部署大模型到底需要多少钱？ 这几乎是所有想把生成式AI引入业务的企业，最焦虑的问题。 选商业API，就要面对持续攀升的token订阅费；自建本地部署，又要承担前期硬件投入与长期运维 开销。既怕超支，又怕浪费，而市场上始终缺乏一个清晰的量化参考框架。 现实的差距惊人：一块1600美元的消费级显卡，就能跑起一个小型开源模型；而高端API的年订阅 账单，却可能飙升至450万美元。 近期，卡内基梅隆大学的研究团队近期给出了系统化的答案。 文章以" 成本测算 "为核心，构建覆盖本地部署（硬件、电费）与商业API（订阅费）的总拥有成本 （TCO）模型，对比Qwen、Llama、Mistral等开源模型与OpenAI GPT-5、Anthropic Claude- 4、Google Gemini 2.5 Pro等商业服务的成本结构，明确不同方案的具体开支，帮企业快速算 清"大模型部署账"： 研究还进一步拆解商业API的定价层级对决策的影响，配套开发在线成本测算工具，能让企业按自身 workload定制分析，为中小企、中型企业、大型企业提供差异化成本指引。 需要说明的是：锦秋基金（公众号：锦秋集；ID：jqcapital）这 ...

今天分享的是：人工智能产业"十四五"复盘与"十五五"展望："两个变局"下的AI要素化跃迁-中国银河 报告共计：49页 《人工智能产业"十四五"复盘与"十五五"展望："两个变局"下的AI要素化跃迁-中国银河》聚焦AI产业在"十四五"期间的发展 成果与"十五五"趋势，围绕技术演进、产业生态、政策支持及应用拓展展开分析。技术层面，大模型成核心突破方向，参数 量增长提速，从2018年GPT-2的15亿参数跃升至2024年GPT-4的1.76万亿参数，2025年呈现"高参数量+轻量化"并行分化，海外 OpenAI、Meta、Google与国内百度、阿里等企业持续推出迭代模型；算力硬件方面，GPU仍占主导（Nvidia占比70%）， ASIC、FPGA等异构芯片加速发展，寒武纪MLU370R-X8等加速卡实现训推一体，海光等企业推动x86与深度计算处理器协 同，液冷等高效散热方案在数据中心普及。产业生态上，AI要素化进程加快，数据经历资源化、资产化、资本化阶段，数据 确权、定价、交易体系逐步完善，政策端2024年数字经济重点工作强调数据要素潜能释放，2025年持续推动标准建设与可信 社会构建；智能体（Agent）生态崛起 ...

Core Viewpoint - The article discusses the competition between two major paradigms in generative AI: Diffusion Models and Autoregressive (AR) Models, highlighting the emergence of Diffusion Language Models (DLMs) as a potential breakthrough in the field of large language models [2][3]. Group 1: DLM Advantages Over AR Models - DLMs offer parallel generation capabilities, significantly improving inference speed by achieving a tenfold increase compared to AR models, which are limited by token-level serial processing [11][12]. - DLMs utilize bidirectional context, enhancing language understanding and generation control, allowing for finer adjustments in output characteristics such as sentiment and structure [12][14]. - The iterative denoising mechanism of DLMs allows for corrections during the generation process, reducing the accumulation of early errors, which is a limitation in AR models [13]. - DLMs are naturally suited for multimodal applications, enabling the integration of text and visual data without the need for separate modules, thus enhancing the quality of joint generation tasks [14]. Group 2: Technical Landscape of DLMs - DLMs are categorized into three paradigms: Continuous Space DLMs, Discrete Space DLMs, and Hybrid AR-DLMs, each with distinct advantages and applications [15][20]. - Continuous Space DLMs leverage established diffusion techniques from image models but may suffer from semantic loss during the embedding process [20]. - Discrete Space DLMs operate directly on token levels, maintaining semantic integrity and simplifying the inference process, making them the mainstream approach in large parameter models [21]. - Hybrid AR-DLMs combine the strengths of AR models and DLMs, balancing efficiency and quality for tasks requiring high coherence [22]. Group 3: Training and Inference Optimization - DLMs utilize transfer learning to reduce training costs, with methods such as initializing from AR models or image diffusion models, significantly lowering data requirements [30][31]. - The article outlines three main directions for inference optimization: parallel decoding, masking strategies, and efficiency technologies, all aimed at enhancing speed and quality [35][38]. - Techniques like confidence-aware decoding and dynamic masking are highlighted as key innovations to improve the quality of generated outputs while maintaining high inference speeds [38][39]. Group 4: Multimodal Applications and Industry Impact - DLMs are increasingly applied in multimodal contexts, allowing for unified processing of text and visual data, which enhances capabilities in tasks like visual reasoning and joint content creation [44]. - The article presents various case studies demonstrating DLMs' effectiveness in high-value vertical applications, such as code generation and computational biology, showcasing their potential in real-world scenarios [46]. - DLMs are positioned as a transformative technology in industries, with applications ranging from real-time code generation to complex molecular design, indicating their broad utility [46][47]. Group 5: Challenges and Future Directions - The article identifies key challenges facing DLMs, including the trade-off between parallelism and performance, infrastructure limitations, and scalability issues compared to AR models [49][53]. - Future research directions are proposed, focusing on improving training objectives, building dedicated toolchains, and enhancing long-sequence processing capabilities [54][56].

