大模型能力技术培训 让数据智能像水电 样简单 语言模型发展历程 大语言模型：包含百亿或更多参数的语言模型 参考文献：https://arxiv.org/abs/2303.18223 • 上世纪90年代：语言模型出现，统计学方法，使用前面的词预测下一个词 • 2003年： Bengio 《A Neural Probabilistic Language Model》 ，首度将深度学习思想融入语言模型 • 2018年： Google提出Transformer神经网络架构， 并通过大量文本训练理解语言规则和模式 • 国外：GPT-3（175B） 、GPT-4 、PaLM（540B） 、Galactica 和 LLaMA 等 • 国内：ChatGLM、文心一言 、通义千问 、讯飞星火等 • 大语言模型和小语言模型（如GPT2）采用相似的架构和预训练任务，但是能力截然不同（涌现能力） • 涌现能力使得大语言模型只使用很少的样本就可以处理全新的任务 对技术领域的影响 对商业领域的影响 参考文献：https://arxiv.org/abs/2303.18223 • 自然语言处理：理解和生成文本，意图理解 、写文章 、 回答问 ...

过去一年，在AI核心业务推动下，科大讯飞出海提速，交出了一份稳健增长的成绩单。2026年2月13 日，在公司年会上，科大讯飞董事长刘庆峰、总裁吴晓如分别发表演讲，系统总结2025年经营成果，并 部署2026年战略重点。 2025年，科大讯飞在资源受限的情况下，不仅实现了归母净利润和经营性现金流的扎实增长，更通过攻 克国产卡训练难关，推动星火大模型持续迭代，近日发布的星火X2大模型在数学、推理、医疗等关键 能力上对标国际一流模型水平。此外，公司去年大模型相关项目中标额超23亿元，再度蝉联大模型"标 王"。 根据科大讯飞此前发布的2025年度业绩预告，公司预计2025年归属于上市公司股东的净利润为7.85亿元 至9.5亿元，较上年同期增长40%至70%；扣除非经常性损益后的净利润为2.45亿元至3.01亿元，同比增 长30%至60%。 在回顾2025年业绩时，吴晓如在年会上披露，2025年科大讯飞经营回款超270亿元，经营性现金流再创 新高达30亿元、同比增长28%。产品端同样亮眼：开放平台大模型日均调用量较上年猛增45倍，毛利增 长244%，单业务毛利首次突破10亿元；讯飞听见用户规模突破1亿，AI营销海外毛利 ...

Core Viewpoint - The domestic AI large model sector is experiencing a vigorous development characterized by simultaneous "technological breakthroughs and commercial implementation," shifting from mere parameter competition to a fierce contest for actual value creation [1]. Group 1: General Large Models - Alibaba released its flagship reasoning model Qwen3-Max-Thinking in January 2026, achieving multiple global records with over one trillion parameters and a new "test-time expansion" mechanism that significantly enhances reasoning efficiency and performance, scoring 58.3 in the "Human Last Test" (HLE), surpassing GPT-5.2-Thinking and Gemini 3 Pro [3]. - Baidu launched its native multimodal model Wenxin 5.0 in December 2025, featuring over 2.4 trillion parameters and topping the LMArena text leaderboard with a score of 1451, marking it as the best in China [3]. - ByteDance's Doubao model version 1.8, released in December 2025, optimized for multimodal agent scenarios, has shown impressive growth with a daily token usage exceeding 50 trillion [3]. Group 2: Vertical Models - Baidu's Wenxin 5.0 has been applied in the education sector, collaborating with publishers to create "AI picture books" for special needs children [5]. - iFLYTEK continues to leverage its voice interaction advantages with its Xinghuo model, achieving significant commercial success in medical and government sectors through voice transcription and meeting minutes generation [5]. - Tencent's Mixuan model has made breakthroughs in 3D generation, with the open-source Mixuan World model 1.1 enabling rapid 3D world creation for industries like gaming and design [5]. - Baichuan Intelligent announced the full opening of the M3 Plus API, with its Baichuan-M3 model ranking first globally in the HealthBench medical evaluation with a score of 65.1, surpassing GPT-5.2 [5]. Group 3: Hardware Developments - The year is marked as the "Battle of Hundreds of Chips," with companies like Birun Technology delivering over 12,000 GPU chips and holding unfulfilled orders worth 822 million yuan as of December 2025 [8]. - Baidu's Kunlun chip is preparing for an IPO on the Hong Kong Stock Exchange, having released multiple AI processing units and achieving significant shipment volumes [8]. - Alibaba's Pingtouge plans to pursue a separate listing, alongside other companies like Moer Thread and Muxi Integrated Circuit, which are also preparing for IPOs [9]. - Current policies are supportive of the AI hardware sector, with overall positive trends in the market, exemplified by the strong performance of the Sci-Tech Semiconductor ETF (588170) [10][11].

