如今，大模型蓬勃发展，有没有指标来衡量AI大模型的"智力能力水平"？近日，清华大学研究团队提出 了大模型的密度法则，相关论文刊发于自然子刊《自然·机器智能》上。密度法则揭示大语言模型的最 大能力密度随时间呈指数级增长，2023年2月至2025年4月，约每3.5个月翻一倍。 计算机领域的"摩尔定律"大家已经耳熟能详：芯片上可容纳的晶体管数量，每隔一段时间就会翻一番。 计算机的强大，不是因为芯片变得像房子一样大，而是因为它在指甲盖大小的面积上集成了天文数字的 计算单元。清华大学计算机科学与技术系助理研究员肖朝军告诉科技日报记者，大模型的智力能力水平 应该也有一个指标，这就是"能力密度"。 研究的核心假设是，采用相同制造工艺、充分训练的不同尺寸模型，其能力密度相同。正如芯片行业通 过提升电路密度实现了计算设备的小型化和普惠化，大模型也在通过提升能力密度实现高效化发展。 肖朝军说，过去，在规模法则的指引下，大家关心一个大模型的"块头"（参数量）有多大，越大就越智 能，就像关心一个举重运动员的体重，体重越重，力量越大。现在，密度法则从另一个视角揭示了大模 型"高效发展"的规律——我们更应该关心它的"能力密度"，即每一单 ...

公众号记得加星标⭐️，第一时间看推送不会错过。 摩尔定律指出，集成电路上的晶体管数量往往每两年翻一番，这一规律的实现很大程度上得益于光刻 技术的进步：光刻技术是一种在硅片上制作微观图案的技术。晶体管尺寸的不断缩小——从20世纪70 年代初的约10000纳米缩小到如今的约20纳米-60纳米——得益于能够制作越来越小图案的光刻方法 的开发。光刻技术的最新进展是采用极紫外 (EUV) 光刻技术，该技术使用波长为 13.5 纳米的光在芯 片上创建图案。 众所周知，极紫外光刻机仅由荷兰ASML公司一家生产，因此，谁能使用这些机器已成为一个重要的 地缘政治问题。然而，尽管机器由ASML制造，但使其得以实现的绝大部分研究工作却是在美国完成 的。美国研发领域一些最负盛名的机构——例如DARPA、贝尔实验室、IBM研究院、英特尔和美国 国家实验室——投入了数十年时间和数亿美元的资金，才使得极紫外光刻技术成为可能。 那么，为什么在美国付出如此多的努力之后，最终实现 EUV 商业化的却是荷兰的一家公司呢？ 半导体光刻技术的工作原理 简而言之，半导体光刻技术的工作原理是利用掩模将光选择性地投射到硅片上。当光穿过掩模（或在 极紫外光刻 ...

Core Insights - The article discusses the challenges and new perspectives in the development of large models, particularly focusing on the "Density Law" proposed by Tsinghua University, which indicates an exponential growth in the maximum capability density of large language models from February 2023 to April 2025, doubling approximately every 3.5 months [1][8]. Group 1: Scaling Law and Density Law - Since 2020, OpenAI's Scaling Law has driven the rapid development of large models, but by 2025, the sustainability of this path is in question due to increasing training costs and the nearing exhaustion of publicly available internet data [1]. - The Density Law provides a new perspective on model development, suggesting that just as the semiconductor industry improved chip density, large models can achieve efficient development through increased capability density [3][4]. Group 2: Implications of Density Law - The research team hypothesizes that different-sized models, when trained adequately, will have the same capability density, establishing a baseline for measuring other models [4]. - The Density Law indicates that the inference cost for models of the same capability decreases exponentially over time, with empirical data showing that the API price for models like GPT-3.5 has decreased by 266.7 times over 20 months, roughly halving every 2.5 months [7][8]. Group 3: Acceleration of Capability Density - An analysis of 51 recent open-source large models revealed that the maximum capability density has been increasing exponentially, with a doubling time of approximately 3.5 months since 2023 [8][9]. - Following the release of ChatGPT, the capability density has increased at a faster rate, doubling every 3.2 months compared to every 4.8 months prior, indicating a 50% acceleration in density enhancement [9][10]. Group 4: Limitations of Model Compression - The research found that model compression algorithms do not always enhance capability density, as many compressed models performed worse than their original counterparts due to insufficient training [11][13]. Group 5: Future Prospects - The intersection of chip circuit density (Moore's Law) and model capability density (Density Law) suggests that edge devices will be able to run higher-performance large models, leading to explosive growth in edge computing and terminal intelligence [14]. - Tsinghua University and the Mianbi Intelligence team are advancing high-density model development, with models like MiniCPM and VoxCPM gaining global recognition and significant download numbers, indicating a trend towards efficient and low-cost models [16].

