一水 发自 凹非寺 量子位 | 公众号 QbitAI 我并不认为Scaling Law已经完全结束了 。 正当学生Ilya为Scaling Law"泼下冷水"时，他的老师、AI教父Geoffrey Hinton却毅然发表了上述截然相反的观点。 这一场面一出，我们不禁回想起了两件有趣的事。 一是Ilya几乎从学生时代起就坚信Scaling Law，不仅一抓住机会就向身边人安利，而且还把这套理念带进了OpenAI。 可以说，Ilya算是Scaling Law最初的拥趸者。 二是Hinton后来在回顾和Ilya的相处时，曾大肆夸赞Ilya"具有惊人的直觉"，包括在Scaling Law这件事上，Hinton曾坦言： 当时的我错了，而Ilya基本上是对的。 比如Transformer确实是一种创新想法，但实际上起作用的还是规模，数据的规模和计算的规模。 但是现在，这对师徒的态度却来了个惊天大反转。 所以，这中间到底发生了什么？ Scaling Law不死派：Hinton、哈萨比斯 其中，最大的挑战无疑是数据缺失问题。 大部分高价值数据都锁在公司内部，免费互联网数据已基本耗尽。 而这个问题将由AI自行解决，即模型通过推 ...

混元团队 投稿 量子位 | 公众号 QbitAI 在语言模型的比拼中，机器翻译一直被视为检验机器理解复杂语义和跨文化对齐能力的"试金石"。 面向端侧场景，12月30日，腾讯混元宣布推出并开源翻译模型1.5，经过量化，可支持端侧直接部署和离线实时翻译，仅需1GB内存即可流 畅运行，并且在参数量极小的前提下，效果超过了大部分商用翻译API。 在常用的中外互译和英外互译测试集Flores200、WMT25以及民汉语言的测试集中，Tencent-HY-MT1.5-1.8B全面超越中等尺寸开源模型 和主流商用翻译API，达到Gemini-3.0-Pro这种超大尺寸闭源模型的90分位水平。在WMT25和民汉翻译测试集上，其效果仅略微差于 Gemini-3.0-Pro，远超其他模型。 模型在效率和性价比也表现突出，与主流商用翻译模型API对比，HY-MT1.5-1.8B推理速度更快，处理50个tokens的平均耗时只有0.18 秒，其他模型的时间在0.4秒左右，显示出明显的速度优势，凭借优化的模型设计和推理逻辑，其领先的效率使其高度适用于即时通讯、智 能客服、移动翻译应用等高吞吐、实时翻译场景。 在大模型厂商纷纷角逐手机等 ...

henry 发自 凹非寺 量子位 | 公众号 QbitAI 2025年的最后一天，上市公司上纬新材董事长 彭志辉 （稚晖君）发布了一款能装进书包的机器人产品—— 上纬启元Q1 。 这是全球首款最小尺寸（0.8m）、实现全身力控的人形机器人，也是智元机器人联合创始人 稚晖君 担任上纬新材董事长以来，发布的首款具 身智能机器人产品。 虽然体型迷你，但大机器人能做的，启元Q1也能做。 大机器人做不了的，启元Q1还能做。 （我骑过狗你骑过吗？） 而前段时间让网友猜疯了的 "大有可为" 神秘海报，也终于在这次的发布视频中正式揭晓答案。 其中醒目的1.88，既不是身高，也不是售价，而是启元Q1的体积（立方米）——一个被压缩到背包级的人形机器人尺寸。 启元Q1是一款怎样的机器人？ 从产品定位上看，稚晖君这次的新作 启元Q1 ，是一款面向个人用户、开发者，科研、陪伴、创作场景的小尺寸人形机器人。 值得一提的是，这种小型化设计，并不只是为了方便携带。更轻的重量，让机器人本身更耐造，也把使用和试错成本一起打了下来，更适合个 人和小团队反复折腾。 在产品能力上，启元Q1反复强调了一个关键词—— 全身力控 。 简单来说，全身力控并不 ...

Core Viewpoint - Elon Musk's xAI is expanding its computing power through the "Macrohard" project, with the acquisition of a third facility named MACROHARD RR, which will have a power supply capacity of 2GW [1][2]. Group 1: Facility Expansion - The new facility MACROHARD RR is located near the existing Colossus II site, which is part of the Macrohard project [15][16]. - Colossus I, the first facility, was built in just 122 days and is currently the largest and most stable computing cluster globally, equipped with approximately 200,000 NVIDIA H100/H200 and 30,000 NVIDIA GB200 NVL72 GPUs [6][7]. - Colossus II is set to deploy 110,000 NVIDIA GB200 GPUs in its first phase, with a final goal of over 550,000 GPUs and a peak power demand exceeding 1.1GW [11]. Group 2: Power Supply and Infrastructure - The 2GW power capacity of the new facility can support around 1.1 million NVIDIA GB200 NVL72 GPUs, based on previous power density and efficiency metrics [2][4]. - xAI has partnered with Solaris Energy Infrastructure to build a permanent gas turbine power plant in Mississippi, which is expected to provide over 1GW of power by early 2027 [18][20]. - To mitigate noise complaints from nearby residents, xAI has constructed a wall between the power plant site and residential areas and deployed 168 Tesla Megapack battery storage systems to support local power needs during peak usage [20]. Group 3: Financial Aspects - xAI is reportedly planning to raise $15 billion at a valuation of $230 billion to support its expansion efforts [22]. - Musk has denied reports regarding the fundraising but has not provided further clarification [23].

