克雷西 发自 凹非寺 量子位 | 公众号 QbitAI 马斯克脑机接口，又有新动态。 Neuralink官方视频当中，一名演示者已经实现了机械臂的意念控制。 只见机械臂稳步抬起，然后伸向演示者嘴边，然后献上了一吻。 马斯克还表示，理论上通过电脑或手机，Neuralink将能够间接控制一切。 该方案计划在他保留现有颅骨植入物的基础上，在脊髓位置新增第二个接口。 这一架构的目的是建立"数字神经桥梁"，大脑端的芯片负责解码运动意图，脊髓端的芯片则负责接收指令并直接刺激肢体神经，从而绕过受损 的生物通路，试图让瘫痪患者重新恢复行走能力。 在硬件稳定性方面，技术团队已针对初代产品暴露出的"电极线回缩"问题完成了关键迭代。 与此同时，Neuralink正在推进一项升级试验。 首位受试者或将收到更新 Neuralink的升级实验将采用"双植入"的方案，首位人类受试者Noland Arbaugh预计将成为首个接受二次手术的患者。 针对首位患者因颅内气隙导致电极线松脱、信号减弱的情况，在针对第二位受试者Alex及后续患者的手术中，团队改进了手术方案，通过严格 控制植入物与大脑表面的间隙（Gap Reduction），并优化了手 ...

OpenDriveLab 投稿 量子位 | 公众号 QbitAI 人形机器人要在人类环境中执行各种任务，需要同时具备两个看似矛盾的能力： 敏捷的动态运动 和 精确的平衡控制 。 AMS从三个关键方面解决动态运动与平衡控制的统一问题： 1. 异构数据源 ：从机器人动作空间直接采样生成可扩展的平衡数据，突破人类数据限制，缓解长尾分布问题。 下面来看详细内容。 人形机器人的"两难困境" 叶问蹲、跳舞、跑步，一个策略全搞定！ 核心思路： 反观人类，却能轻松自然的实现这种协同——比如在动态行走后精确放置物体，或者在单腿站立时用自由肢体作为临时支撑去够取物体。 近日，来自香港大学、NVIDIA和清华大学的联合研究团队提出了一种名为 AMS （Agility Meets Stability） 的统一人形机器人全身控制框 架， 首次 实现了在单一策略中同时具备动态运动跟踪和极限平衡控制能力。 2. 混合奖励机制 ：选择性应用平衡先验奖励，精准平衡指导不牺牲敏捷性，化解优化目标冲突。 3. 自适应学习策略 ：动态调整采样概率，同时对每个动作"因材施教"，实现高效的自适应学习。 然而，对于人形机器人来说，同时实现这两种能力却是一 ...

鱼羊 发自 凹非寺 量子位 | 公众号 QbitAI 在面对复杂任务时，消耗的token数偏多，甚至可能会出现"又长又错"的答案。 比如，同样解决一个问题，Gemini只用了2万个token，而Speciale需要花费7.7万个。 这是怎么一回事？ 没有被纠正的"长度偏见" 有研究者指出，这其实是自DeepSeek-R1-Zero以来，DeepSeek系列模型一直存在的一个"bug"。 DeepSeek-V3.2很强很火爆，但随着讨论的深入，还是有bug被发现了。 并且是个老问题：浪费token。 △ 图源：x@Hangslin 不少网友都提到，DeepSeek-V3.2的长思考增强版Speciale，确确实实以开源之姿又给闭源TOP们上了压力，但问题也很明显： 结果就是：模型会故意生成"又长又错"的答案，看起来像是在"认真推理"，其实是在"凑字数躲惩罚"。 难度偏见 ：太简单或太难的题被过度关注 GRPO会根据"同一批题的得分标准差"调整权重。比如一道题所有人都做对（标准差小），或所有人都做错（标准差也小），这道题会被当成 "重点" 反复训练；而中等难度、有人对有人错的题（标准差大），反而被忽略。但实际训练 ...

