机器之心编辑部 马斯克于 2023 年与另外 11 位联合创始人共同创办的 xAI，如今已有 6 人离开。 最新消息，xAI 联合创始人 Jimmy Ba 周二表示，他已经离开了这家 AI 初创公司。 Jimmy 写道：这是我在 xAI 的最后一天。xAI 的使命是推动人类提升卡尔达舍夫等级（Kardashev tech tree）。我非常荣幸能在公司创立之初共同参与这一历程。由 衷感谢 @elonmusk 将我们聚集在一起，开启了这段不可思议的旅程。我为 xAI 团队所取得的成就感到无比自豪，未来也将作为团队的朋友继续保持紧密联系。感 谢大家一起并肩奋斗的岁月，和同事情谊才是我真正的财富。2026 年注定将是疯狂的一年，也极有可能是人类物种未来最繁忙（且最具影响力）的一年。 据知情人士透露，Jimmy 直接向马斯克汇报工作，并在去年年底之前一直负责公司相当大的一部分业务。直到去年晚些时候，他的多项职责被拆分，分别交由另 外两位联合创始人吴宇怀（Tony Wu，已经官宣离职）和 Guodong Zhang 负责。 Jimmy 是在不到 48 小时内第二位离开 xAI 的联合创始人。吴宇怀已于周一晚间宣布从这家 ...

2月9日到2月10日，埃隆·马斯克（Elon Musk）在社交媒体平台X上连续发帖，扔出了一个重磅消息：他 旗下的太空探索技术公司SpaceX，要调整战略重心了。 那个被他反复提及十余年的"火星殖民"愿景，正阶段性让位于一个更近、更具紧迫性的目标，在月球上 建造一座城市。 马斯克在社交媒体发帖解释道："对于不了解最新进展的人来说，SpaceX的重点已经转向月球，并计划 在那里建设一座可自我发展的城市。我们有可能在不到10年内实现这一目标，而真正实现火星殖民，则 仍需要更多的时间"。 随后他继续发帖，进一步道出这一战略转向背后的深层忧虑："之所以改变优先事项，是因为我担心自 然灾害或人为灾难可能切断来自地球的补给，从而导致殖民地走向灭亡"。 那个曾断言"月球只是干扰"、主张"直奔火星"的马斯克，为何突然将目光重新投向月球？这座设想中的 月球城市将如何建成？它真的可能在十年内从愿景走向现实吗？ 01 星际梦想绕道："火星先让路月球" 马斯克的这个决定，并非一时兴起，而是基于一套冷酷的现实计算。用他自己的话说，核心原因就两个 字：速度。 去火星，太难也太慢。由于行星运行轨道的关系，从地球发射飞船前往火星的最佳窗口期 ...

Core Viewpoint - SpaceX is advancing its vision of transforming the night sky into a solar-powered AI brain by proposing a satellite system that could deploy up to one million satellites to support AI computing needs in space [1][8]. Group 1: FCC Application and Approval - The Federal Communications Commission (FCC) has accepted SpaceX's proposal for a "non-geostationary satellite system" aimed at relocating high-energy AI computing to space [2]. - The FCC is seeking public comments on the application, with a deadline set for March 6, 2026 [4]. Group 2: Technical Details of the Satellite System - The proposed satellite system will operate at altitudes between 500 kilometers and 2000 kilometers, utilizing a sun-synchronous orbit to maximize solar energy [9]. - The network will connect with SpaceX's existing Starlink system, allowing for data routing and processing in orbit before transmission to ground stations [10]. Group 3: Strategic Implications and Future Plans - SpaceX's satellite project is described as a step towards achieving a "Type II Kardashev civilization," which involves harnessing energy from a star system [11]. - The project is also a key driver for SpaceX's potential initial public offering (IPO), which could raise billions of dollars [11].

Core Viewpoint - The next significant leap in artificial intelligence may occur in space rather than on Earth due to infrastructure and energy supply bottlenecks [1][3]. Group 1: Elon Musk's Vision - Elon Musk proposed that localized AI computing satellites in low-latency sun-synchronous orbits could become the most cost-effective way to transmit AI data within three years [2][3]. - Musk envisions launching 1 million tons of satellites annually, each equipped with 100 kW of power, which could add 100 GW of AI computing capacity each year without operational or maintenance costs [3]. - The concept includes building satellite factories on the Moon and using mass drivers to accelerate AI satellites to escape velocity, potentially achieving over 100 TW of AI computing capacity annually [3][4]. Group 2: Challenges in AI Development - The rapid growth in AI model sizes and demand is putting significant pressure on existing data centers, fiber networks, and power systems [5]. - There is a need for new energy sources to keep pace with the increasing demand, while also considering factors like latency, climate risks, and political obstacles [5]. Group 3: Competing Projects - Google's "Project Suncatcher" aims to create orbiting computing nodes powered by continuous solar energy, which could operate machine learning models more efficiently than ground-based data centers [6]. - Amazon's "LEO" project seeks to establish a global broadband network of low Earth orbit satellites, potentially providing edge computing services for AI tasks in underserved areas [7]. - Musk is also conceptualizing an "orbital computing farm" for xAI and SpaceX, which would not only run AI models but also train them, addressing the challenges of continuous energy supply and resource consumption [7].

Core Insights - Google's next-generation flagship model, Gemini 3, has gained significant attention even before its official release due to its impressive capabilities [1] - Users have shared positive feedback on Gemini 3's performance, particularly in front-end development and multi-modal tasks [1][3] Group 1: Model Features - Gemini 3.0 Pro utilizes a MoE architecture with trillions of parameters, activating only 15-20 billion parameters per query, and has expanded its context window from 1 million to several million [4] - The model includes a deep thinking mode and supports multi-modal inputs such as video, images, and audio [4] - In a challenging general intelligence test, Gemini 3.0 achieved an accuracy rate of nearly 35%, outperforming other models which scored below 20% [6] Group 2: Performance Highlights - User tests indicate that Gemini 3.0 excels in programming and multi-modal tasks, particularly in front-end and interface design [9] - The model successfully generated a visually appealing website for an ancient art museum and handled complex tasks related to the Kardashev scale [9] - In the "HLE benchmark," considered one of the toughest tests, Gemini 3.0 scored 32.4%, surpassing GPT-5 and Grok 4 [10] Group 3: Visual and Creative Capabilities - Gemini 3.0 demonstrated significant improvements in rendering images, producing high-quality game backgrounds, and creating intricate SVG graphics [12][15] - The model effectively understood physical principles and accurately counted objects in images, showcasing its advanced reasoning capabilities [21] Group 4: Anticipated Release - There are speculations about the release date of Gemini 3.0, with rumors suggesting it may launch on October 22, following earlier incorrect predictions [23][24] - Internal documents allegedly leaked from Google have fueled these release date discussions [24]