Group 1 - Nvidia is reportedly planning to acquire Israeli AI startup AI21 Labs for $2-3 billion to recruit over 200 top AI talents [1][2][40] - AI21 Labs, founded in 2017, specializes in developing large language models and has a strong founding team with notable backgrounds in AI and technology [3][7][27] - The company's valuation in 2023 is approximately $1.4 billion, with a recent funding round led by Nvidia and Google raising $300 million [4][6] Group 2 - The acquisition reflects Nvidia's strategy of "talent acquisition" rather than traditional business mergers, allowing it to bypass strict regulations on business monopolies [41][42] - AI21 Labs has developed its own models, including the Jurassic series and the recently launched Jamba, which is an open-source large model [28][29] - The partnership between Nvidia and AI21 Labs aims to combine Nvidia's computational infrastructure with AI21's application solutions, enhancing enterprise-level generative AI deployment [36][50] Group 3 - Nvidia's recent acquisitions, including Groq, demonstrate a pattern of acquiring companies primarily for their talent rather than their technology [43][45] - The acquisition of AI21 Labs is expected to further solidify Nvidia's strategic position in Israel, integrating hardware, software, and AI applications [50][51] - Nvidia's ambition extends beyond being a chip company, as it seeks to control the entire AI industry chain through strategic acquisitions [52][54]

Core Viewpoint - MiniMax, a Chinese AI company, is set to go public with an IPO aiming to raise over $600 million, valuing the company at over HKD 46.1 billion, and is expected to list on January 9, 2026 [2][7]. Group 1: Company Overview - MiniMax is positioned as a global artificial general intelligence (AGI) technology company, with services covering over 200 countries and regions, and 70% of its revenue coming from international operations [12]. - The company has a strong backing from 14 cornerstone investors, including Alibaba and the Abu Dhabi Investment Authority, with total subscriptions amounting to approximately HKD 27.23 billion [7][8]. Group 2: Market Context - December 2025 marks a significant period for IPOs in Hong Kong, with 25 companies having completed listings, making it the busiest month since 2019 [9]. - MiniMax and another company, Zhiyuan, are both entering the market around the same time, creating a competitive atmosphere that splits investor attention [10]. Group 3: Financial Performance - MiniMax's revenue has shown remarkable growth, reaching $3.46 million in 2023 and projected to soar to $30.52 million in 2024, representing a year-on-year increase of 782.2% [35]. - For the first nine months of 2025, revenue surged by 175% to $53.44 million, significantly surpassing the previous year's total [36]. - The company has improved its gross margin from -24.7% in 2023 to 23.3% in the first nine months of 2025, indicating a positive trend in profitability [38]. Group 4: Product Development - MiniMax has released several models, including the M1 and M2 text models, with M2 achieving top rankings in performance metrics [20][21]. - The company has also developed a voice model, Speech 01, and its upgraded version, Speech 02, which supports over 40 languages and has generated over 2.2 million hours of speech [24]. - MiniMax's video model, Hailuo, has been recognized for its capabilities in generating videos and has helped create over 590 million videos globally [28]. Group 5: Investment and Support - MiniMax has raised over $1.5 billion in funding from various strategic investors, including major tech companies and venture capital firms, positioning it as a leading player in the AGI space [50]. - The company has a cash reserve of $1.102 billion as of September 30, 2025, which is sufficient to sustain operations for over 53 months without additional funding [46].