Core Viewpoint - The article discusses the introduction of Improved MeanFlow (iMF), which addresses key issues in the original MeanFlow (MF) model, enhancing training stability, guidance flexibility, and architectural efficiency [1]. Group 1: Model Improvements - iMF reformulates the training objective to a more stable instantaneous velocity loss, introducing flexible classifier-free guidance (CFG) and efficient in-context conditioning, significantly improving model performance [2][14]. - In the ImageNet 256x256 benchmark, the iMF-XL/2 model achieved a FID score of 1.72 in 1-NFE, a 50% improvement over the original MF, demonstrating that single-step generative models can match the performance of multi-step diffusion models [2][25]. Group 2: Technical Enhancements - The core improvement of iMF is the reconstruction of the prediction function, transforming the training process into a standard regression problem [4]. - iMF constructs the loss from the perspective of instantaneous velocity, stabilizing the training process [9][10]. - The model simplifies input to a single noisy data point and modifies the prediction function's computation, removing dependency on external approximations [11][12][13]. Group 3: Flexibility and Efficiency - iMF internalizes the guidance scale as a learnable condition, allowing the model to adapt and learn average velocity fields under varying guidance strengths, thus enhancing CFG flexibility during inference [15][16][18]. - The improved in-context conditioning architecture eliminates the need for the large adaLN-zero mechanism, optimizing model size and efficiency, with iMF-Base reducing parameters by about one-third [19][24]. Group 4: Experimental Results - iMF demonstrates exceptional performance on challenging benchmarks, with iMF-XL/2 achieving a FID of 1.72 in 1-NFE, outperforming many pre-trained multi-step models [26][27]. - In 2-NFE, iMF further narrows the gap between single-step and multi-step diffusion models, achieving a FID of 1.54 [29].

Core Insights - The MEET2026 Smart Future Conference will focus on cutting-edge technologies and industry developments that have garnered significant attention throughout the year [1] - The theme "Symbiosis Without Boundaries, Intelligence to Ignite the Future" emphasizes how AI and smart technologies penetrate various industries, disciplines, and scenarios, becoming a core driving force for societal evolution [2] Group 1: Conference Highlights - The conference will cover hot topics in the tech circle this year, including reinforcement learning, multimodal AI, chip computing power, AI in various industries, and AI going global [3] - It will feature the latest collisions between academic frontiers and commercial applications, showcasing leading technological achievements from infrastructure, models, and product industries [4] - The event will also include the authoritative release of the annual AI rankings and the annual AI trend report [5][116] Group 2: Notable Speakers - Zhang Yaqin, President of Tsinghua University's Intelligent Industry Research Institute and an academician of the Chinese Academy of Engineering, will be a key speaker [11] - Sun Maosong, Executive Vice President of Tsinghua University's Artificial Intelligence Research Institute, has led numerous national projects [15] - Wang Zhongyuan, Director of the Beijing Academy of Artificial Intelligence, has extensive experience in AI core technology research [19] - Wang Ying, Vice President of Baidu Group, oversees several key business units including Baidu Wenku and Baidu Netdisk [24] - Han Xu, Founder and CEO of WeRide, has led the company to become a leader in autonomous driving technology [28] Group 3: Annual AI Rankings and Trends - The "Artificial Intelligence Annual Rankings" initiated by Quantum Bit has become one of the most influential rankings in the AI industry, evaluating companies, products, and individuals across three dimensions [117] - The "2025 Annual AI Top Ten Trends Report" will analyze ten AI trends that are releasing significant potential, considering factors like technological maturity and practical application [118] Group 4: Event Details - The MEET2026 Smart Future Conference is scheduled for December 10, 2025, at the Beijing Jinmao Renaissance Hotel, with registration now open [119] - The conference aims to attract thousands of tech professionals and millions of online viewers, establishing itself as an annual barometer for the smart technology industry [122]

