梦瑶 发自 凹非寺 量子位 | 公众号 QbitAI 不是，谁也没跟我说今年的AI春节大战搞得这么猛猛猛啊！？！ 年还没到呢，可灵就超绝不经意甩出一个「过大年计划」：推出 可灵3.0 多模态全家桶。 让每个人，都能上桌当——大导演 。 我主打一个先尝为快！先看我导的这出《拳王》（赛博版）动作大戏，10秒钟狂切6个分镜头： 不光我一个人玩嗨了，各位脑洞大开的网友们也纷纷share自己的大作了，下面这小哥直接搓出来了个超燃篮球赛大片，并直言够逼真！！！ 还有网友感慨，以后拍电影怕是都不用找真人演员了，这不嘛，人家直接找AI演了波超抽象的家庭大戏，脑洞太大了… 反正这波实测下来我最直观感受就是: 智能分镜能力确实夯，模型确实更能理解镜头语言了，像文字和人物的一致性上表现也蛮超出预期。 具体哪些功能最好用、适合啥样的使用场景，我也帮友友们整理好了（省流版）： 好莱坞大片也是手拿把掐，10秒钟切换7个镜头，从引擎轰鸣火花飞溅，再到男女主激烈争执，让我这个导演有点汗流浃背了... 灾难片自然我也不在怕的，浓雾封城、街道废弃、广告牌疯狂摇晃……咋样，是不是有点《后天》内味儿了： 1） 智能分镜｜音画同步｜主体一致性： 特别 ...

小米Mimo大模型团队投稿 量子位 | 公众号 QbitAI 小米MiMo大模型团队，加入AI拜年战场—— 推出 HySparse，一种面向Agent时代的混合稀疏注意力架构 。 HySparse创新使用极少的全注意力 （Full Attention） 层提供"token选择+KV Cache"，其余稀疏注意力 （Sparse Attention） 层直接复 用这些信息，实现高效精准的长上下文建模。 在总共49层的80B-A3BMoE模型实验中， 仅保留5层Full Attention仍能保持甚至提升模型能力，同时显著降低KVCache存储与计算开销 ，实现效果与效率的兼顾，展示出混合稀疏注意力在超长上下文建模中的巨大潜力。 HySparse的设计灵感来源于学术界已有研究工作的经验和观察之上 。 一部分是显著token在相邻层之间相对稳定。 已有工作如TidalDecode等，观察到连续层的 "重要 token" 会高度重合，因此可以在某层识别重要token并在后续层复用。 HySparse将这一观察提升用于模型结构设计并直接训练。 还有部分受启发于跨层KV Cache共享能显著省显存且不显著伤性能 ，YOC ...

Core Viewpoint - Wang Huiwen is actively seeking to engage with entrepreneurs and projects related to OpenClaw, indicating a strong interest in the AI sector and local agent development [1][2][30]. Group 1: OpenClaw and Its Impact - OpenClaw has gained significant traction, with its GitHub stars increasing from 100,000 to 171,000, nearly double that of PyTorch [10][11]. - The platform represents a new product approach in the AI era, focusing on integrating APIs into user workflows rather than just model development [8][9]. - OpenClaw's community engagement is evident, with over 1,000 fans attending its first offline gathering in San Francisco, showcasing strong user loyalty and brand presence [42][46]. Group 2: Wang Huiwen's Investments - Wang Huiwen has invested approximately $70 million in Kimi, which recently announced free access to its Kimi K2.5 model, marking a significant development for users and investors alike [36][37]. - His investment strategy spans various AI sectors, including foundational infrastructure, model development, and application layers, indicating a comprehensive approach to the AI landscape [68]. - Wang's previous ventures include the acquisition of OneFlow and investments in Trooly.AI, reflecting his ongoing commitment to AI innovation and entrepreneurship [60][67]. Group 3: Market Dynamics and Trends - The AI market is witnessing a surge in interest, with products like Moltbook and rentahuman.ai gaining popularity, although some skepticism exists regarding their user metrics and authenticity [20][25]. - The local agent sector is emerging as a hot trend, with significant potential for commercialization and user engagement, as evidenced by the sustained interest in OpenClaw [46][69]. - The rapid evolution of AI products and platforms suggests a dynamic market environment, where user engagement and community building are becoming increasingly important for success [41][45].

La ST₀ 团队 投稿 量子位 | 公众号 QbitAI 近日， 至简动力、北京大学、香港中文大学、北京人形机器人创新中心 提出了一种名为LaST₀的全新隐空间推理VLA模型，在基于 Transformer混合专家架构的快慢系统中，实现了 隐空间时空思维链 （Latent Spatio-Temporal CoT） 过程，实现了对物理世界的高效 推理，并且保持了高频的动作预测能力。 LaST₀提供了一种在具身大模型中引入高效隐空间推理的全新范式，在双臂、移动操纵、人形灵巧手上均实现SOTA水平，超越Pi0.5。 论文链接： https://arxiv.org/abs/2601.05248 项目主页： https://vla-last0.github.io/ 视觉-语言-动作 （VLA） 模型近期展现出了强大的泛化潜力，部分前沿方法尝试在执行前显式生成语言推理链或预测未来状态。然而，这 种显式推理往往会引入不可忽视的推理延迟，从而限 制了机器人操控所需的控制频率；更重要的是，此类推理受限于语言空间，难以精准 刻画那些"不可言说"的物理属性（物理规律、环境动态、几何关系等），形成了表示瓶颈。 为了解决这些挑战，团 ...