Core Viewpoint - The lawsuit filed by adult film giant Strike 3 Holdings against Meta highlights the issue of copyright infringement in the context of AI training, specifically focusing on the unauthorized use of adult film content for developing AI models [2][3]. Group 1: Lawsuit Details - Strike 3 Holdings and Counterlife Media accuse Meta of systematically pirating 2,396 adult films since 2018 for training its AI models, potentially leading to a compensation claim of $359 million (approximately 2.6 billion RMB) [2][3][16]. - The lawsuit marks a significant case as it is the first to address the use of adult film content in training video generation AI, differing from previous copyright disputes involving text and images [2][3]. Group 2: Impact on the Industry - The plaintiffs express concern that Meta's AI could replicate their unique production style at a fraction of the cost, threatening the viability of traditional adult film studios that invest in high-quality production [5][16]. - The lawsuit reveals that Meta allegedly utilized a "tit-for-tat" mechanism on the BT network to not only download but also distribute pirated content, which could significantly enhance download speeds [6][7][8]. Group 3: Evidence and Allegations - The lawsuit cites data from the plaintiffs' VXN Scan tracking system, which indicates that 47 Facebook-registered IPs were involved in illegal distribution, with over 100,000 instances of infringement verified [10][12]. - Meta is accused of constructing a piracy network using "shadow data centers" and non-human usage patterns, suggesting a deliberate strategy to collect training data for AI [11][12][14][15]. Group 4: Legal Proceedings and Reactions - The plaintiffs are seeking a jury trial, asserting that Meta's actions constitute both direct and indirect copyright infringement [16]. - Meta has publicly denied the allegations, but the evidence presented by the plaintiffs is considered substantial, leading to speculation about a potential out-of-court settlement [18].

Core Insights - Baidu leads the domestic AI search industry with 322 million monthly active users, as reported by QuestMobile [2] - The company announced a significant upgrade to its search intelligence framework, integrating multi-modal tools like AI writing and problem-solving capabilities [2][3] - The shift in product presentation, featuring younger product managers, reflects a broader organizational restructuring aimed at enhancing transparency and user engagement [2][4][7] Product Development and Communication - Baidu's recent product launch emphasized a new communication approach, where product managers articulate the logic behind AI-generated content directly to users [4][5] - The team focused on three main narrative pillars: the logic of generated search, fact-checking mechanisms, and AI tool integration capabilities [5] - A collaborative environment was fostered, allowing for user feedback to be incorporated into product iterations, moving away from a closed development model [6] Industry Trends - Other tech companies like ByteDance and Alibaba are also shifting towards younger representatives for AI product presentations, indicating a trend in the industry [3][8] - The communication styles of various companies differ, with some emphasizing technical details while others focus on strategic narratives [8][10] - The choice of spokespersons reflects deeper organizational values regarding transparency, power distribution, and user relationships in the AI product landscape [11]