Core Insights - The AGI-Next summit confirmed a significant divergence in the Chinese large model industry, moving beyond a simple ToB vs. ToC debate to different approaches in realizing the value of intelligence [1][2] - The divergence is characterized by two paths: one focusing on embedding models into production processes for productivity certainty (ToB), and the other emphasizing real-life scenarios for user experience and scale (ToC) [1] ToB vs. ToC Divergence - The core discussion at the summit revolved around the clear differentiation between ToB and ToC, with experts agreeing that user perceptions of AI value differ significantly between the two [4] - ToC users often do not require extreme intelligence, using AI more like an enhanced search engine, while ToB users are willing to pay a premium for high-performing models that enhance productivity [4] - The business models also differ: ToC requires tight integration of models and products, while ToB tends to favor a layered approach where strong models attract application layer companies [4] Market Dynamics - The recent IPOs of companies like Zhiyuan and MiniMax exemplify the divergence, with Zhiyuan focusing on ToB and MiniMax on ToC, indicating a clear trend in the industry [7] - In the ToB space, companies like iFlytek are integrating AI into core production processes, demonstrating the practical application of high intelligence in enhancing productivity [7][8] - The financial sector is also seeing the emergence of ToB models, such as the intelligent risk control platform launched by Tongyi Qianwen, which improves credit risk identification and loan approval processes [8] Future Directions - The ToC path faces challenges in capturing user needs and ensuring sustainable commercialization, with competition extending beyond large model companies to traditional internet platforms [14] - The evolution of ToC is shifting towards extreme personalization and deep scenario engagement, focusing on high-frequency, strong-demand areas like education and content creation [14] - The relationship between ToB and ToC is expected to become more complementary, with ToB needing ToC's scenario understanding and ToC relying on ToB's foundational models [15] Global Competitive Landscape - The differentiation in the Chinese large model industry will significantly impact its international competitiveness, with experts noting that while Chinese companies show strong execution, they still lag in leading new paradigms [16] - The potential for Chinese companies to innovate in the application layer is highlighted, suggesting that they may lead in ToC product forms globally [16] Conclusion - The divergence in the large model industry is a sign of maturity, with ToB focusing on efficiency and endurance, while ToC emphasizes experience and scale, with time being the ultimate judge of success [18]

Core Insights - The AI industry is experiencing a significant transformation, moving from conceptual hype to a focus on practical value and commercial applications, particularly in the realm of large models [1][2][3] Group 1: Industry Trends - The large model sector in China is witnessing a "Matthew effect," where resources are increasingly concentrated among leading companies, marking the end of the "hundred model battle" [3] - Major players like DeepSeek and ByteDance are leading the charge with innovative products, such as DeepSeek-R1 and the Doubao model, which have significantly lowered the barriers for AI application in enterprises [3][4] - The introduction of AI applications in various sectors, including health care, education, and finance, is rapidly increasing, with over 200 new applications incorporating AI features [7] Group 2: Technological Advancements - The development of large models has transitioned from mere technical validation to practical tools that enhance productivity in workplaces, as evidenced by user experiences in financial analysis and software development [6][7] - The AI hardware market is gaining momentum, with significant investments in AI glasses and smartphones, indicating a shift towards integrating large models with hardware for improved human-computer interaction [8][9] Group 3: Market Performance - The AI sector in the A-share market has seen a cumulative increase of over 35% in 2025, reflecting strong investor interest and confidence in the industry's growth potential [11]

