本篇内容转载自「锦秋集」 semianalysis写了一篇文章，通过深入分析DeepSeek和Anthropic两家公司的策略选择，揭示了一 个行业共同面临的根本挑战：计算资源的稀缺。 DeepSeek R1发布128天后的数据呈现出一个看似矛盾的现象：官方平台用户流失，但第三方托管 的模型使用量却暴增20倍。为什么用户会抛弃价格极低的官方服务，转而选择第三方平台？ 本文通过Token经济学这一分析框架找到了答案。 文章指出，AI服务的定价本质上是三个性能指标的权衡游戏。 第一是延迟，即用户发送请求到收到第一个字符的等待时间； 第二是吞吐量，即模型每秒能生成多少个token，直接影响对话的流畅度； 第三是上下文窗口，决定了模型能"记住"多少对话历史，对于分析长文档或大型代码库至关重要。 关键洞察在于：通过调整这三个参数，服务商可以实现任何价格水平。 以下为原文内容。 原文： https://semianalysis.com/2025/07/03/deepseek-debrief-128-days-later/ 超 9000 人的「AI 产品市集」社群！不错过每一款有价值的 AI 应用。 邀请从业者、开发人员和 ...

Core Viewpoint - CloudWalk Technology Co., Ltd. (stock code: 9678.HK) has successfully gone public, becoming the first company in the Hong Kong stock market to focus on General Artificial Intelligence (AGI) as its main business, with a market capitalization exceeding HKD 23 billion in its first week of trading [1][2]. Financial Performance - The company issued 1.561 million shares at an IPO price of HKD 205 per share, raising a net amount of approximately HKD 206 million [3]. - In its first week, the stock price peaked at HKD 338.6, closing at HKD 329.4, representing a 60.6% increase from the issue price, with a total market capitalization of HKD 233.7 billion [2][3]. Technological Investment - Despite rapid expansion, the company maintains a high level of investment in R&D, with projected R&D expenses of HKD 280 million in 2024, accounting for over 30% of revenue [5]. - The company's computing power has reached 184 PFLOPS, enabling real-time training and inference of large models with hundreds of billions of parameters [5]. Commercial Strategy - The company employs a "lighthouse customer" strategy, deeply binding with industry leaders such as Gree, Ping An Technology, and Xiamen Metro, increasing its customer base in lifestyle scenarios to 411 and covering over 500 medical institutions [7]. - The average revenue per project has increased by 38% year-on-year, reflecting a strategic shift towards high-value customers and improved project quality [7]. - The revenue share from end-users has decreased from 52.7% in 2022 to 44.9% in 2024, while the share from system integrators/agents has risen to 55.1%, indicating a flexible market expansion strategy through ecosystem collaboration [7]. Future Plans - The company plans to use the net proceeds from the IPO for four main areas: investment in core technology R&D such as Atlas AI infrastructure, expansion into vertical industry applications like smart healthcare and transportation, acceleration of international expansion focusing on emerging markets in Southeast Asia and the Middle East, and strategic partnerships along with operational funding [7].

Core Viewpoint - The release of Grok 4 by xAI marks a significant advancement in AI capabilities, with claims of achieving postdoctoral-level proficiency across multiple disciplines, potentially leading to groundbreaking scientific discoveries within the year [2][8]. Group 1: Product Details - Grok 4 is available in two subscription versions: Grok 4 at $30/month and Grok 4 Heavy at $300/month, with the latter's annual fee exceeding 20,000 RMB [4][5]. - Grok 4 Heavy scored 44.4% in the Human Last Exam (HLE), outperforming the previous top model, Gemini 2.5 Pro, which scored 26.9% [5][8]. Group 2: Performance and Testing - Grok 4 excelled in the HLE test, which spans 100 disciplines and includes 2,500 doctoral-level questions, indicating a significant breakthrough in complex knowledge systems and deep thinking capabilities [8]. - The model has achieved top scores in various prestigious tests, including HMMT, USAMO, and GPQA, and received a perfect score in the AIME25 [13][14]. Group 3: Technological Advancements - The training volume from Grok 2 to Grok 4 increased by 100 times, with enhanced training efficiency through data selection and algorithm optimization [9]. - Grok 4's reasoning ability improved by 10 times compared to its predecessor, aided by the use of the world's top supercomputing clusters and increased reinforcement learning investments [9]. Group 4: Future Developments - xAI plans to release additional models, including a coding model in August, a multi-model agent in September, and a video generation model in October, focusing on enhancing visual capabilities [19][20].