Core Insights - Musk announced that his AI company xAI will open source its flagship chatbot Grok 2's source code next week, continuing its strategy of promoting transparency in the AI field [1][3] - Grok 2 is built on Musk's proprietary Grok-1 language model and is positioned as a less filtered and more "truth-seeking" alternative to ChatGPT or Claude, with the ability to pull real-time data from the X platform [1][3] - The chatbot offers multimodal capabilities, generating text, images, and video content, and is currently available to X Premium+ subscribers [3] Group 1 - The core competitive advantage of Grok 2 lies in its deep integration with the X platform, allowing it to respond uniquely to breaking news and trending topics [3] - The open-sourcing of Grok 2 will enable developers and researchers to access its underlying code and architecture, facilitating review, modification, and further development based on this technology [3] - This strategic move may strengthen Musk's business network and create integration possibilities among his companies, including Tesla, SpaceX, Neuralink, and X [3] Group 2 - The decision to open source Grok 2 aligns with the industry's trend towards open-source AI models, positioning xAI as a counterbalance to major AI companies like OpenAI, Google, and Anthropic [4] - However, Grok's relatively lenient content restriction policies have previously sparked controversy, raising concerns about the potential amplification of risks associated with open-sourcing [4] - There are industry worries regarding the misuse of this technology in sensitive areas such as medical diagnostics or autonomous driving systems, which could lead to severe consequences [4]

Group 1: Development Trends of Large Models - The rapid development of large models in the AI field is transforming the understanding of AI and advancing the dream of AGI (Artificial General Intelligence) from science fiction to reality, characterized by two core trends: continuous leaps in model capabilities and increasing openness of model ecosystems [1][4]. - Continuous improvement in model capabilities is achieved through iterative advancements and technological innovations, with examples like OpenAI's ChatGPT series showing significant enhancements in language understanding and generation from GPT-3.5 to GPT-4 [1][2]. - The breakthrough in multimodal capabilities allows models to natively support various data types, including text, audio, images, and video, enabling more natural and rich interactions [2][3]. Group 2: Evolution of AI Applications - The rapid advancement of large model capabilities is driving profound changes in AI application forms, evolving from conversational AI to systems capable of human-level problem-solving [5][6]. - The emergence of AI agents, which can take actions on behalf of users and interact with external environments through tool usage, marks a significant evolution in AI applications [6][8]. - The recent surge in AI agents, both general and specialized, demonstrates their potential in solving a wide range of tasks and enhancing efficiency in various domains [8][9]. Group 3: AI Middleware Opportunities and Challenges - AI middleware is emerging as a crucial layer that connects foundational large models with specific applications, offering opportunities for agent development efficiency, context engineering, memory management, and tool usage [13][19][20]. - The challenges faced by AI middleware include managing complex contexts, updating and utilizing persistent memory, optimizing retrieval-augmented generation (RAG) effects, and ensuring safe tool usage [26][29][30]. - The future of AI middleware is expected to focus on scaling AI applications, providing higher-level abstractions, and integrating AI into business processes, ultimately becoming the "nervous system" of organizations [39][40].

Core Insights - DeepSeek has emerged as a significant player in the AI large model landscape, driving widespread adoption among individuals, enterprises, and governments due to its low cost, high performance, and open ecosystem [1] - The large-scale application of AI models is crucial for rapid iteration and development in China, but it faces challenges such as low stability of underlying frameworks, barriers to cross-industry integration, and limited ecological support [1] - The current strategic opportunity period for AI development in China necessitates efforts in technological breakthroughs, industry adaptation, and risk warning to create a conducive environment for AI model applications [1] Group 1: Challenges in AI Model Application - The complexity and lack of interpretability in AI models, particularly deep neural networks, pose significant challenges for industry applications, leading to unreliable outputs and "hallucinations" [2] - Specific industries, such as manufacturing, face adaptation difficulties due to the complex and multimodal nature of their data, which existing models struggle to accurately interpret [3] - The fragmented approach to integrating AI models across industry chains increases long-term collaboration costs, as many companies overlook the importance of coordinated applications [4] Group 2: Economic Impact and Efficiency - The high operational costs associated with AI models, such as DeepSeek-R1, can lead to significant financial losses for companies, highlighting the need for cost-effective solutions [4] - Data integration across the supply chain can dramatically enhance operational efficiency, with reported improvements in order response speed and anomaly handling when fully integrated [5] - The rapid penetration of AI models into industries may lead to exponential increases in the costs for latecomers, limiting their ability to catch up with established players [6] Group 3: Regulatory and Ethical Considerations - The current ecosystem for AI model application is underdeveloped, with weak foundations in data, standards, and ethics, which could hinder the promotion of AI models [6] - The scarcity of high-quality training data, particularly in sensitive areas like healthcare, poses a significant barrier to effective AI model training and deployment [6] - The lack of a robust standard system for addressing ethical, legal, and social implications of AI models is a critical issue, as highlighted by the EU's AI regulatory draft [6][7]

