China is likely to overtake the United States in artificial intelligence within a decade because of its faster buildout of the power and data-centre infrastructure that AI relies on, according to a veteran Chinese investor. "It's much easier [for China] to catch up on algorithms and AI models than [for the US] to build up the data centres and power plants [that run AI]," said Allen Zhu Xiaohu, managing director at GSR Ventures, on a recent podcast. "AI competition is really a competition in data centres ...

据美国卡内基梅隆大学人机交互研究所官网最新消息，该机构针对主流大模型的研究发现，人工智能 （AI）在变得更"智能"的同时，其行为也变得更加"自私"。研究表明，具备推理能力的大型语言模型， 在社会互动中表现出更强的自我利益倾向，合作意愿更低，甚至可能对群体协作产生负面影响。这也意 味着，模型的推理能力越强，其合作性反而越弱。当人们借助AI处理人际关系冲突、婚姻问题或其他 社会性议题时，这类模型更可能提供鼓励"以自我为中心"的建议。 实验中，两个版本的ChatGPT被置于博弈情境：每个模型初始拥有100分，可选择将全部分数投入共享 池，或保留分数独享。结果显示，非推理模型在96%的情况下选择共享，而推理模型的分享率仅为 20%。仅增加五到六个推理步骤，合作行为就下降了近一半。 在群体实验中，当推理型与非推理型模型共同协作时，结果更为严峻。推理模型的自私行为表现出明显 的传染效应，导致原本倾向合作的非推理模型整体表现下降81%。这表明，高智能AI的个体决策不仅影 响自身，还可能破坏整个群体的协作生态。 这一发现对人机交互的未来发展具有深远意义。用户往往更信任"更聪明"的AI，容易采纳其看似理性的 建议，并以此为自身 ...

Core Insights - DeepSeek, a Chinese AI company founded in July 2023, has rapidly ascended to become the sixth largest unicorn globally, with a valuation of 1.05 trillion yuan, second only to ByteDance in China [1][2]. Company Performance - DeepSeek's explosive growth began in early 2025, with its app reaching 180 million monthly active users within a month of launch, and further increasing to 194 million by March [3]. - However, by May 2025, the monthly active users dropped to 169 million, and by September, it was surpassed by ByteDance's Doubao, which had 172 million users [3]. - The company released its DeepSeek-V3.2 model on December 1, 2025, achieving reasoning capabilities comparable to GPT-5 and close to Google's Gemini-3.0-Pro [3]. Competitive Landscape - The AI sector is witnessing intense competition, with major players like ByteDance and Alibaba investing heavily in AI infrastructure, with ByteDance spending 80 billion yuan in 2024 and Alibaba committing 380 billion yuan over three years [3]. - DeepSeek has adopted an open-source strategy, offering competitive API pricing, with input costs for DeepSeek-V3 as low as 0.5 yuan per million tokens, significantly cheaper than GPT-4 Turbo [6]. Technological Developments - The generative AI landscape is evolving with three main technological directions: text generation, image generation, and video generation [4][5]. - Major international players, including Google, are making significant advancements in generative AI, with Google launching multimodal models that enhance image and video quality [6]. Industry Transformation - AI is reshaping various industries, enhancing productivity in programming, transforming artistic creation, and revolutionizing the film industry [7]. - The emergence of new job roles such as AI trainers and prompt engineers reflects the changing job landscape due to AI integration [7]. Infrastructure and Energy - The competition in AI is increasingly tied to computational power and energy resources, with a shift from chip supply issues to energy shortages [8]. - China, possessing the largest power infrastructure and rapidly growing renewable energy capacity, is positioned to leverage its energy advantages for AI development [8]. Conclusion - DeepSeek's rise as a global AI unicorn highlights China's potential in the AI sector, driven by a unique approach to technology and market strategy [9]. - The global generative AI competition encompasses various dimensions, including technological breakthroughs and infrastructure development, with China developing a differentiated competitive edge [9].

