机器人的大脑：从 LLM 到世界模型 20250811 摘要 具身智能模型架构主要分为三类：完全端到端大模型（如特斯拉 FSD、 谷歌 RT、英伟达 Grok01），多模态融合大模型（VLM、VOA），以 及加入触觉传感器的数据融合大模型（VL-T）。完全端到端模型依赖海 量数据和强大算力，多模态融合模型集成文本、图像和动作信息，触觉 传感器模型提升操作精度。 数据在具身智能中至关重要，数据量决定场景泛化能力，多模态数据 （包括物理量信息如动作和触觉数据）使机器人能执行实际操作。数据 需求和多模态数据处理对算力提出极高要求，限制了高级系统的开发。 企业采用人形机器人模型面临的主要挑战包括数据需求量大、算法复杂 度高以及系统之间的连接和解耦问题。分层端到端模型通过创新架构解 决系统连接和解耦问题，提高任务执行效率。 分层端到端大模型被广泛应用于人形机器人领域，如 Finger AI 的 Helix 模型，通过将感知、决策和运动控制分层实现智能化操作，提高了任务 执行效率。 Q&A 什么是具身智能模型，它与非具身智能模型有何区别？ 具身智能模型与非具身智能模型的主要区别在于输入和输出模态的不同。非具 身智能模型， ...

谷歌（Google）前首席商务官莫·乔达特（Mo Gawdat）表示，"人工智能将创造就业岗位"的说法纯属"彻头彻尾的废 话"。他警告称，人工智能将摧毁就业岗位，甚至"能力欠佳的首席执行官"和"心怀不轨"的政府领导人也难以幸免。图 片来源：Kate Green / Stringer / Getty Images 科技巨头们一直坚称，人工智能将引领人类进入"黄金时代"，届时所有疾病都能被治愈，人们尽享富足 生活，劳动者将拥有"超级人类"能力。然而，前谷歌高管却对"人工智能不会成为就业杀手，反而会为 人类创造新工作"的观点展开了猛烈抨击。 尽管"公司完全由机器人运营、没有人类参与"的场景极具反乌托邦色彩，但这位前谷歌高管并不畏惧未 来。这位58岁的高管认为，人工智能并非导致失业的罪魁祸首——实际上，是那些贪婪的首席执行官们 为追求经济利益而任由技术接管工作，这才是问题的根源所在。 "在我看来，这种说法简直就是彻头彻尾的废话，"谷歌X前首席商务官莫·乔达特最近在《首席执行官日 记》播客中表示，"无论身处何种岗位，顶尖人才将会留下来。真正了解架构、掌握技术的顶尖软件开 发者会留下来——至少在一段时期内是这样。" 乔达 ...

整理 | 美股研究社 在这个快速变 化的时代，人工 智能技术正以前所未有的速度发展，带来了广泛的机会 。 《AI日 报 》致力于挖掘和分析最新的AI概念股公司和市场趋势，为您提供深度的行 业 洞察和 价 值 分 析。 A I 快 报 1 . C3.ai美股盘前跌近30%，公司将其营收展望削减33%，重组全球销售团队。 2.美国银行月度基金经理调查显示，得益于乐观的盈利表现以及全球经济情绪的改善，投资者重 新涌入股市，持有美国大型科技股再度成为最热门交易。 8月的调查涵盖169位参与者，他们管 理的资产规模达4130亿美元。其中45%的受访者认为，最拥挤的交易是"做多'七大科技巨 头'"——这一组合包括英伟达和微软等美国大型科技股。 七 巨 头 日 报 【标准普尔 500 指数的近 50% 权重集中在 20 只股票上】 阿波罗首席经济学家托尔斯滕·斯洛克 (Torsten Sløk) 表示，标准普尔 500 指数对少数几只超大盘股的集 中度已达到历史最高水平。 阿波罗最新数据显示，英伟达（NVDA）目前在该基准指数中的个股权重为1981年有记录以来的最高水 平，凸显了市场对人工智能（AI）领军企业的高度关注。这 ...

Group 1 - Alphabet (NASDAQ: GOOG) (NASDAQ: GOOGL) stock has faced challenges this year due to concerns about losing market share to OpenAI in the search sector [1] - The company is currently involved in an ongoing antitrust case with the Department of Justice (DOJ) [1] - A volatile market environment has also contributed to the difficulties faced by the company [1]

