「停止研究 RL 吧，研究者更应该将精力投入到产品开发中，真正推动人工智能大规模发展的关键技术是互联网，而不是像 Transformer 这样的模型架 构。」 前 OpenAI 研究员 Kevin Lu 最近更新了一篇博客长文《The Only lmportant Technology ls The Internet》，直指互联网才是推动人工智能进步的核心技术， 是 next-token 预测的完美补充。 Kevin Lu 认为，没有 Transformer 架构，我们可能也会拥有 GPT-4.5 级别的大模型。在 GPT-4 模型以来，基础模型的能力并没有显著的提升，我们可能会 像 2015-2020 年时代的 RL 研究一样，重蹈覆辙，正在进行无关紧要的 RL 研究。 而互联网提供了丰富而海量的数据来源，这些数据具有多样性、能提供自然的学习课程、代表了人们真正关心的能力，并且是一种经济上可行的规模化部 署技术。相比之下，单靠优化模型结构、手工制作数据集或微调算法，都难以带来模型能力质的飞跃。 有趣的是，Kevin Lu 此前在 OpenAI 任职时的主要研究方向之一正是 RL。在推特上，有博主评论道，「当 ...

作者 | 悦木Ivy 编辑 | 自动驾驶之心 原文链接： https://www.zhihu.com/question/637090848/answer/3363233578 点击下方 卡片 ，关注" 自动驾驶之心 "公众号 戳我-> 领取 自动驾驶近15个 方向 学习 路线 >>自动驾驶前沿信息获取 → 自动驾驶之心知识星球 本文只做学术分享，如有侵权，联系删文 先看下Chat和Agent的差别。 Chat（聊天）：纯粹的Chat，像是一个主要由"大脑和嘴"构成的智能体，专注于信息处理和语言交流。比如 ChatGPT这样的系统，它能够理解用户的查询，给出有用和连贯的回答，但它本身不直接执行任务。 Agent（代理）：像一个具有"手、脚"的智能体，它能够进行思考、决策，并且能执行具体的任务。 我们可以简单粗暴的理解为， Chat强调的是"说"，Agent强调的是"做"。 要你你选哪个？ 这么讲吧， 但凡人类对于躺平的美好追求还在，"让机器替我干活"的念头就不可能熄灭。 自ChatGPT发布后，从plugin的推出，到Function Calling再到Assistant API的面世，OpenAI这一系列动作 ...

Core Viewpoint - Robinhood Markets is expanding its offerings by introducing tokenized shares of both public and private companies, aiming to attract a broader customer base and enhance trading activity on its platform [2][6][10]. Group 1: Company Growth and Popularity - Robinhood has seen significant growth, with over 25 million funded accounts in 2024, up from 10 million five years ago [1]. - The company has increased its revenue from $1.8 billion in 2021 to just under $3 billion in the past year, indicating a strong financial performance [10]. Group 2: Tokenized Shares Offering - The introduction of tokenized shares allows users to buy and sell contracts that track the price of underlying assets on a blockchain, with minimal costs [4]. - Tokenized shares are currently available in Europe, providing users exposure to popular U.S. stocks, while U.S. regulations limit their availability domestically [4][9]. - Robinhood is also offering tokenized shares of private companies like OpenAI and SpaceX, which raises questions about price tracking and regulatory compliance [5][8]. Group 3: Market Potential and Future Outlook - The move to offer tokenized shares could significantly increase trading activity on Robinhood's platform, especially in Europe [4][6]. - The potential for tokenization in the U.S. is being recognized, with favorable comments from the SEC Chairman, suggesting future opportunities in the crypto space [9]. - Robinhood's stock has increased over 310% in the past 12 months, reflecting strong investor interest, although it is trading at more than 50 times earnings [11][12].

陈白/文 硅谷如今最抢手的是什么？答案可能超出所有人的想象——不是AI项目，也不是芯片算力，而 是华人软件工程师。 比如就在7月10日，在埃隆·马斯克旗下AI公司Grok 4的发布会上，坐在马斯克身旁的就是两位华人AI技 术负责人。更早一些的7月8日，苹果基础模型团队负责人、著名华人工程师庞若鸣（Ruoming Pang）宣 布离职加入Meta，Meta为他开出了每年数千万美元的薪酬方案。在此之前，扎克伯格还以上亿美元 的"签字费"从OpenAI、苹果等大公司挖走了一批华人学者，他们很多都来自清华大学、北京大学等国 内顶尖高校，且具备OpenAI、Google等科技巨头的核心算法经验。其中有人主导过ChatGPT研发，有人 负责自动驾驶感知系统，堪称AI行业的"梦之队"。 马斯克、扎克伯格们的选择变化有技术积累的原因。从Transformer架构的奠基性论文到ResNet（残差网 络）等影响深远的突破，其中都有华人科学家和工程师的身影。国内顶尖高校对基础学科的严苛训练， 以及其毕业生在OpenAI、Google等机构参与核心算法研发的履历，更是一种权威背书。大模型时代对 复杂数学建模、高效算法设计与海量数据（ ...