【环球时报报道 记者 陈子帅 杨沙沙 王冬】 编者的话： 2025年，世界科技舞台的聚光灯频频对准中国，其中人形机器人和大模型的表现尤为亮眼。从 DeepSeek以卓越的实用能力引发全球人工智能（AI）开发者社区热议，到多场人形机器人赛事"极限破圈"，再到年底中国AI开源大模型集体"霸榜"国际榜单 ——2025年的中国经济正以AI高科技为新名片迈出国门。多位受访专家表示，中国高科技应用不仅是单项技术偶发突破，更是来自新型举国体制统筹下的 系统集成能力与集群生态优势。 双线领跑，中国 AI 吸引全球聚光灯 "2025年最大的变化，是具身智能和大模型应用都从好看的技术展示，转向能落地的工程体系和产业模式。其中，具身智能行业真正从'实验室表演'走向'工 厂干活、园区干活'的阶段。赛迪研究院未来产业研究中心人工智能研究室主任钟新龙对《环球时报》记者分析称，这一年，以人形机器人产品为具体表现 的具身智能产业关键突破体现在"产业规模化推进和应用闭环能力的显著增强"。 走出实验室，来到聚光灯下，中国人形机器人在2025年的不同"秀场"亮相——2025年央视春晚，宇树科技的人形机器人以"扭秧歌"形式惊艳登台；4月，北 京亦庄 ...

Core Insights - The report titled "2025 Generative Marketing Industry Research: From AIGC to AIGD" highlights the evolution of generative AI in marketing, transitioning from AIGC (AI-Generated Content) focused on content creation to AIGD (AI-Generated Decision) centered on decision support, indicating a shift from AI as an "efficiency tool" to a "strategic partner" in marketing [2][4]. AIGC: Marketing Supply Explosion - By 2025, generative AI has matured in creating marketing content, including copy, images, videos, and digital personas, significantly enhancing content supply [3]. - The emergence of new models and products like DeepSeek and Manus has lowered technical barriers, evolving AI from a "creative assistant" to an "execution agent" [3]. - However, the explosion of content supply raises challenges for businesses in selecting optimal solutions and ensuring the authenticity and reliability of AI-generated content, as AIGC addresses production issues but not effectiveness [3]. AIGD: Systematic Upgrade in Marketing Decisions - AIGD aims to resolve decision-making challenges from the AIGC era by integrating classic marketing theories with AI tools, creating a complete loop from "insight—generation—validation—decision" [4]. - On the consumer side, over 68% of consumers are influenced by AI recommendations in their purchasing decisions, indicating a shift in decision-making power towards AI [4]. - On the enterprise side, AI is utilized for core marketing tasks such as environmental analysis, brand positioning, and demand exploration, enhancing the scientific and efficient nature of marketing decisions [4]. AI Practices: Deepening Industry Applications - Generative AI has been implemented across various industries: - **Food and Beverage**: Companies like Mengniu and Yili use AI for health models and ad testing [5]. - **Beauty and Personal Care**: L'Oréal optimizes formula development with AI, while Proya builds ROI-driven decision systems [5]. - **Automotive**: AI enhances lead management and user profiling [5]. - **Alcohol Industry**: AI integrates into brewing processes and enhances cultural experiences through digital personas and the metaverse [5]. - **Dining and Retail**: Smart ordering and AI-driven site selection are core to digital transformation [5]. - **Apparel, Home Appliances, and Digital Products**: AI assists in design, sales forecasting, virtual fitting, and customer service [5]. Future Outlook: Human-AI Collaboration with Decision Priority - The future of generative marketing emphasizes human-AI collaboration rather than AI replacing humans, necessitating organizations to build "AI-ready" structures that deeply integrate AI into strategy, operations, and innovation processes [6]. - The transition from AIGC to AIGD represents a shift from "content-driven" to "decision-driven" approaches, where effective use of AI for decision-making will determine market success [6].