美国公司正以史无前例的速度回购股票，推动股价上涨并创造新的市场纪录。 据Birinyi Associates数据，美股公司今年迄今已宣布9836亿美元的股票回购计划，创下自1982年有记录 以来的最佳开局，预计全年回购总额将超过1.1万亿美元。 本轮回购浪潮的领军者主要是大型科技公司和银行。苹果公司、谷歌母公司Alphabet，以及摩根大通、 美国银行和摩根士丹利等金融巨头都在积极回购。 推动这一趋势的核心动力源于企业充裕的现金流。强劲的盈利增长为企业带来了丰厚利润，而围绕贸易 政策的持续不确定性，使得许多公司暂停了新的投资计划，这使得股票回购成为运用现金的更具吸引力 选择。 分析师认为，这一买入热潮有望推动股指继续上涨，为当前的股市涨势提供动力。目前，标普500指数 中91%已公布二季度业绩的公司中，约有82%超出预期。 科技与金融巨头引领回购浪潮 此次股票回购表现出明显的头部集中效应，排名前20位的公司占据了近一半的回购总额。 大型科技公司是今年回购授权规模最大的群体，AI热潮推高其市值的同的同时，也让它们的资产负债 表迅速膨胀。 今年5月，苹果公司承诺将投入高达1000亿美元用于回购自家股票。根据其7月 ...

在科技界，埃隆・马斯克和拉里・佩奇都是绕不开的名字。前者创办了 PayPal、特斯拉、SpaceX等企 业，后者是谷歌的联合创始人。这两位颇具影响力的企业家，在成功路上有着相似的思维特质 —— 都 极度依赖物理学原理分析问题。 从具体实践来看，物理学原理是他们破解难题的钥匙。马斯克思考火箭制造时，先从力学原理分析动力 与阻力，再用能量守恒原理推演，最终让 SpaceX 实现了火箭回收，将卫星发射成本降至原来的十分之 一以下。 有人将这种思维模式总结为 "十倍进化法则"：先锁定目标对象，再基于物理、数学、化学原理开发新 产品，以十倍优势取而代之。马斯克瞄准美国航天局的高成本发射模式，佩奇则聚焦杂乱的网络信息， 前者实现了发射成本的大幅降低，后者让信息检索效率和丰富度提升十倍，各自在领域内开辟了新天 地。这种成功的思想可以被复制，供企业家们参考。 马斯克对基础学科的重视，在他 2016 年创办的秘密学校 Ad Astra 上体现得淋漓尽致。这所位于 SpaceX 办公楼内的学校，招收了包括马斯克 6 个儿子在内的公司高管子女，课程表上只有数学、物理学、化 学、工程学、伦理学五门课。在他看来，数理化知识是解决复杂问 ...

Investment Rating - The report does not explicitly state an investment rating for the industry [2] Core Insights - The TMT sector experienced varied performance in the week of August 4-8, with Media up by 2.30%, Electronics by 1.65%, Communication by 1.30%, and Computers down by 0.41% [3] - The release of Google's Genie3 marks a significant breakthrough in AI simulation, featuring real-time interaction and long-term consistency, which is expected to benefit game development and film production in the short term, while providing high-quality training data for AI agents and robots in the long term [3] - OpenAI's GPT-5 has been launched with improved cost-effectiveness, showing enhanced accuracy in various tests compared to its predecessor, and is expected to accelerate AI penetration in programming and healthcare applications [3] - Anthropic's Claude Opus 4.1 has been released, showcasing upgrades in programming capabilities and solidifying its market leadership in AI programming with a projected 42% market share by mid-2025 [3] Weekly Market Review - The TMT sector's performance for the week included notable gains in sub-industries such as Brand Consumer Electronics (up 6.81%), Gaming III (up 5.96%), and Communication Cables and Accessories (up 4.43%), while sectors like Horizontal General Software and Printed Circuit Boards saw declines of -3.86% and -2.59% respectively [3][5] - The report includes detailed weekly performance metrics for various TMT sub-industries, highlighting the fluctuations in stock prices and market trends [7][10] Industry News and Key Company Announcements - Significant industry events include the 2025 ChinaJoy and the release of various AI models and applications by major companies, indicating a vibrant landscape for AI development and deployment [15][17] - Key company announcements include TSMC's sales growth of 25.8% year-on-year and the performance updates from major players like SMIC and Huahong Semiconductor, reflecting the ongoing advancements and competitive dynamics in the semiconductor industry [24][25]

