🌐 This transcript was automatically translated to English from the original.
Good morning, this is the AI circle. Let’s learn about AI with a group of people. I’m Yun Ye. Did you know? A company’s revenue increased 80 times year-on-year, but it said it has not yet reached our expectations. This company is called Anthropic. Do you know what they just did to the world of programmers? Do you know? OpenAI shut down the robot team and said that AGI can be realized faster without robots. Five years later, they are back. This time they want to build a personal robot for everyone. Did you know? A model worth 550 billion parameters has been open sourced. The inference speed is five times faster than similar models, and the cost is 30% lower. The market value of the company that released it has just dropped by one Tencent. These things may seem unrelated, but when strung together, you will find that AI is changing from talking to doing. The words robot, giant intelligence, and physical AI are moving from the laboratory to the press conference, to the financing road, and to your life. The depth of AI at the end of the program We will continue to talk about why Nvidia, OpenAI, and Anthropic are all suddenly pushing in the same direction. How will this coming physical AI war reshape the entire industry? There is more AI information for us to pay attention to. Today is June 2, Tuesday. Let’s take a look at today’s AI timeline. Today’s AI timeline, we start from a company that has raised 13 billion US dollars but is still saying that its growth rate has not met expectations. In 2021, Anthropic is just a small team spun off from OpenAI. The founders were worried about the security risks of AI and decided to do it themselves. At that time, their opponent was OpenAI with a market value of nearly US$1 trillion. It sounded like ants versus elephants. Today, 5 years later, Anthropic CEO Dario revealed that in the first quarter of 2026, the company's revenue and usage increased 80 times year-on-year. It was not 8%, but 80 times. This number did not meet internal expectations, which was originally planned to grow faster. According to the adjusted pricing, as of early April, Anthropic's annualized sales had reached 30 billion US dollars. What concept? 相当于腾讯或阿里一个季度的收入 但更有意思的是他们发布的Cloud Code更新 一整套让AI自己干活的框架 Auto模式让Cloud自己决定要不要调用工具 Worktrees让他自己建分支,自己销毁分支 Routines让他按计划自动执行任务 换句话说,不是你在操作AI,而是AI在替你上班 为什么?这很重要 因为就在同一天,Github说他们希望缓存命中率达到94%以上 Verso用更智能的模型简化系统设计 Bong展示了能自动复现bug并创建修复请求的机器人 这意味着程序员世界里会用AI工具和不会用AI工具的差距 正在从效率差距变成存在性差距 不是工具升级,这是工作方式的范式转移 而Anthropic正在抢占定义者的位置 除了让AI自己干活这个大主题 今天还有这些事值得重点关注 新闻一,阿里云签下欧足联6年独家合作 阿里巴巴与欧足联签署6年协议 阿里云成为欧冠、欧联、欧协联及2028年欧洲杯的官方独家AI 云计算及电商合作伙伴 欧足联出了名的挑剔 多年拒绝过多家科技巨头提案 阿里能拿下,靠两件事 一是扛得住压力,体育直播不能当机 洪峰承载能力要求极高 二是读得懂比赛 AI要在数秒内解析复杂赛事规则并给出准确答案 底气来自七年奥运营运会服务经验 东京奥运首次大规模云转播 北京冬奥核心系统百分之百上云 巴黎奥运会云上转播规模首次超过卫星 拿下欧足联合同 意味着阿里完成了云全球化能力 与签问AI商业化能力的双重验证 新闻二,英伟达开源,5500亿参数 Nemotion 3 Ultra推理速度比同类快5倍 英伟达发布Nemotron 3 Ultra 5500亿参数混合专家架构 主要帮智能体更快速完成复杂任务 英伟达说 相比同级别开源模型 推理速度最快快5倍 使用成本能降低30% 模型已可在OpenClaw,OpenHands等平台直接部署 Crowstrike用它自动检测漏洞 Palantir用它支持AI平台自主工作 英伟达说6月4日发布 届时在Hugging Face,Modelscope,OpenRouter和Build.NVIDIA.com都能找到 推理快5倍,成本降30% 