Core Viewpoint - The article emphasizes the ongoing AI boom and invites individuals to join the company "Quantum Bit," which focuses on tracking AI advancements and has established itself as a leading content platform in the industry [1]. Group 1: Job Opportunities - The company is hiring for three main directions: AI Industry, AI Finance, and AI Product, with positions available for both experienced professionals and fresh graduates [2][4]. - Positions are open for various levels, including editors, lead writers, and chief editors, with a focus on matching roles to individual capabilities [6]. Group 2: Job Responsibilities - **AI Industry Direction**: Responsibilities include tracking innovations in infrastructure, such as chips, AI infrastructure, and cloud computing, as well as interpreting technical reports from conferences [6][7]. - **AI Finance Direction**: Focuses on venture capital, financial reports, and capital movements within the AI industry, requiring sensitivity to data and strong logical structuring [11]. - **AI Product Direction**: Involves monitoring AI applications and hardware developments, conducting product evaluations, and engaging with entrepreneurs and product experts [11]. Group 3: Benefits and Growth - Employees can expect to gain exposure to the latest AI technologies, enhance their work efficiency through new tools, and build personal influence in the AI field [6]. - The company offers competitive salaries, comprehensive benefits, and a supportive team environment that encourages growth and mentorship [6][11]. Group 4: Company Impact - By 2025, Quantum Bit aims to have over 2.4 million subscribers on WeChat and more than 7 million users across platforms, with a daily reading volume exceeding 2 million [12].

Core Viewpoint - The article discusses the increasing integration of AI in transportation safety, particularly through the "Eagle Eye" warning system developed by Gaode, which enhances driver awareness and reduces accident risks during the Spring Festival travel season [2][4][6]. Group 1: Spring Festival Travel - The Spring Festival travel volume is expected to reach a record high, with an estimated 9.5 billion trips over 40 days, and 80% of travelers opting for self-driving [1]. - The article highlights the unique aspects of this year's travel, emphasizing the role of AI in enhancing safety during journeys [1]. Group 2: AI's Role in Safety - The "Eagle Eye" system, developed in collaboration with the China Safety Production Science Research Institute, provides real-time risk awareness by detecting 24 types of potential hazards, including sudden braking and adverse weather conditions [4][6]. - The system has been upgraded to offer broader coverage and faster alerts, ensuring that it is accessible across various vehicles and road types [7]. Group 3: Technical Implementation - The core of the "Eagle Eye" system is the TrafficVLM model, which utilizes real-world traffic data to create a digital twin for training purposes [8][10]. - TrafficVLM enhances the system's ability to predict traffic conditions and provide timely warnings to drivers, thereby improving overall road safety [13][15]. Group 4: Performance Metrics - As of February 1, 2026, the "Eagle Eye" system has issued 11.2 billion warnings, averaging 88 million warnings per day, contributing to a 10% reduction in daily accident rates during peak travel times [16][18]. - The system's effectiveness is validated by real-world data, demonstrating its ability to help drivers avoid potential accidents [16][19].

Core Viewpoint - The article discusses the launch of OpenScholar, an AI system developed by the Allen Institute for AI and the University of Washington, which aims to eliminate the issue of false citations in academic writing by leveraging a vast database of 45 million scientific papers [2][5]. Group 1: OpenScholar's Features - OpenScholar connects to a large database called ScholarStore, which contains full texts and abstracts of 45 million papers, significantly reducing the false citation rate of traditional large language models (LLMs) [9][11]. - The system employs Retrieval-Augmented Generation (RAG) technology to ensure that each knowledge point is backed by a real paper, enhancing the accuracy of citations [12][13]. - OpenScholar's feedback loop allows it to refine its outputs by searching, generating, self-reviewing, and revising, which helps confirm the existence of supporting literature [12][13]. Group 2: Performance Comparison - In a benchmark test called Scholar QABench, OpenScholar-8B outperformed GPT-4o by 5% in correctness and matched human expert citation accuracy [16]. - A double-blind experiment showed that 51% of OpenScholar's answers were rated better than those written by human researchers, with an upgraded version achieving a 70% success rate [18]. - Experts noted that OpenScholar's strengths lie in its comprehensive information coverage, clearer structure, and stronger logical coherence compared to traditional models [19].