Investment Rating - The industry investment rating is "Positive" and is maintained [3][4]. Core Insights - Recent breakthroughs in AI models have been significant, with Google DeepMind releasing the new general world model Genie 3 and OpenAI launching its latest AI model GPT-5, indicating a competitive landscape among leading players [6]. - The commercialization of AI applications is accelerating, particularly in vertical sectors such as gaming, advertising, e-commerce, education, and IP [6]. Summary by Sections AI Model Developments - Google DeepMind's Genie 3 can generate dynamic interactive worlds at a speed of 20-24 frames per second in 720p resolution, supporting real-time interaction and maintaining content coherence for several minutes [6]. - OpenAI's GPT-5 has shown significant improvements in various benchmarks, outperforming previous models in programming, mathematics, writing, health, and visual perception, with a 45% reduction in factual errors compared to GPT-4 and an expanded context capability of 400k tokens [6]. AI Content Creation and Commercialization - Tencent's ima has introduced new features, including an AI podcast function that converts articles into dialogue format, enhancing user experience [6]. - Kuaishou's AI project "New World Loading" has achieved over 1.97 billion views, showcasing the potential of AI in film and content creation [6]. - Google's Gemini AI chatbot has added a "Storybook" feature that generates illustrated storybooks based on user prompts, highlighting advancements in AI-driven content generation [6]. AI Gaming Sector - The upcoming AI game "Whispers from the Star," led by Mihayou's co-founder, is set to launch on August 15, featuring natural language interaction and a rich narrative experience [6]. - The report suggests focusing on AI interactive products, particularly AI toys, which are expected to commercialize quickly due to enhanced user experiences and broader audience reach [6]. Investment Opportunities - The report highlights the potential of AI interaction products, the advantages of large tech companies in data and model development, and the replicable business models from overseas in advertising, e-commerce, and education sectors [6].

Summary of Key Points from Conference Call Industry Overview - The conference call primarily discusses the AI computing power industry, focusing on advancements in AI models, particularly OpenAI's GPT-5, and the competitive landscape among major tech companies in North America such as Google and Meta [1][7][8]. Core Insights and Arguments - **GPT-5 Enhancements**: GPT-5 has not revolutionized application paradigms but has significantly improved efficiency, pricing, and context length. The API pricing is set at $1.25 per million tokens for input and $10 for output, making it cost-effective compared to other AI models [1][4]. - **Market Demand for Computing Power**: There is a strong and persistent demand for computing power in the market, with expectations of increased capital expenditures from major cloud providers. The total capital expenditure for North America's top four cloud companies is projected to reach $366.1 billion in 2025, a year-on-year increase of approximately 47% [3][23]. - **Domestic Market Trends**: Despite uncertainties in capital expenditures, domestic internet giants are expected to continue investing in computing power, with a notable increase in investments in the domestic patent industry chain anticipated in the second half of the year [8][10][9]. - **Technological Innovations**: Domestic computing power manufacturers are focusing on innovations in interconnect technology, superpoint architecture, and large-scale system solutions to support large models comprehensively [11]. - **Performance of U.S. Tech Hardware Companies**: U.S. tech hardware companies reported second-quarter results that generally met expectations, driven by surging demand for coding and agent-related inference and pre-training [12]. Additional Important Insights - **Competitive Landscape**: Major tech companies like Google and Meta are heavily investing in model updates and optimizations, with Google potentially leading in frontier models. Meta is also increasing its investments to enhance its capabilities [7]. - **Emerging Technologies**: The call highlights the competitive dynamics among interconnect technologies, including PCIe and Ethernet interconnects, with Broadcom's advancements in PCIe 6.0 and Ethernet Scale-Up technology being particularly noteworthy [2][13]. - **Future Projections**: The demand for AI GPUs and ASICs is expected to rise significantly in 2026 and 2027, with recommendations for core industry players that exhibit high elasticity and clear performance delivery [18]. - **Arista's Performance**: Arista's stock surged over 10% following its second-quarter earnings report, attributed to better-than-expected performance and an optimistic revenue outlook for AI [19][20]. Conclusion - The AI computing power industry is experiencing robust growth driven by advancements in AI models, significant investments from major tech companies, and a strong demand for computing power. The competitive landscape is evolving with new technologies and strategies, indicating a promising outlook for the sector.

Core Viewpoint - Linus Torvalds publicly criticized a Google engineer's RISC-V code submission for Linux 6.17, labeling it as "garbage" due to its poor quality and late submission [2][3][5]. Group 1: Criticism of Code Quality - Torvalds emphasized that the newly submitted RISC-V code contained unnecessary and poor-quality additions that do not pertain specifically to RISC-V [6][9]. - He provided examples of problematic code, such as the "make_u32_from_two_u16()" helper function, which he argued complicates understanding and makes the code worse [7][9]. - Torvalds insisted that such low-quality submissions should not be sent, especially late in the merge window, and warned developers against future late submissions [10][9]. Group 2: Developer Communication - Torvalds's comments were not merely harsh; he aimed to provide constructive feedback to improve code quality and adherence to submission timelines [3][4]. - He advised developers to submit their pull requests early in the merge window and to ensure that the code is of high quality, stating that "no more garbage" should be submitted [10][9]. - The response from social media regarding Torvalds's outburst was mixed, with some appreciating his directness and others questioning his approach [3].