Core Insights - The article discusses the evolution of large models in AI, highlighting the challenges posed by increasing training costs and the potential end of pre-training as currently understood by 2025 [1] - It introduces the "Densing Law" from Tsinghua University, which suggests that the maximum capability density of large language models is growing exponentially, doubling approximately every 3.5 months from February 2023 to April 2025 [1] Group 1: Scaling Law and Densing Law - The Scaling Law proposed by OpenAI indicates that larger model parameters and training data lead to stronger intelligence capabilities, but sustainability issues arise as training costs escalate [1] - The Densing Law provides a new perspective on model development, revealing that the capability density of large models is increasing exponentially over time [1][6] Group 2: Key Findings from Research - The research team analyzed 51 recent open-source large models and found that the maximum capability density has been doubling every 3.5 months since 2023, allowing for the same intelligence level with fewer parameters [9] - The inference cost for models of the same capability is decreasing exponentially over time, with empirical data showing that the API price for GPT-3.5 has dropped by 266.7 times over 20 months, approximately halving every 2.5 months [12] Group 3: Implications of Densing Law - The capability density of large models is accelerating, with a notable increase in the rate of doubling from 4.8 months before the release of ChatGPT to 3.2 months afterward, indicating a 50% acceleration in density enhancement [14] - Model compression algorithms do not always enhance capability density, as many compressed models have lower density than their original counterparts, revealing limitations in current compression techniques [16] - The intersection of chip circuit density (Moore's Law) and model capability density suggests significant potential for edge computing and terminal intelligence, leading to a transformative shift in computational accessibility from cloud to edge devices [18] Group 4: Future Developments - Tsinghua University and Mianbi Intelligence are advancing high-density model research based on the Densing Law, releasing several efficient models that have gained global recognition, with downloads nearing 15 million and GitHub stars approaching 30,000 by October 2025 [20]

摩尔定律是由英特尔的共同创办人高登·摩尔在1965年提出，并非自然科学定律，而是趋势预测。 摩尔定律是指在相同面积的积体电路（芯片）上，可容纳的电晶体数量大约每18个月到24个月会增加一 倍，提升效能，降低成本。 半导体先进制程推进，业界不时传出摩尔定律将死或终结，不过，ASML台湾暨东南亚区客户行销主管 徐宽成昨（19）日指出，摩尔定律走不下去是不对的，预期未来15年制程蓝图会继续推进。此外，记忆 体应用方面不论NAND或DRAM都需要键合，改善异质整合。 他提到，新的芯片架构上，先进封装技术利用"矽中介层"堆叠下，新的机台曝光效率升级对ASML是重 要里程碑，ASML在后段封装较少为外界关注，但实际上已推出XT:260设备，并于今年第3季首度出货 应对客户需求，在过往芯片微缩以外，放大面积造成的生产效率提升也是另一个议题。 ASML旗下EUV协助芯片制造商进行线宽微缩包含Low NA EUV（NXE:3800E）：提升先进制程客户的 生产效率和设备可用以及High NA EUV：具备更高成像品质与简化流程（单次曝光）的优势。 ...