Core Viewpoint - Nvidia has indefinitely postponed the release of the RTX 50 Super and the next-generation RTX 60 series due to a global shortage of memory chips, prioritizing AI GPU production instead [2][15][18]. Group 1: Nvidia's Product Delays - Nvidia has historically released new gaming GPUs every other year, but this year, it has broken tradition by not launching the RTX 50 Super as expected [8][10]. - The RTX 50 Super was reportedly already designed under the codename "Kicker," but the release was delayed as of December last year [12][13]. - The delay will also affect the planned production of the RTX 60 series, originally set for late 2027 [14]. Group 2: Market Impact and Pricing - The price of the RTX 5090 has surged from an initial MSRP of $1999 to between $3500 and $4000, with projections suggesting it could reach $5000 by the end of the year [27][28]. - The ongoing chip shortage is causing price increases across all PC gaming components, leading to a potential shift in consumer behavior towards cloud services or delaying hardware upgrades [26][35]. - Companies like Valve and Nintendo are also reevaluating their product pricing and release schedules due to the memory shortage, with Nintendo having already increased the prices of its Switch models [29][33].

Core Viewpoint - The discussion highlights Elon Musk's vision for the future of AI and space, emphasizing the potential of space as a hub for AI infrastructure and energy production, which could surpass Earth's capabilities within the next five years [5][36][96]. Group 1: Space as AI Infrastructure - Musk predicts that within 30 to 36 months, space will become the preferred location for AI infrastructure due to its advantages in energy efficiency and scalability [5][12][16]. - The anticipated annual AI computing power in space is expected to exceed the cumulative total on Earth within five years, with projections of reaching around 1 terawatt (TW) of power generation [36][58]. - Space solar panels are estimated to be five times more efficient than those on Earth, eliminating the need for batteries and significantly reducing costs [12][28][30]. Group 2: Energy Supply Challenges - Musk identifies energy supply as a critical issue, noting that while chip production is rapidly increasing, energy production is stagnating outside of China [6][7]. - The construction of data centers on Earth faces significant regulatory and logistical challenges, making space a more viable option for expansion [11][12][19]. - The average power consumption in the U.S. is around 500 gigawatts (GW), and Musk emphasizes the difficulty of scaling energy production to meet the demands of large data centers [17][58]. Group 3: Chip Production and Supply Chain - Musk discusses the need for large-scale chip manufacturing facilities, suggesting a project akin to "TeraFab" to meet future demands for AI chips [48]. - The current chip supply chain is constrained, with existing foundries unable to meet the anticipated demand, leading to potential bottlenecks in AI deployment [56][57]. - Musk expresses concerns about memory production, indicating that the path to producing sufficient memory for logic chips is less clear than that for logic chip manufacturing [52]. Group 4: Competitive Landscape and Innovation - Musk warns that without breakthrough innovations, the U.S. risks losing its competitive edge to China, which is rapidly advancing in manufacturing and energy production [96]. - The discussion touches on the importance of maintaining a skilled workforce and the challenges posed by China's larger population and manufacturing capabilities [92][95]. - Musk believes that the future of companies will increasingly rely on AI and robotics, which will outperform human-involved companies in efficiency and productivity [80].

Core Viewpoint - The domestic intelligent computing infrastructure is crossing a critical watershed, marking a significant advancement in the deployment and operationalization of large-scale AI computing clusters in China [1][5]. Group 1: Scale of Deployment - The first domestic AI computing pool with a deployment of 30,000 cards has officially formed, representing the largest operational domestic AI computing pool [3][5]. - The transition from single-point breakthroughs to large-scale deployment indicates a maturity in engineering capabilities for domestic supercomputing clusters [5][9]. Group 2: System Engineering Capabilities - The competition in the large-scale computing cluster market has shifted from merely achieving high card counts to focusing on system-level collaboration, including network, storage, cooling, power supply, and operational optimization [8][9]. - Key capabilities such as controllable cycles, reproducible performance, fault localization, and cost accounting are essential for long-term players in the market [8][12]. Group 3: Real-World Application - The true test for large-scale computing clusters lies in their ability to convert computing power into real business productivity, moving beyond mere demonstrations to practical applications [14][15]. - Successful large-scale applications require open compatibility, making it easier for users to adapt and integrate into existing systems [17][19]. Group 4: Public Infrastructure - The perception of large-scale computing clusters is evolving from being exclusive to large corporations to becoming public infrastructure, serving various sectors such as manufacturing, energy, transportation, research, education, healthcare, and finance [20]. Group 5: Performance in Various Scenarios - The scaleX computing cluster demonstrates its value across three dimensions: stability in large model training, practical service capabilities in high-throughput inference, and accelerated scientific discovery in AI for Science applications [23][26]. - The cluster has successfully supported over 400 mainstream models and thousands of applications, showcasing its ability to integrate with the AI industry ecosystem [21][26]. Group 6: Future Outlook - The release of three scaleX computing clusters signals that merely stacking computing power is no longer the core competitive advantage; future winners must ensure that these clusters continuously deliver value in real-world business applications [27][28].