Core Viewpoint - Meta is heavily investing in artificial intelligence, including a $14.3 billion investment in Scale AI, to build a "superintelligent" team aimed at becoming a leading personalized AI assistant by 2025 [1][4][22] Investment and Financial Performance - Meta's core AI efforts have enhanced the profitability of its application series, justifying its spending plans as operating and free cash flow continue to expand [2][12] - The company has raised its annual capital expenditure guidance from $62.5 billion to $68 billion, which is expected to account for 36.4% of total revenue by FY2025 [12][22] - Despite losses in the Reality Labs division, Meta's overall operating income remains strong, with a family of apps operating income of $21.03 billion and an operating margin of 41% [9][10] Talent Acquisition and Competition - Meta is actively recruiting AI talent from competitors, having hired at least 15 individuals from OpenAI, Google, and Apple, indicating a fierce talent war in Silicon Valley [6][7] - The cost of this talent acquisition is substantial, estimated at $20 billion, which includes the investment in Scale AI and high salaries for new hires [7][22] Future Projections - Analysts have begun to revise revenue and earnings forecasts upward, with expectations of continued mid-to-low double-digit revenue growth, reaching $238.6 billion by FY2027 [18][23] - The consensus earnings per share is projected to grow to $32.76 by FY2027, with a forward PE ratio calculated at 28.57 [20][23] Market Sentiment and Stock Valuation - Meta's stock is currently above its five-year average expected PE ratio but has room to grow, potentially reaching a target price of $832, reflecting a 14% increase from current levels [22][23] - Investor confidence is bolstered by Meta's strong financial discipline, with free cash flow and operating cash continuing to grow despite increased capital expenditures [16][22]

Core Viewpoint - Nvidia has acquired Canadian AI startup CentML for over $400 million, aiming to enhance its AI capabilities and talent pool [2][14]. Group 1: Acquisition Details - Nvidia's acquisition price for CentML is over $400 million, with the base purchase estimated at over $300 million and additional payments tied to performance targets [14]. - CentML was founded by a team led by a 95-born Chinese PhD, focusing on reducing AI training costs through software solutions [3][4]. - The acquisition includes 19 employees, including four co-founders, who will join Nvidia [12]. Group 2: CentML's Technology and Impact - CentML's core technology, Hidet, acts as a "translator" between AI models and chips, potentially increasing inference speed by up to 8 times and significantly lowering computational costs [7][10]. - The technology is expected to be integrated into Nvidia's TensorRT inference platform, enhancing its performance in edge computing and autonomous driving [11]. Group 3: Funding and Valuation - CentML has completed multiple funding rounds, including a $3.5 million Pre-seed round and a $27 million seed round, achieving a post-money valuation of $300 million [8][9]. - Prior to the acquisition, CentML had raised approximately $30.9 million in venture capital, indicating a strong return for its investors [15]. Group 4: Talent Acquisition in AI Industry - Nvidia's recent acquisitions, including CentML, are part of a broader strategy to secure top AI talent, with the company reportedly acquiring more firms in 2024 than in the previous four years combined [22]. - The competition for AI talent is intensifying, with other tech giants like Meta also actively recruiting top researchers from leading institutions [27][28].

值得注意的是，在发布会上同马斯克一起分享的两位xAI的核心创始成员，都是华人。这 也并非他们首次与马斯克同台。 早在今年2月的发布会上，他们就已经以核心成员的身份 亮相 。 记者丨董静怡 孔海丽 编辑丨骆一帆 当地时间7月9日，马斯克的xAI公司发布最新旗舰模型Grok 4，直接跳过Grok 3.5版本，以 25.6万tokens的上下文窗口、10倍于前代的推理能力，以及多模态支持能力震撼业界。 坐在中间位置的吴宇怀，是出生于杭州建德的"95后"，今年刚满30岁，却已在AI领域崭露头 角 。 他 在 多 伦 多 大 学 攻 读 机 器 学 习 博 士 ， 师 从 " 深 度 学 习 之 父 " 杰 弗 里 · 辛 顿 （ Geoffrey Hinton）。博士期间曾在Google DeepMind和OpenAI实习，毕业后加入Google研究院，并在 斯坦福大学从事博士后研究。 2023年xAI成立时，他作为联合创始人加入，成为团队中最年 轻的核心成员之一 。 坐在右侧的Jimmy Ba，不仅是xAI的创始成员，也是吴宇怀的导师之一。他同样毕业于多伦 多大学，其2014年与Diederik P. Kingma ...