意味着企业部署智能体的门槛正在快速下降 新闻3 简制机器人完成数亿元多轮融资 聚身智能无本体数据赛道最大规模 简制机器人宣布完成连续多轮融资 总金额达数亿元 蚂蚁集团,滴滴,德联,资本联合领头 这是聚身智能无本体数据领域迄迄今为止最大规模的融资 简制的方案是从模型定义数据标准 用高保真、多模态、人类行为数据作为核心解决方案 公司自主研发了Ego认知中枢产品系列 包括仿生双指、灵巧无指、工业甲爪等多种数据采集硬件 累计订单突破1万台 首部追踪精度稳定在1厘米 6D姿态感知达亚毫米级 数据产现覆盖3000多名采集用户和1万多处真实场景 累计沉淀超百万小时真实场景数据 本轮融资将重点投入核心技术研发和全链路闭环构建 接下来是今天的AI资讯速递 在近日举办的Code with Cloud会议上 Anthropic发布了一系列重要更新 公司CEO Dario透露 以年化口径计算 2026年第一季度收入和使用量增长了80倍 远超原计划的10倍规模 截至4月初 公司调整定价模式后 年化销售额已达到300亿美元 会议重点展示了Cloud Code的新功能 远程控制功能 让绘画可以在不同设备间无缝切换 桌面界面加入了分栏式图和自动生成目录等特性 在自主能力方面 AUTO模式通过分类器管理权限决策 Worktreets允许Cloud自行创建和销毁隔离分支 Routines则支持按计划或触发条件执行任务 多位合作伙伴分享了时间案例 Github通过缓存优化提升效率 目标是保持94%以上的命中率 Verso使用更智能的模型简化系统设计 OPUS虽然只占约20%的使用量 却贡献超过70%的支出成本 Bong展示了能够自动复现问题 并创建修复请求的机器人能力 Anthropic联合创始人强调 开发者是Cloud最重要的用户群体 公司致力于在强大能力与安全护栏之间保持平衡 关于未来发展 Alex Albert提到Cloud在SYBench测试中的成绩 已从一年前的62%提升至现在的87% 展示了能力的快速进步 OpenAI正在重新启动机器人团队 公司首席执行官Sam Altman表示 他希望将来每个人都能拥有自己的个人机器人 这个团队正在招聘硬件、运营、系统和机器学习方面的工程师 短期内 机器人将帮助专业人员建造基础设施 长期来看 OpenAI想象 每个人都能拥有一个可以完成任何需求的个人机器人 机器人团队源于Aditya Remisch领导的World Simulation研究项目 该项目在Sora视频应用被关闭后 也吸收了Sora团队 OpenAI在2020年关闭了机器人部门 当时认为不借助机器人可以更快地实现通用人工智能 而且机器人训练数据太少了 从2025年1月开始 这个团队被重新组建 目标是开发通用机器人来推动通用人工智能的发展 目前还不清楚OpenAI这样做的具体目的 特别是公司最近宣布了向人工智能代理应用转型的战略 Ultman设定的目标可能还需要很多年才能实现 真正的目的可能是获取训练数据和利用巨身人工智能模型 来探索新的人工智能方法 据知情人士透露 英伟达CEO黄仁勋将在本周访问韩国 计划与当地多家大型企业的负责人进行会谈 他将在参加完台北电脑展2026后 于周四晚上抵达首尔 随后与SK集团会长崔太元 LG集团会长聚光魔 以及Annever创始人李海珍等人会面 现代汽车集团的执行会长郑艺轩 也在考虑参与相关会议 不过三星集团会长李在荣 因海外行程安排将不会出席 这次会谈的内容不仅涉及AI半导体方面的合作 还有望拓展到机器人以及物理AI等新兴领域 此外 Never正在与英伟达协商 安排黄仁勋在下周一访问其第二办公大楼 今年3月 AMD董事长苏兹峰曾到访该地点 并与Never签署了关于AI基础设施的合作协议 业内分析认为 黄仁勋此次韩国之行 将进一步推动双方在人工智能领域的深度合作 2026年5月 阿里巴巴与欧足联签署6年合作协议 阿里云成为欧冠、欧联、欧协联 及2028年欧洲杯的官方独家AI 云计算及电商合作伙伴 欧足联在科技合作伙伴选择上速来苛刻 多年拒绝过多家科技巨头提案 此次选中阿里主要基于两方面的证明 其一是扛得住压力 体育赛事直播不能出现 服务器当机 实时处理和洪峰承载能力要求极高 其二是读得懂比赛 AI需在数秒内解析复杂赛事规则 并给出准确答案 阿里云之所以具备这些能力 源于多年服务奥运会的积累 2017年起 阿里成为国际奥委会Top合作伙伴 东京奥运会首次大规模启用云转播 北京冬奥会核心系统百分之百上云 巴黎奥运会云上转播规模 首次超过卫星转播 这些经历验证了阿里云的基础设施成熟度 和大规模并发处理能力 AI方面 签问大模型已在米兰冬奥会实现全量内容自动化打标 并驱动360度实时回放等特效 帮助国际奥委会打造奥运史上首个官方大模型 拿下欧族联合同 对阿里而言意味着完成了云全球化能力 与签问AI商业化能力的双重验证 将为其进一步拓展欧洲市场提供重要背书 英伟达推出了一款名为Cosmos 3的大模型 这是专门为物理人工智能设计的产品 物理人工智能就是让机器人和自动驾驶汽车 能够像人一样感知和理解真实世界的技术 Cosmos 3是全球第一个完全开源的全模态大模型 意味着它可以同时处理和生成文本、图片、视频、声音和动作等多种形式的内容 这款模型采用了混合Transformer架构 能够帮助机器人学习如何理解周围环境并做出正确的动作决策 以前训练一个智能机器人可能需要好几个月的时间 现在有了Cosmos 3 只需要几天就能完成训练 英伟达还联合了多家科技公司组成联盟 共同推动这项技术的发展 目前Cosmos 3已经推出了Super和Nano两个版本 其中Nano版本可以在几秒钟内完成复杂的视频分析和动作推理任务 另一个Edge版本也即将发布 主要用于边缘设备上的实时推理 经过多个行业基准测试 Cosmos 3在物理仿真精度、动作策略能力和视觉理解能力 第三个方面都表现优秀 排名都是第一 英伟达最近推出了一款名为NemoTron 3 Ultra的大型语言模型 这个模型有5500亿个参数 属于混合专家类型 它的主要作用是帮助智能体更聪明、更快速地完成各种复杂任务 相比同级别的其他开源模型 它的推理速度可以快最多5倍 使用成本能降低最多30% 这个模型已经准备好了 可以直接在多个智能体平台上使用 比如Hermis Agent、LangChain Deep Agents、OpenClaw、OpenHands和OpenCode 企业可以直接拿来部署 不用再做复杂的调整 目前有两家大公司已经开始使用这个模型了 Crowstrike用它来自动检查电脑漏洞 评估风险等级 还能帮助修复配置错误 这样网络安全人员就不用那么辛苦了 Talenteer则把它用在自己的AI平台上 让人工智能能够自主完成复杂的工作 并且能根据实际使用情况不断学习改进 这个模型预计会在6月4日发布 到时候可以在Hugging Face、Modelscope、OpenRouter和Build.NVIDIA.com这些平台上找到 英伟达还会通过自己的云计算合作伙伴网络来提供这个服务 在2026台北国际电脑展上 NVIDIA CEO黄仁勋做了主题演讲 他强调从产业角度看 Token就是一种资产 因为可以带来利润 AI公司因此希望生产更多Token 进而建造更多AI工厂 会上 NVIDIA推出了NVIDIA DSX平台 旨在为建设者提供完整的人工智能工厂解决方案 该平台整合了开源模块化软件库 API 参考设计 NVIDIA加速计算平台 以及合作伙伴技术 使得AI工厂的设计 部署和运营更加高效 简制机器人近日宣布完成连续多轮融资 