Core Viewpoint - The article discusses a novel training paradigm for large language models (LLMs) called POET (Reparameterized Training via Orthogonal Equivalence Transformation), which aims to enhance training efficiency and stability based on first principles [2][3]. Group 1: POET Methodology - POET introduces structural reparameterization of each neuron by incorporating two learnable orthogonal matrices and a fixed random weight matrix, maintaining the singular value distribution of weights during training [3][11]. - The method combines singular value invariance with minimal hyperspherical energy, providing a new paradigm that offers both physical interpretability and generalization capability for large model training [3][11]. - POET's training process is designed to stabilize the optimization process and significantly improve model generalization performance [3][11]. Group 2: Advantages of POET - POET maintains the spectral properties of the weight matrix throughout training, ensuring that the singular values remain consistent with the randomly initialized matrix [17]. - The method allows for efficient parameter control and avoids the issue of excessively large singular values that can occur in standard LLM training [17]. - Two new initialization strategies, normalized Gaussian initialization and uniform spectrum initialization, are proposed to ensure bounded singular values in the generated weight matrices [17]. Group 3: Training Dynamics and Performance - The article presents experimental results demonstrating POET's superior performance in training large language models, including improvements in perplexity and training efficiency compared to traditional methods like AdamW [20][24]. - POET's training process is divided into three phases: conical shell searching, stable learning on the conical shell, and final adjusting, which reflects the evolution of the orthogonal matrices during training [40][41]. - The use of a fully stochastic sampling approach in POET allows for a significant reduction in memory costs compared to traditional methods, enhancing scalability [26][27].

Core Viewpoint - The article discusses the evolving landscape of large models in autonomous driving, highlighting the focus on lightweight solutions, hardware adaptation, knowledge distillation, and advanced reasoning paradigms like CoT and VLA+ reinforcement learning as key areas for future development [1][2]. Group 1: Course Introduction - The course aims to explore cutting-edge optimization methods for large models, focusing on parameter-efficient computation, dynamic knowledge expansion, and complex reasoning [2]. - It addresses the core challenges in model optimization, including pruning, quantization, retrieval-augmented generation (RAG), and advanced reasoning paradigms [3]. Group 2: Problems Addressed by the Course - The course provides a systematic understanding of large model knowledge, helping students build a coherent theoretical framework [3]. - It assists students in combining theoretical knowledge with practical coding skills, enabling them to replicate research papers and develop new models [3]. - The course offers guidance on writing and submitting academic papers, addressing common challenges faced by students [3]. Group 3: Enrollment Information - The course limits enrollment to 6-8 students per session [4]. - It targets individuals with a background in deep learning or machine learning, familiarity with Python, and a passion for research [6]. Group 4: Course Outcomes - Participants will gain insights into classic and cutting-edge papers in the field, enhancing their understanding of key algorithms and principles [9]. - The course includes a structured approach to writing and revising academic papers, culminating in the production of a draft [9]. Group 5: Course Structure - The course spans 12 weeks of online group research followed by 2 weeks of paper guidance and a 10-week maintenance period [9]. - It covers various topics, including model pruning, quantization, and advanced reasoning techniques, with a focus on practical applications [19].