例如，对一段生产工艺数字化时，系统除了需要录屏、收音并抓取CNC操作日志外，还可配合传感器 获取工具寿命和工件良率，并结合上述五种模态数据降低知识入库门槛。 当对不同行业、不同业务线积累的数据足够庞大后，即可通过"智能数据套件"对其进行统一封装，带 着"业务属性"去检查数据的一致性、唯一性，通过"智能数据体检"后，将数据治理结果与国标、企业标 准及历史数据关联，构建出专属于指定工厂的"工艺知识图谱"，进而提升数据质量，改善大模型应用效 果。 除了数据积累与运用层面的提升，涉及多个生产环节的智能体协同，也是当前工业大模型应用的另一重 点提升方向。 据行业测算，当前企业内部AI应用的发展速度，很有可能符合"摩尔定律"的规律，即每隔18个月，AI相 关应用的数量将会翻一倍，这对企业管理者、执行者与中控平台对不同智能体的协同调度能力产生了考 验。 近日，2025"雅典娜杯"两岸青年人才创新创业大赛决赛在浙江绍兴举行，在两岸300支参赛队伍中选拔 出的19支队伍展示了其创新产品与产业应用思路。 作为为大赛提供数智原生底座的数字化服务商，鼎捷数智执行副总裁刘波在大赛现场的主题演讲中表 示，AI大模型时代，在为创业者们搭建 ...

存储设备公司成长性："价格周期"和"技术周期"共振 带来高斜率 20251117 摘要 全球半导体设备市场由少数头部供应商主导，薄膜沉积领域通常有三家 左右，国内未来预计有两到三家主要公司。泛林集团通过拓展 CVD、ALD 等薄膜沉积技术，全球市占率接近 20%，干法刻蚀方面接 近 40%-50%。 泛林集团收入从 2014 年的 48.6 亿美元增长至 2024 年的 162 亿美元， 年复合增长率 12.8%，利润从 7.2 亿美元增至 42.9 亿美元，年复合增 长率约 20%。预计 2024 年至 2028 年收入复合增长率仍将保持在 10%左右，毛利率计划到 2028 年达到 50%左右。 3D NAND 技术推动了对刻蚀和薄膜沉积设备的需求，范林在 NAND 领 域的收入从 2014-2015 年的 16.3 亿美元增长到 2022 年的 74.7 亿美 元。长江存储被制裁后，中微公司和拓荆公司正在填补市场缺口。 存储器未来发展趋势是 NAND 向更高层数堆叠发展，DRAM 从平面结构 向 3D 结构演进，这将增加对刻蚀与薄膜沉积等装备的需求。逻辑芯片 制程不断演进，单晶圆片资本开支密度大幅提升。 ...

Core Viewpoint - The recent tensions between the Netherlands and China, highlighted by the Nexperia incident, underscore the fragility of the semiconductor supply chain and the importance of dialogue to prevent escalation [2][3]. Group 1: Nexperia Incident - The Nexperia situation illustrates the critical nature of the semiconductor industry and the ecosystem's vulnerability, emphasizing the need for responsible actions and dialogue among stakeholders [2]. - Nexperia, owned by China's Wingtech Technology, primarily supplies power control chips to automotive manufacturers like BMW and Volkswagen. The Dutch government's sudden takeover of the company's key decision-making authority led to retaliatory actions from Beijing, disrupting the supply of critical automotive components [2]. - Recent developments indicate a thawing of relations, with China resuming some exports of Nexperia chips and the Dutch government planning to send a delegation to seek a mutually acceptable solution [2]. Group 2: ASML's Position - ASML, as the sole producer of advanced extreme ultraviolet (EUV) lithography machines, plays a pivotal role in the semiconductor industry, providing equipment to major companies like TSMC and Intel [3][5]. - The company reported a net sales figure of €28.3 billion (approximately $33.1 billion) for 2024, with a market capitalization exceeding €350 billion (around $406 billion), making it the most valuable company in Europe [5]. - ASML's success is attributed to significant investments in EUV technology, which required breakthroughs in physics, optics, and materials science, supported by direct investments from major industry players like Intel, TSMC, and Samsung [6]. Group 3: Leadership and Culture - ASML's CEO, Christophe Fouquet, emphasizes the company's strong sense of responsibility within the industry and the importance of long-term vision and restraint in leadership [6][8]. - The company fosters a culture of openness and collaboration, which is seen as a cornerstone of its innovation, allowing employees to communicate freely across all levels [8]. - The leadership style at ASML is characterized by humility and a focus on creating value for customers, recognizing the broader impact of their work on the world [8][9]. Group 4: Geopolitical Context - Geopolitical factors increasingly influence ASML's future, with export controls, subsidies, and strategic alliances playing a critical role alongside technological advancements [8]. - The company recognizes the necessity of adapting to macroeconomic and geopolitical uncertainties while maintaining strong relationships with customers and entering vital markets [9].

