Shannon 信息论 本节先归纳一下 Shannon 的主要结论和方法论启示。下图是一般通信系统的原理图。 作者 | 白铂 博士 白铂 博士 华为 2012 实验室理论研究部主任 信息论首席科学家 引言 在本系列的第二篇《 信号处理篇 》中，我们引入了一些信息论的概念和方法来理解语义嵌入/向量化。本篇将完全从信息论的角度出发，深入解读原论文，探讨大 模型背后的第一性原理 ¹ 。 1948 年，Shannon 发表了题为 A Mathematical Theory of Communication 的划时代论文，奠定了现代数字通信的理论基础，推动了人类迈向信息时代 ² 。论文的主要 目标是用 数学方法 解决有噪声的数字通信系统的可靠传输问题。以此为起点，Shannon 及后来的专家学者建立了一套完备的数学框架与理论体系，这便是后来众所 周知的 信息论 。1949 年，Weaver 与 Shannon 合著了一篇论文，文中明确将通信问题分为三个层级 ³ ： 受到 Shannon 方法论的启发，本文尝试从推理的视角出发探讨大模型的可解释理论。我们发现，只要将 Shannon 的理论从以 BIT 为中心转换为以 T ...

Token Issuance - PM (预测市场平台) 有 99% 的概率发行 Token [1] - Token 发行价格约为 150 亿 (15 Billion) 美元 [1] - Token TGE (Token Generation Event) 一年后解锁 [1] - TGE 时间预计在明年 S1 赛季或世界杯之后 [1] Institutional Involvement - 两家香港券商正在销售 PM 的 Token [1] - 其中一家券商为胜利证券 [1] Token Details - 出售的 Token 不包含股权 [1]

Summary of Key Points from the Conference Call Industry Overview - The conference focuses on the AI computing power sector, highlighting the significant growth driven by inference and training demands, emphasizing the importance of large models and applications rather than solely relying on industry chain data [1][2][3]. Core Insights and Arguments - **AI Computing Demand**: The demand for computing power is closely linked to the volume of tokens, with increasing computational needs driving this trend. The growth in overseas computing companies cannot be explained solely by traditional performance metrics, necessitating a deeper analysis of how token volume influences computing demand and future trends [1][4]. - **Agentic AI Concept**: Introduced as a new paradigm derived from reasoning models, agentic AI emphasizes task distribution, execution, and planning to achieve specific goals, capable of handling complex or simple tasks through a multi-step process [1][6]. - **GTC Conference Attendance**: The GTC conference saw a 50% increase in attendance compared to the previous year, with a notable rise in AI industry participants, indicating the growing importance of the event for the AI sector [3]. - **Token Explosion**: The global token volume is experiencing explosive growth, significantly impacting computing demand. The relationship between token consumption and computing power is complex and non-linear, with a potential for exponential increases in demand [12][17][21]. Important but Overlooked Content - **Skin Law**: Huang Renxun introduced the concept of "skin law," which describes the inflation of computing demand across three phases: pre-training, post-training, and test time, each contributing to increased computational needs [8][10]. - **Future Drivers of Computing Demand**: The shift from CPU to GPU architectures and the need for capital investment in software rather than just human resources are identified as key factors driving future computing demand [34][35]. - **Market Dynamics**: The competition among major tech companies to enhance user experience through faster response times and accurate outputs is leading to increased investments in computing power, indicating a shift towards a model where software relies heavily on computational resources [26][38]. Market Predictions - **Data Center Market Growth**: The data center market is expected to exceed $1 trillion by 2028, with 2025 being a pivotal year for rapid growth in demand [32]. - **GPU Demand**: Major cloud service providers have shown significant demand for GPUs, purchasing millions of units, indicating sustained growth in computing needs [31]. Conclusion - The AI computing power sector is at a critical juncture, with emerging paradigms like agentic AI and the explosive growth of token consumption reshaping the landscape. Understanding these dynamics is essential for accurately predicting future trends and making informed investment decisions in the sector [5][43][45].

Summary of Conference Call Records Industry and Company Overview - The conference call discusses the development of Multi-Channel Platforms (MCP) in the AI application sector, particularly focusing on Alibaba's initiatives and products like DingTalk and Quark [1][2][3]. Key Points and Arguments MCP Development and Market Position - Domestic MCP development lags behind international counterparts, particularly in multi-task planning and ecosystem construction. International AI agents like Manners and CodeBot can independently execute complex tasks, while domestic applications are still developing [1][2]. - The Manas super agent shows significant token consumption when handling complex tasks, with daily token usage reaching 350 billion to 450 billion, indicating strong market demand [1][5]. - Zinus, priced at $199 to $299 per month, has received positive market feedback, with similar daily token usage as Manas, but may face future price competition [1][6]. Strategic Positioning of DingTalk and Quark - DingTalk is positioned as a ToB AI entry application focusing on commercialization and revenue, while Quark targets ToC with an emphasis on daily active user growth and token consumption [1][7]. Future Projections and Cost Adjustments - Alibaba's Qianwen model is expected to reduce costs by 30% to 50% by 2025 to enhance market competitiveness and encourage more enterprises to adopt the model for business optimization [1][9]. Token Consumption Trends - Alibaba's token consumption has shown exponential growth, with daily usage projected to reach 10 trillion by the end of Q2 2025, driven by both internal and third-party models [3][12]. MCP Protocol and Application Integration - The MCP protocol serves as a standardized interface that facilitates the integration and deployment of AI agents across various applications, enhancing operational efficiency [17][18]. Additional Important Insights Challenges in Domestic MCP Adoption - The slow adoption of MCP in China is attributed to model capability issues, ecosystem limitations, and conservative strategies among major tech companies [2][4]. - The current penetration rate of AI applications in enterprises is low, with many companies still in a wait-and-see approach [22][26]. Future Trends in AI Agents - The future of AI agents is expected to see a surge in capabilities, leading to broader applications across various sectors, which will drive digital transformation and innovation opportunities [19][20]. Knowledge Management in AI - Knowledge structuring and tuning are critical components of AI capabilities, as they involve processing complex data types that current models struggle to interpret [30][31]. Market Dynamics and Competition - The competitive landscape is evolving, with companies like Tencent focusing on consumer applications while Alibaba and ByteDance compete in the ToB space [21][22]. This summary encapsulates the key discussions and insights from the conference call, highlighting the current state and future potential of MCP and AI applications within Alibaba's ecosystem.