Core Viewpoint - The article discusses the emergence of "Hey Tuya," an AI life assistant developed by Tuya Smart, which integrates software and hardware to create a seamless smart home experience, moving beyond simple command responses to proactive engagement in daily life [6][46]. Group 1: AI Life Assistant Features - "Hey Tuya" allows users to control smart home devices through a unified interface, enabling actions like adjusting lighting and temperature with simple voice commands [8][16]. - The assistant can manage home security by providing real-time updates and alerts, such as recognizing delivery personnel through camera feeds [23][24]. - It offers energy management solutions, allowing users to set automated energy-saving strategies for their devices [30][31]. Group 2: User Interaction and Personalization - The AI can create to-do lists and reminders based on natural language conversations, adapting to users' habits and schedules [32][35]. - "Hey Tuya" can simulate a user's voice in meetings, recording discussions and generating structured meeting notes [37][38]. - It provides nutritional analysis of meals through image recognition, helping users track their dietary habits [40][42]. Group 3: Technological Infrastructure - The underlying architecture of "Hey Tuya" is based on Tuya's Physical AI Engine (PAE), which enables real-time collaboration between AI and physical devices [46][47]. - PAE includes three core engines: Conversational AI Engine, Vision AI Engine, and IoT Intelligence Engine, facilitating complex interactions across various environments [50][51]. - The system employs a long-term memory mechanism to learn user preferences and behaviors over time, enhancing its responsiveness [53][54]. Group 4: Company Background and Market Position - Tuya Smart, founded in June 2014, has evolved from a device connectivity platform to a leading AI cloud service provider, focusing on AIoT integration [56][59]. - As of September 30, 2025, Tuya has over 1.622 million registered developers across more than 200 countries, showcasing its extensive global reach [60]. - The company's AIoT ecosystem encompasses over 3,000 product series across eight categories, positioning it as a key player in the smart home and IoT market [61][62].

编辑部 发自 凹非寺 量子位 | 公众号 QbitAI 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。 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： 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容，建立个人知名度，成为AI领域的意见领袖。 拓展行业人脉 ：与AI领域大咖零距离接触，参与重要科技活动和发布会，拓展行业视野。 获得专业指导 ：应届新人会由主编级编辑出任mento ...

Core Viewpoint - The article discusses the growing anxiety surrounding the "AI bottleneck" as the third anniversary of ChatGPT approaches, questioning whether current technological paradigms can effectively utilize increased computational power to develop models significantly stronger than GPT-4 [1][2]. Group 1: Nature of Intelligence and Its Measurement - Intelligence is fundamentally about energy conversion, where AI has transformed electricity into reusable intelligence over the past decade, but the efficiency of this conversion is now under scrutiny [6]. - The essence of intelligence is not explanation but prediction, characterized by the ability to forecast future states and bear the consequences of those predictions [7][10]. - The current models derive their intelligence primarily from the pre-training phase, which consumes the most energy and computation, raising questions about the stability of intelligence growth with continued computational investment [15][20]. Group 2: Computational Paradigms and Their Limitations - The article emphasizes that the real bottleneck is not the cessation of computational growth but rather the diminishing returns in the relationship between computational power and intelligence growth [22][27]. - It challenges the mainstream narrative by suggesting that pre-training, fine-tuning, and reinforcement learning are fundamentally about gradient computation and parameter updates, rather than distinct methodologies [12][11]. - The success of the Transformer architecture is attributed to its compatibility with GPU systems, which has enabled a stable feedback loop between computational growth, model scaling, and capability enhancement [16][18]. Group 3: Future Directions and Exploration - Future AI infrastructure should focus on the overall scalability of parallel computing systems rather than just single-chip performance, with an emphasis on maintaining or improving the ratio of computational to communication costs [24][25]. - Multiple exploration directions are proposed, including higher precision, advanced optimizers, and more scalable architectures or loss functions, all aimed at ensuring that increased computational investments yield proportional intelligence enhancements [25][26]. - The article concludes that as long as more efficient computational organization methods can be found, the upper limits of intelligence are far from being reached [27].

Core Insights - The article discusses the emergence of numerous keywords in the AI product sector in China by 2025, highlighting the rapid evolution and innovation in AI technologies [4] - The "AI 100" list by Quantum Bit Think Tank aims to evaluate and recognize the top AI products that represent China's AI capabilities [4][12] Group 1: AI 100 List Overview - The "AI 100" list is divided into three main categories: "Flagship AI 100," "Innovative AI 100," and the top three products in ten popular sub-sectors [6] - The "Flagship AI 100" focuses on the strongest AI products of 2025, emphasizing those that demonstrate significant technological breakthroughs and practical application value [7] - The "Innovative AI 100" aims to identify products that are expected to emerge in 2025 and have the potential to lead industry changes in 2026 [8] Group 2: Sub-sector Focus - The ten sub-sectors for the top three products include AI Browser, AI Agent, AI Smart Assistant, AI Workbench, AI Creation, AI Education, AI Healthcare, AI Entertainment, Vibe Coding, and AI Consumer Hardware [9] - This categorization is designed to provide a more precise reflection of the development trends within each sub-sector [9] Group 3: Application and Evaluation Criteria - The evaluation of the "AI 100" list employs a dual assessment system combining quantitative and qualitative measures [13] - Quantitative metrics include user data such as user scale, growth, activity, and retention, with over 20 specific indicators considered [13] - Qualitative assessments focus on long-term development potential, evaluating factors like underlying technology, market space, functionality, monetization potential, team background, and growth speed [13]