就是希望能有一个真正的手机通用Agent。它应该像一个随身助理，不管你在做什么，都随时响应你的 需求，同时能够在后台默默工作，不打断你正在进行的任何事情。 其实早在Manus刷屏之前，智谱就已经在Agent赛道上埋头苦干了。我们之前测过他们的初代AutoGLM，印象还不错。而就在刚刚，智谱再次升级了 AutoGLM Agent功能。 带着这样的疑问，我们想看看这款Agent能否把"手机自动驾驶"这个概念变成现实。 一句话就能让AI帮你打卡追剧点奶茶，AutoGLM Agent开启手机自动驾驶 0:00 / 0:21 据智谱官方介绍，AutoGLM Agen ...