尽管性能领先，马斯克坦言Grok 4仍存在"致命缺陷"——图像理解能力不足。他解释称，当前版本的 Grok在处理抽象图像、复杂场景和动态视觉信息时，准确率较文本任务低40%以上，例如难以识别手绘 草图中的隐喻含义，或区分照片中微小物体的功能属性。 "这就像一个天才数学家看不懂漫画书。"马斯克比喻道，"我们需要为Grok植入更强大的'视觉大 脑'。"据xAI技术团队透露，改进方案包括升级多模态大模型架构、引入3D空间感知算法，并扩充包含 10亿张高分辨率图像的专用训练集。马斯克承诺，新版本将于"一个月内"完成部署，届时Grok将实现文 本、图像、视频的跨模态无缝交互。 2025年7月10日，智通财经电——在今日举行的xAI直播活动中，特斯拉与SpaceX创始人埃隆·马斯克公 开表示，其旗下人工智能公司xAI开发的聊天机器人Grok已在"大多数领域展现出超越人类的智能水 平"，但当前版本的Grok 4仍存在图像理解盲区，需通过改进后的图像训练模块优化，预计问题将在一 个月内解决。 马斯克在直播中直言："Grok 4在逻辑推理、多模态交互和复杂任务处理上已显著优于普通人类，甚至 在部分专业领域接近顶尖专家水平。"他 ...

萤石蓝海大模型持续进化，在2025年6月迎来了全新升级的2.0版本，聚焦技术能力与场景应用的双重落 地，通过多维融合、模态扩展与专项记忆三大技术路径，实现感知、理解、记忆能力的三重增强，同时 助力垂直领域，做到专项能力、定向优化、精准服务，解决用户的痛点和难点。 近日，萤石自研的"萤石蓝海大模型"正式获得由知名权威咨询机构CIC灼识咨询颁发的"家居行业首个具 身智能大模型"市场地位确认证书。 根据中国计算机协会，具身智能大模型指一种基于物理身体进行感知和行动的智能系统，通过智能体与 环境交互获取信息、理解问题、做出决策并实现行动，产生智能行为和适应性。家居行业具身智能大模 型定义为应用于家居行业且获得完整国家备案的人工智能大模型。 萤石蓝海大模型自2024年6月首次亮相便确立了面向物联网场景的具身智能方向，并于9月完成生成式人 工智能服务安全备案。其命名"蓝海"寓意深刻——正如地球生命起源于浩瀚的蓝色海洋，萤石期待在复 杂多变的家居物联海洋中，孕育出全新的智能形态，最终奔向通用人工智能（AGI）的广阔未来。本次 获得的"家居行业首个具身智能大模型"市场地位确认，正是对其践行这一理念的阶段性肯定。 在通用大模型 ...

来源：环球网 【环球网科技综合报道】7月9日，据外媒报道，在AI行业人才争夺战持续升温之际，OpenAI于近日完 成一次关键性人才收割——成功从特斯拉、xAI和Meta挖来四位顶尖工程师及研究员，直指其核心战略 领域：构建支撑通用人工智能（AGI）的超级基础设施。这一动作被视为对Meta近期大规模挖角行动的 反制，同时可能进一步激化与马斯克的矛盾。 根据OpenAI联合创始人Greg Brockman在内部的公告，新成员包括：David Lau（前特斯拉软件工程副总 裁），主导自动驾驶系统开发；Uday Ruddarraju（前xAI/X基础设施负责人），曾主导构建全球最大AI 超算Colossus；超算系统优化专家Mike Dalton（前xAI基础设施工程师）；多模态大模型训练专家 Angela Fan（前Meta AI研究员）。 外媒指出，此次挖角被视为对Meta近期行动的回应。据悉，Meta CEO扎克伯格今年已从OpenAI挖走至 少7名核心研究员，开出"天价薪酬+无限算力"条件。OpenAI CEO Sam Altman上周向员工透露，将调整 研究人员薪资结构以应对竞争。 "今天的AI竞赛，本质是 ...

Core Viewpoint - Yunzhisheng Intelligent Technology Co., Ltd. has successfully listed on the Hong Kong Stock Exchange, marking a significant milestone for Chinese AGI technology companies and injecting confidence into the industry [1][4]. Group 1: Stock Market Performance - On the listing day, Yunzhisheng's market value surged over 40%, reaching HKD 21 billion, and in the first week, it accumulated a total increase of 60.6%, surpassing HKD 23 billion in market capitalization [1][3]. - The global offering of H-shares amounted to approximately 1.561 million shares, representing about 2.2% of the total share capital post-listing, with net proceeds of approximately HKD 206 million (around RMB 189 million) [3]. Group 2: Technological Advancements - Yunzhisheng has successfully implemented its end-side intent understanding model in a well-known domestic automotive company, marking the first application of this model in cockpit scenarios [4]. - The technology allows vehicles to establish a seamless natural language interaction system even in offline environments, enhancing the AI assistant's capabilities [4]. - The end-side intent understanding model is based on the self-developed Shanhai model, which has been optimized to reduce its parameters to 0.5 billion, significantly lowering hardware requirements for inference [4]. Group 3: Industry Significance - As the first stock in Hong Kong focused on general artificial intelligence, Yunzhisheng's stock performance, technological iteration capabilities, and international expansion will serve as an important case study for observing the transition of Chinese AI companies from "technological breakthroughs" to "value realization" [6].