Group 1 - China's "5G + Industrial Internet" has achieved systematic breakthroughs in the past five years, with over 20,000 construction projects covering all 41 industrial categories, significantly boosting manufacturing development [1] - The investment scale in "5G + Industrial Internet" has exceeded 50 billion RMB, with over 1,200 5G factories built, leading to notable improvements in product quality, operational costs, and production efficiency [1] - China has established the world's largest and highest-quality mobile broadband network, with China Tower constructing 9.53 million base stations, saving over 210 billion RMB through resource sharing [1] Group 2 - China leads in industrial standard construction, having released the world's first industrial 5G international standard and over 100 national and industry standards [2] - The cost of 5G modules has decreased by 90% compared to the initial commercial phase, facilitating the large-scale application of "5G + Industrial Internet" [2] - AI is driving the transformation of "5G + Industrial Internet," with applications in various industries, such as the "Industrial Brain" project by Kaos, which has empowered 160,000 enterprises [2] Group 3 - In the energy sector, AI has generated direct benefits of 1.9 billion RMB for the National Energy Group, with 40% of new developers focusing on the industrial field [3] - In the automotive sector, generative AI has improved design efficiency by over 8%, with nearly 90% of automotive companies adopting AI assistance [3] - The average efficiency of 5G factories has increased by 19.6%, while operational costs have decreased by 14.5%, leading to significant improvements in quality and efficiency [3]

Core Insights - Singapore is shifting its AI strategy from using US-based Meta's models to adopting China's Alibaba's Qwen model, marking a significant pivot in its national AI initiative [3][5][9] - This transition highlights China's growing influence in the AI sector, particularly in Southeast Asia, as it begins to overshadow US technology [3][10][12] Group 1: Strategic Shift in AI Models - Singapore's national AI program (AISG) has officially announced a major change, opting for Alibaba's Qwen model over Meta's [5][7] - The new project, "Sea-Lion v4," will utilize Alibaba's Qwen3-32B open-source architecture, which is tailored for Southeast Asian languages [5][10] - The collaboration will involve Alibaba providing core model technology and ongoing technical support for the "Sea-Lion" model [7][9] Group 2: Performance Comparison - Meta's Llama series model had significant shortcomings, particularly in Southeast Asian languages, with only 0.5% of training data dedicated to these languages [9][10] - In contrast, Alibaba's Qwen3-32B model covers 119 languages and has over 100 billion tokens of training data specifically for Southeast Asian languages, achieving top rankings in language capability [10][12] Group 3: Investment and Growth in AI - Singapore has invested approximately 7 billion SGD (about 1.5% of its GDP) in AI from 2019 to 2023, making it a global leader in AI investment [19][25] - The country aims to become a leading AI economy by 2030, with plans to expand its AI workforce and enhance its infrastructure [26][30] - The AI sector in Singapore is projected to generate 15 billion SGD in value by 2024, contributing 5.8% to the GDP, reflecting its role as a new economic engine [30][33] Group 4: Global AI Landscape - The rise of Alibaba's Qwen has led to a phenomenon termed "Qwen Panic" among Silicon Valley giants, indicating its competitive edge in the global market [12][17] - Major companies, including Airbnb and Nvidia, have expressed reliance on Qwen for its superior performance and cost-effectiveness compared to other models [14][17] - The collaboration between Singapore and China in AI is seen as a strategic move that could position Singapore as a key player in the Southeast Asian AI market [43][44]

Core Viewpoint - The company, Beixin Source (300352.SZ), has not established direct cooperation with Alibaba or Ant Group in the field of artificial intelligence as of November 26 [2]. Group 1: Company Developments - Beixin Source has developed its own secure communication platform called Xinyuan Mixin, which is capable of integrating with AI-driven intelligent dialogue robots [2]. - The Xinyuan AI capability platform allows for dialogue and secure communication between robots and humans [2]. - The company has already integrated with several domestic AI products, including Baidu's Wenxin Yiyan, Alibaba's Tongyi Qianwen, Zhipu's ChatGLM, Kimi, DeepSeek, and iFLYTEK's Xinghuo large model [2].