Core Insights - Amazon Web Services (AWS) showcased an unprecedented number of product launches at the re:Invent 2025 event, with CEO Matt Garman challenging himself to release 25 products in 10 minutes, ultimately unveiling 40 new products in just over two hours, emphasizing practicality and addressing challenges in AI applications [1][7][9]. Group 1: AI Computing Power - AWS has restructured its AI computing supply model by focusing on self-developed chips, specifically the Trainium series, which has grown into a multi-billion dollar business with over 1 million chips deployed, outperforming competitors by four times in speed [15][20]. - The latest Trainium3 Ultra Servers, based on 3nm technology, offer a 4.4 times increase in computing performance and a 3.9 times increase in memory bandwidth compared to the previous generation [18]. - The upcoming Trainium4 chip promises significant advancements, including a 6 times increase in FP4 computing performance and a 4 times increase in memory bandwidth, tailored for large model training needs [20][22]. - AWS introduced AI Factories, allowing clients to deploy AWS AI infrastructure within their data centers, thus maintaining control and security while accessing top-tier AI computing power [23][24]. Group 2: Model Development and Integration - AWS launched Amazon Bedrock, a flexible and customizable model platform, which now includes Chinese models Kimi and MiniMax, marking their entry into the global developer ecosystem [26][28]. - The new Amazon Nova 2 series includes various models designed for different tasks, with Nova 2 Light focusing on cost-effectiveness and low latency, Nova 2 Pro excelling in complex tasks, and Nova 2 Sonic optimizing real-time voice interactions [30][32]. - Nova Forge introduces the concept of Open Training Models, allowing enterprises to integrate their proprietary data with AWS's training datasets, creating specialized models that retain general reasoning capabilities while understanding unique business knowledge [40][41]. Group 3: AI Agents - AI Agents emerged as a key focus, with Garman stating that the era of AI assistants is being replaced by AI Agents, which will be widely adopted across companies [45][46]. - AWS introduced several new Agents, including Kiro Autonomous Agent for complex development tasks, AWS Security Agent for proactive security measures, and AWS DevOps Agent for continuous system monitoring and troubleshooting [50][52][56]. - AWS provides tools like AWS Transform Custom for code migration and Policy in AgentCore for defining agent behavior, ensuring that agents operate within controlled parameters [58][61]. Group 4: Strategic Vision - AWS's strategy emphasizes the importance of practical applications of AI technologies, focusing on building a comprehensive, secure, and scalable enterprise-level infrastructure rather than solely on technological breakthroughs [68][70]. - The company aims to address challenges related to computing costs, model understanding of proprietary knowledge, and the controllability of AI Agents through its innovative solutions and partnerships [70].

Jay 发自 凹非寺 量子位 | 公众号 QbitAI 创投风口赛道上的具身智能，让垂直场景上的玩家有了「焕新」机遇。 这不，主打机器人控制系统的 仙工智能 ，又一次来到了港交所门前。 仙工智能在5月就首次向港交所递交了招股书，但后面没能转换为实质进展。 这一次带着更新的业绩和数据，再次冲刺港股IPO。 过去三年，仙工智能的营收其实在一路走高： 2022年1.84亿元、2023 年2.49亿元 ，到了2024年，这一数字再次跃迁——来到3.39亿元。 大概就是现在的仙工智能每天都有将近一百万元进账。 但营收规模逐年扩大的背后，盈亏平衡节点还没到来。 公司连续三年亏损，累计已达1.22亿元。 仙工智能做什么？ 仙工智能，一家以机器人控制系统为核心的智能机器人公司。 事实上，仙工智能所做的，并不是这个行业里最炫目的那颗明珠—— 他们没有造To C全尺寸人形机器人，也没有展示能在舞台上翻跟头、跳舞的惊艳demo。 相比这些，仙工智能更聚焦于为工厂提供解决方案。 但从商业逻辑上看，比起通用韧性，工业场景或许也是眼下相对容易跑通的方向。 仙工智能的产品矩阵主要包含四个象限—— 控 制 器、软件、机器人、配件。 运行于仙 ...