Core Insights - Kevin Kelly, known as K.K., discusses his new book "2049: The Possibilities of the Next 10,000 Days," which envisions a hopeful future 25 years from now, covering various aspects such as technology, society, and the role of China in the world [4][5][7][8]. Group 1: Future Scenarios and Predictions - K.K. emphasizes that the book is not about making predictions but constructing scenarios that explore optimistic futures, including advancements in AI and technology [10][11]. - The book aims to inspire readers to actively imagine and work towards a desirable future rather than passively waiting for it to happen [12][13]. - K.K. believes that technology is a double-edged sword, and while it can bring positive changes, it also poses challenges that need to be addressed [7][15]. Group 2: Education and Workforce - K.K. argues that education systems must adapt to a future where lifelong learning and AI assistance are the norms, emphasizing the importance of teaching students how to learn effectively [18][19]. - He suggests that the focus should be on developing the ability to ask good questions and optimize learning processes rather than acquiring specific knowledge or skills [19][21]. - K.K. reassures that AI will not replace jobs but will change the nature of work, encouraging graduates to pursue their passions rather than solely focusing on financial rewards [22][24]. Group 3: Media and Content Creation - K.K. asserts that the media industry will still require human journalists to manage AI-generated content and ensure accountability [26][30]. - He predicts that new forms of content creation will emerge, including immersive experiences that enhance the way news is reported [27][28]. - K.K. advises content creators to remain adaptable and willing to experiment with new formats and technologies to stay relevant in the evolving landscape [30][31]. Group 4: Technological Advancements - K.K. identifies five key technologies that will shape the future: robotics, autonomous driving, space exploration, life sciences, and brain-computer interfaces [49]. - He expresses skepticism about the widespread adoption of humanoid robots in households, suggesting they will first be used in controlled environments like warehouses [51][55]. - K.K. believes that while autonomous vehicles will become more prevalent, the need for human drivers will still exist, albeit in a more specialized capacity [57][58]. Group 5: The "Mirror World" Concept - K.K. introduces the idea of a "mirror world," where augmented reality will blend the physical and digital realms, facilitated by smart glasses [32][34]. - He outlines four layers of this mirror world, including virtual reality experiences, spatial digital avatars, and enhanced environments for AI interactions [35][36][39]. - K.K. predicts that the development of this mirror world will depend on the affordability and efficiency of AI technology [37][40]. Group 6: Future of Smart Glasses - K.K. anticipates that the competition in the smart glasses market will be fierce, with potential for Chinese companies to emerge as leaders due to their manufacturing advantages [42][46]. - He highlights the significant technical challenges that need to be overcome for smart glasses to become mainstream, including battery life and user experience [43][45]. - K.K. believes that the timeline for achieving widely accepted smart glasses could extend to 10-25 years [46]. Group 7: Societal Implications of AI and Longevity - K.K. discusses the potential for AI to contribute to longer, healthier lives, while also acknowledging the societal challenges posed by an aging population [70][71]. - He emphasizes the need for innovative solutions to address the demographic shifts, particularly in cultures with a high proportion of elderly individuals [72][73]. - K.K. expresses cautious optimism about advancements in life sciences, suggesting that breakthroughs in health technology could lead to significant improvements in quality of life [68][69].

搜索引擎曾是 PC互联网时代的流量王者，百度凭此稳坐了多年的中文互联网"头把交椅"。即便腾讯、360等追赶者推出多款桌面浏览器产品，也难以动摇 其地位。 智能手机的普及催生了移动互联网时代。超级APP们割裂出一个个数据孤岛，让搜索引擎的高效只能局限于自家地盘。阿里巴巴旗下的夸克再一次发起在 搜索和浏览器赛道的冲锋，且取得了不小战果。 如今随着AI时代的加速降临，百度引以为豪的搜索框与腾讯、阿里后发的浏览器，正在被打破边界融合在一起。一个名为Agentic Browser（智能体浏览 器）的新物种悄然诞生，一场新的流量战争也在酝酿之中。 在今年5月底的谷歌I/O大会上，号称重构了谷歌传统搜索逻辑的AI Mode备受瞩目。它基于先进的推理和多模态分析能力，不再仅仅专注于回答问题，而 是为搜索带来了全新的智能体验。 无独有偶，前不久的百度AI DAY活动中，百度搜索也官宣了十年来最大的改版升级，过去的搜索框变大为"智能框"，可以满足超千字的长文本搜索、文 件搜索、AI工具调用、语音搜等多模态、更丰富的搜索需求。 与此同时，腾讯的QQ浏览器、阿里的夸克浏览器，都在生成式大语言模型的技术加持下，不断进行AI化的改造和迭 ...

AI前沿每周关键词Top50 （ 0707-0711 ） 扫码 加入AGI数据库，AI智能问答 腾讯研究院 ： 《算法破茧｜腾讯研究院三万字报告》 | 类别 | Top关键词 | 主体 | | --- | --- | --- | | 模型 | Grok4 | 马斯克 | | 模型 | Phi-4新版本 | 微软 | | 模型 | 开放权重模型 | OpenAI | | 模型 | SmolLM3 | Hugging Face | | 模型 | Skywork-R1V 3.0 | 昆仑万维 | | 模型 | BlueLM-2.5-3B | vivo | | 模型 | DeepSeek-R1外挂 | 上海交大 | | 模型 | HumanOmniV2 | 阿里 | | 模型 | Skywork-Reward-V2 | 昆仑万维 | | 模型 | DeepSeek增强版 | 德国TNG公司 | | 模型 | Sekai数据集 | 上海AILab | | 应用 | AI浏览器Comet | Perplexity | | 应用 | MedGemma 27B | 谷歌 | | 应用 | 生肖企鹅AI共创 | 腾讯 | | ...