以下文章来源于CSDN ，作者唐小引 CSDN . 成就一亿技术人 作者 | 唐小引 出品 | CSDN（ID：CSDNnews） 1997 年，AI 正处于第二次寒冬，这次寒潮的时间有点长，从 20 世纪 90 年代直至 21 世纪的第一个十年。 寒冬之下，有人坚持，有人幸运，也有人不是那么地有运气。 李飞飞说："我觉得自己是世界上最幸运的人，因为我的整个职业生涯始于 AI 寒冬的末期，也就是 AI 开始腾飞的时候，所以我真的非常幸运，也有点 自豪。" 而在现任 清华大学基础科学讲席教授、 心理与认知科学系主任、人工智能学院教授、 北京智源人工智能研究院首席科学家 刘嘉 的世界里，则不一 样。 1997 年，对人工智能深感困惑的青年刘嘉远赴美国麻省理工学院 （ MIT ） 深造，慕名拜访 人工智能之父马文·明斯基 （ Marvin Minsky ）， 想要从这位鼎鼎大名的 AI 教父身上寻求良方。 明斯基很 Nice，完全没有任何架子，让刘嘉前往 MIT 人工智能实验室面谈。但面谈的结果却让刘嘉备 受打击，" 我记得当时人工智能实验室在 MIT 是一栋很高的楼，我 上去的时候是仰视 ，终于到了那种殿堂级的地 ...

Core Insights - Yunzhisheng has officially listed on the Hong Kong Stock Exchange, becoming the first general artificial intelligence (AGI) company to do so, with an initial share price of HKD 205 and a first-day closing price of HKD 296.4, representing a 44.59% increase [1] - The company's stock price reached a high of HKD 338.6 within the first week, closing at HKD 329.4, marking a 60.6% increase from the initial price, with a market capitalization exceeding HKD 233.7 billion [1] Company Background - Founded in 2012, Yunzhisheng has focused on technology as its core driver, being one of the first to apply deep learning algorithms in commercial voice recognition [2] - The company has received investments from notable institutions such as Lenovo Ventures and has developed a strategic focus on the medical and IoT sectors since 2014 [2] Financial Performance - According to the prospectus, Yunzhisheng's revenue is projected to grow from CNY 601 million in 2022 to CNY 939 million in 2024, with a compound annual growth rate (CAGR) exceeding 25% [5] - The company ranks as the fourth largest AI solution provider in China, holding a market share of 0.6%, with strong performances in both consumer and medical AI solutions [5] Technology and Commercialization - Yunzhisheng has established a commercial closed-loop model that integrates technology research and development, scene implementation, and data feedback [6] - The company has launched several products, including the UniOne AI voice chip in 2018 and the "Shanhai" large model in 2023, adapting to the needs of the generative AI era [6][8] Market Strategy - The company's strategy of aligning technology with market demands has allowed it to build unique advantages in niche areas, enhancing user experience and operational efficiency in consumer applications [8] - In the medical sector, its diagnostic assistance systems are implemented in over 800 hospitals nationwide, improving diagnostic efficiency [8]

机器之心报道 机器之心编辑部 反击从这里开始？ 最近 Meta 挖 AI 人才成了科技圈最大的瓜。有人找到了山姆・奥特曼，问他有关扎克伯格和 Meta 挖人的看 法。奥特曼说：还行吧。 实际上表面的尴尬之下，激烈的斗争早已开始。 本周二《连线》杂志获悉，OpenAI 从竞争对手公司挖来了 四位 备受瞩目的工程师加入其扩展团队，其中 包括特斯拉前软件工程副总裁 David Lau。 这一消息来自 OpenAI 联合创始人 Greg Brockman 本周二在内部 Slack 上的通知，他也是扩展团队的负责 人。 加入 OpenAI 的还包括： xAI 和 X 公司的前基础设施工程负责人 Uday Ruddarraju； xAI 的基础设施工程师 Mike Dalton Meta 的 AI 研究员 Angela Fan。 Dalton 和 Ruddarraju 此前还在 Robinhood 共事过。在 xAI 期间，Ruddarraju 和 Dalton 曾共同参与构建 Colossus —— 一台由超过 20 万块 GPU 组成的大型超级计算机。 OpenAI 发言人 Hannah Wong 表示：「我们非常高 ...