公众号记得加星标⭐️，第一时间看推送不会错过。 遵循摩尔定律，集成电路中晶体管尺寸的持续缩小——微芯片上的晶体管数量大约每两年翻一番—— 是一项非凡的工程壮举，突破了基础物理学的极限。晶体管是关键元件，它通过开关电流来调节端子 （导电连接点）之间的电流。当晶体管尺寸减小时，开关速度加快，从而使集成电路能够更快地处理 信息。 然而，随着晶体管尺寸缩小到纳米级，互连线——连接晶体管和微芯片上其他电路元件的金属导线 ——成为处理速度的主要瓶颈。因此，提升下一代电子设备集成电路的性能不能仅仅依靠缩小晶体管 尺寸来实现，还需要开发新型互连材料。 互连线用于将信号从一个电路元件传输到另一个电路元件。信号在导线中传输所需的时间称为电阻- 电容 (RC) 时间延迟，它取决于互连材料的固有电阻和周围介质元件（一种可被电场极化的电绝缘 体）的电容。因此，铜作为导电性最好的金属之一，一直是互连线的标准材料。然而，仿真结果表 明，目前使用的最小互连线（宽度约为 15 nm）的 RC 时间延迟可能高达晶体管开关速度的 20 倍。 互连线的RC时间延迟为何如此之大？铜的电阻会随着尺寸的减小而增大。金属内部电子的运动是造 成这种尺寸效应影 ...

Core Insights - The AI industry is facing a critical juncture where the marginal returns of large models are diminishing, leading to a shift from a parameter race to an efficiency revolution [1][4][11] - Training costs for cutting-edge AI models have skyrocketed, with expenses for models like GPT-4 exceeding $100 million and approaching $1 billion for the most advanced models, making it a domain dominated by capital-rich giants [1][2] - Smaller models, such as DeepSeek R1-0528, are demonstrating that they can outperform larger models while significantly reducing operational costs, indicating a potential paradigm shift in AI development [2][4] Industry Trends - The transition from "Cloud First" to "Local First" is underway, as the limitations of Moore's Law have prompted tech giants to seek new paths for efficiency and performance [5][6][7] - Companies like Apple and NVIDIA are innovating in chip design and architecture to adapt to the new landscape, focusing on vertical integration and parallel processing capabilities [6][7] - The emergence of small language models (SLMs) is challenging the dominance of large language models (LLMs), with SLMs achieving comparable or superior performance in various tasks at a fraction of the cost [2][4] Challenges in AI Deployment - The current AI landscape faces three major pain points: lack of closed-loop productivity experiences, high token costs limiting application scalability, and network dependency restricting usage scenarios [9][10] - Users are increasingly concerned about data privacy and the inability to utilize AI in offline environments, which has led to a demand for local AI solutions [10][11] GreenBitAI's Innovations - GreenBitAI is pioneering a Local Agent Infra that allows for professional-grade AI applications to run entirely offline on consumer-grade hardware, addressing privacy concerns and operational efficiency [15][32] - The company has developed a series of low-bit models that maintain high accuracy while significantly reducing computational requirements, demonstrating the viability of local AI solutions [19][22] - GreenBitAI's product, Libra, showcases the potential for local AI applications to handle complex tasks traditionally reserved for cloud-based solutions, marking a significant advancement in the field [32][33] Market Potential - The global market for AI PCs is projected to grow significantly, with estimates suggesting that by 2026, AI PCs will account for over 55% of the total PC market [35][36] - GreenBitAI aims to capture a substantial share of the emerging local AI market, positioning itself as a foundational infrastructure provider for future AI applications [37][38]