总金额达数亿元 由蚂蚁集团 滴滴 德联资本 三家机构联合领头 顺维资本 百度风投 90智能等原有股东继续跟头 这是聚身智能无本体数据领域迄今为止最大规模的融资项目 简制机器人也因此成为该赛道累计融资最多的企业 聚身智能是只让机器通过身体与环境互动来学习和发展智能的技术 而高质量的数据是训练这类模型的关键 简制机器人采用从模型定义数据标准的方法 将高保真多模态的人类行为数据作为核心解决方案 公司自主研发了视觉模组 无线通讯等技术 实现多设备同步延迟小于一毫秒的行业领先水平 并首创多摄像头感知矩阵系统 通过多个RGB相机实现全视野高精度环境与行为记录 公司还开发了EGO认知中枢产品系列 包括Fingers仿生双指 DeX灵巧无智 Gripper工业甲爪等多种数据采集硬件 累计订单已突破1万台 是行业首个覆盖从头手到全身高精度数据获取的产品系列 在数据生成方面 公司首创了Data Foundation Model architecture realizes an end-to-end closed loop from training to true value verification. The head tracking accuracy is stable at the level of one centimeter. 6D attitude perception reaches the millimeter level. The company's true ADP professional giant intelligent data production line has covered more than 3,000 collection users and more than 10,000 real scenes, including homes, factories, commerce, logistics, laboratories, hospitals, etc. It has accumulated more than one million hours of real scene data and more than 2,000 daily human practical skills. Among them The monthly high-precision data collection capacity of Pitcher Collaboration has exceeded 100,000 hours. The company has established in-depth business cooperation with more than 30 leading artificial intelligence companies at home and abroad. This round of financing will focus on core technology research and development, including continued exploration of multiple modalities of human behavior data, building a data-based large model technology system, and opening up a full-link closed loop from data production to training adaptation to effect evaluation. Jianzhi Robot also plans to accelerate global market layout and deepen ecological cooperation with industry leading companies. Well, the AI news express is here Next, let’s get into today’s AI depth. In this issue, we won’t talk about specific companies or products. Let’s talk about an ongoing megatrend. Why are NVIDIA, OpenAI, and Anthropic? These three companies, which seem to have completely different paths, are suddenly rushing in the same direction. This direction is physical AI, allowing robots and self-driving cars to perceive, understand, and act in the real world like humans. NVIDIA has launched Cosmos 3, which is the world’s first fully open source, full-modal large model that can process text, images, videos, sounds, and actions at the same time. Nvidia said that it used to take months to train an intelligent robot. Now with Cosmos 3, it only takes a few days. This model has already won the first place in three benchmark tests: physical simulation accuracy, action strategy ability, and visual understanding. OpenAI is restarting the robot team. CEO Sam Altman said that his vision is to allow everyone to have their own personal robot in the future. The team's short-term goal is to help professionals build infrastructure. The long-term vision is that everyone has a personal robot that can complete any needs. This team originated from the World Simulation research project led by Aditya Remisch. After the Sora video application was shut down, it also absorbed the Sora team. OpenAI closed the robotics department in 2020. At that time, they believed that AGI could be achieved faster without the help of