AI热潮还在汹涌，但如果你还不知道如何参与……那为什么不来 量子位 呢？ 我们是一家以 追踪AI新进展 为核心的内容平台，经过8年积累，目前拥有顶流影响力，广泛且备受认可的产业资源，以及时代风口的最佳观 测和学习生态位。 编辑部 发自 凹非寺 量子位 | 公众号 QbitAI 目前，我们有 三大方向 岗位招聘，希望你是 （或者能成为） 这三个方向的内容专家： 岗位均为全职，工作地点：北京中关村。 岗位面向： 加入我们，你可以获得： 以下是岗位详情： 所有岗位不同能力层级职位均在开放，欢迎结合个人履历和经验申请。 AI产业方向 岗位职责： AI产业方向 ：关注基建层创新，包含芯片、AI Infra、云计算； AI财经方向 ：关注AI领域创投和财报，跟踪产业链资本动向； AI产品方向 ：关注AI在应用和硬件终端方向的进展。 社招：覆盖编辑、主笔、主编各个层级，按能力匹配岗位； 校招：应届毕业生，接受实习且可转正。 站在AI浪潮之巅 ：第一时间接触和了解AI领域最新技术和产品，构建完整的AI认知体系。 玩转AI新工具 ：将各种AI新技术、新工具应用于工作，提升工作效率和创造力。 打造个人影响力 ：通过撰写独家原创内 ...

henry 发自 凹非寺 量子位 | 公众号 ChatGPT三岁生日这一天，硅谷热议的新模型来自 DeepSeek 。 准确说是 两款开源 模型—— DeepSeek-V3.2 和 DeepSeek-V3.2-Speciale 。 这俩模型火到什么程度呢？ 有网友表示，在去圣地亚哥的（疑似赶场NeurIPS 2025）航班上，有30%的乘客都在对着DeepSeek的PDF两眼冒光。 其中，标准版DeepSeek-V3.2在推理测试中，达到了GPT-5的水平，仅略低于Gemini-3.0-Pro。 而"特别版"DeepSeek-V3.2-Speciale不仅全方位超越了GPT-5，还能在主流推理任务中和Gemini-3.0-Pro掰掰手腕。 此外，V3.2-Special还拿下了IMO、CMO、ICPC及IOI的金牌，并在ICPC和IOI上达到了人类选手第二名与第十名的水平。 而上周嘲讽DeepSeek "昙花一现"的推特更是在发布的当晚被刷到了 500万 浏览。 除了普通网友，奥特曼也是急急急急：不仅启动红色警报，还临时推迟了在ChatGPT上投放广告的计划。 与此同时，那一头的谷歌也没被放过。 网友直接 " ...

Core Insights - OpenAI is under pressure to enhance ChatGPT in response to Google's Gemini 3 Pro, leading to an internal "code red" alert to prioritize resources for this task [2][34] - The competition has intensified, with Google's Gemini series gaining significant traction, resulting in a decline in ChatGPT's traffic and user engagement [10][15] Group 1: Competitive Landscape - OpenAI's ChatGPT traffic dropped by 6% within a week following the release of Gemini 3 Pro and Nano Banana Pro [10] - Gemini's monthly active users surged from 450 million in July to 650 million in October, indicating rapid growth and increased competition [14] - OpenAI's market share in AI assistants remains at 70%, but growth is slowing, raising concerns about its competitive edge [15][16] Group 2: Financial Challenges - OpenAI is not yet profitable and faces increasing financial pressure, needing to raise approximately $100 billion to sustain operations amid high cash burn [19][21] - Projected revenues from ChatGPT are $10 billion for this year, $20 billion next year, and $35 billion by 2027, but these figures are insufficient to cover expenses [20][21] - The company must achieve $200 billion in revenue by 2030 to reach profitability, which poses a significant challenge given its current financial trajectory [21] Group 3: Product Development and Strategy - OpenAI plans to release a new reasoning model next week, with the internal codename "Garlic," which is expected to outperform Gemini 3 in evaluations [5][24] - The development of Garlic has reportedly made significant strides in pre-training, allowing for better performance in coding and reasoning tasks [28] - OpenAI is also working on additional models, including one named "Shallotpeat," to further enhance its product offerings and respond to competitive pressures [27][31]