robots, and there was too little robot training data. Now they are back, with the goal of promoting AGI by developing general robots. Anthropic is doing something similar in the Cloud Code Auto mode. WorkTrees Routines are essentially allowing AI to gather together in the digital world, create branches by themselves, perform tasks by themselves, and make decisions by themselves. Although it is not a physical robot, this is a framework that allows AI to work on its own. The underlying logic is the same as what the robotics team is doing. So why now? Three reasons. First, the available data has increased. The company Jianji Robot told us that they have accumulated more than one million hours of real scene data. The head tracking accuracy is stable at one centimeter. 6D posture perception reaches sub-millimeter level. When the data quality is high enough and the scale is large enough, training a model that can work in the real world becomes possible. Second, the cost curve has reached a critical point. Cosmos 3 can compress the training time from months to days. Nemotron 3 Ultra can increase the reasoning speed by 5 times and reduce the cost by 30%. This means that for the first time, making a hands-on AI has become cheap enough and fast enough for small and medium-sized companies to participate. Third, the business logic has changed. Huang Renxun said something at the Taipei International Computer Show. From an industrial perspective, Token is an asset. What does this mean? Your ability to produce AI determines the size of your assets. You can either produce more Tokens to gain more profits, or you can only watch others occupy this market, while hands-on AI robots gather intelligence. Essentially, it is extending the ability of AI from speaking to doing things. The token increase brought by this expansion is exponential. So you see, NVIDIA open source Cosmos 3, OpenAI restarted the robot team, and Anthropic released the AUTO mode. On the surface, it is three things, but it is essentially the same thing. To seize the position of physical AI, the most important battlefield in the next ten years, how will this war be fought? I think there will be three routes. The first one is the infrastructure route. NVIDIA represents the GPU model, platform, and developer ecosystem. I want to be the one selling shovels. No matter who wins, I can win. The second one is the application route. OpenAI is the representative. I want to be the largest agent platform. Users only need to tell me what you want and I will complete everything, including controlling the physical world. The third one is the data route. Startups like Ji Robot are representatives. But one rule is constant. In this industry, companies that can define problems will always make the biggest profits. Companies that can solve problems will eat the soup. Companies that can only follow the trend will pay. Which route are you on now? Well, today’s AI sound is here for a while. Thank you for listening. This is the AI circle. Let’s understand AI with a group of people. I am Yun Ye. We will continue at the same time tomorrow. The above is all about the AI circle. We will see you tomorrow.