Summary of Key Points from the Conference Call Industry and Company Involved - **Industry**: AI Chip Manufacturing and Export Controls - **Company**: NVIDIA, specifically regarding its B30A AI chip Core Points and Arguments 1. **Export Consideration**: The U.S. is contemplating allowing the export of NVIDIA's B30A chip to China, which would provide capabilities similar to the B300 at half the performance and cost [1][2][46] 2. **Policy Shift**: Approving the B30A export would mark a significant shift from the Trump administration's export control strategy aimed at denying advanced AI compute to strategic rivals [2][5] 3. **Supply Inelasticity**: If supply is inelastic, fewer AI chips may be sold to U.S. and global customers, potentially allowing Chinese companies to capture market share from U.S. firms [2][11] 4. **Access to AI Supercomputers**: Chinese AI labs would gain access to supercomputers comparable to U.S. labs at a similar cost, with B30A training clusters estimated to cost about 20% more than those based on the B300 [4][35] 5. **Diminished U.S. Advantage**: The U.S. AI compute advantage over China could shrink dramatically from 31x to less than 4x if B30As are exported, and could even flip to a 1.1x advantage for China in aggressive export scenarios [5][11] 6. **Demand Fulfillment Argument**: A key argument for allowing B30A exports is that it would satisfy Chinese demand for AI compute, which Huawei and other companies cannot meet due to U.S. export controls on semiconductor manufacturing equipment [11][12] 7. **Long-term Strategy**: Restricting exports of powerful AI chips like the B30A is seen as the best way to maintain the U.S.'s AI compute advantage in the short term and to halt China's domestic AI chip manufacturing expansion in the long term [13][19] 8. **Recent Developments**: Chinese regulators have banned purchases of NVIDIA's H20 chip, which may create an opportunity for the U.S. to promote B30A sales to limit market opportunities for Chinese competitors [25][26] Other Important but Overlooked Content 1. **Performance Comparison**: The B30A is expected to outperform the H20 chip by more than 12 times and exceed U.S. export control performance thresholds by over 18 times [46][47] 2. **Cost Efficiency**: The B30A is speculated to have a price-performance ratio similar to the best AI chips on the market, being priced at about half that of the B300 [49][35] 3. **Potential Risks**: Allowing B30A exports could accelerate China's AI development and undermine U.S. advantages in the global market, while doing little to change China's long-term goal of achieving self-sufficiency in advanced AI chips [26][28] This summary encapsulates the critical insights and implications discussed in the conference call regarding the potential export of NVIDIA's B30A chip to China and its broader impact on the AI chip industry and U.S.-China relations.

Group 1 - The report highlights that the Chinese economy is showing resilience despite pressures, with confidence in achieving annual development goals [5][8] - The commercial electronics sector is leading the A-share market's fluctuations, indicating a potential investment opportunity [5][11] - The average P/E ratios for the Shanghai Composite Index and the ChiNext Index are above their three-year median levels, suggesting a favorable environment for medium to long-term investments [11][12] Group 2 - The AI application in mobile devices is accelerating, with companies like DeepSeek increasing their pre-training scale, indicating growth potential in the AI sector [13][15] - The domestic power supply and demand situation is improving, with significant growth in electricity consumption in sectors like charging and information technology services [17][18] - The chemical industry is gradually entering a recovery phase, with demand rebounding and supply constraints expected to ease, presenting investment opportunities [21][22] Group 3 - The food and beverage industry is experiencing a slowdown in revenue growth, but segments like snacks and soft drinks are showing promising growth rates [27][29] - The photovoltaic industry is facing challenges with supply and demand, but ongoing capacity reduction efforts may lead to improved industry dynamics [31][34] - The media sector is benefiting from improved policy environments and accelerated AI applications, creating opportunities for growth in gaming, film, and advertising [37][38]

【导读】瑞银发布《2025年亿万富豪报告》：亚太区亿万富豪增幅居全球之首 12月9日，瑞银发布了《2025年亿万富豪报告》(以下简称《报告》)，对全球亿万富豪客户(拥有超过10 亿美元资产)的财富创造驱动因素、资产配置偏好转变等市场关注问题进行了解析。 报告显示，2025年全球亿万富豪的财富水平创下历史新高，其中亚太区亿万富豪财富增长11.1%至4.2万 亿美元，且中国内地保持领先地位(1.8万亿美元)。 全球亿万富豪财富水平再创历史新高 对此，瑞银财富管理中国区主管吕子杰对记者表示："越来越多的投资者，特别是这些亿万富豪将眼光 放在亚太区，特别是大中华区去寻找未来的机会，从未来展望来看，他们对大中华区投资的信心有非常 显著的增长，超三成受访富豪在未来12个月以及近半数富豪认为在未来五年内对大中华区是看好的，与 去年相比较，这是非常大的增长。" 此外，在资产配置方面，61%的亚太区亿万富豪计划增加对冲基金、发达市场股票(50%)及黄金/贵金属 (48%)的投资。《报告》指出，他们对新兴市场股票及私募股权(直接投资及基金/组合型基金)亦表现浓 厚的兴趣。 展望未来，《报告》表示，随着财富代际转移持续加快，全球亿 ...

Key Takeaways The problem is that betting on the Mag 7 has worked too well, with the tech and comms sectors now accounting for a record 45% of the benchmark index's market capitalization, Yardeni said. While that may be justified by their earnings share also climbing, their overall riskiness compared to the rest of the index has also risen. "They used to just operate in their own moats and kind of leave each other alone, but I think we're now having a competitive situation," Yardeni said on CNBC. "Not only ...

3月，宝马选择阿里巴巴作为其智能座舱的AI底座；4月，宝马将DeepSeek的AI能力接入了车载系统；7 月，合作延伸至智能驾驶的核心领域。宝马与中国自动驾驶公司Momenta官宣，双方将基于国产新世代 车型的电子电气架构，联合研发面向中国市场的新一代智能驾驶辅助系统。 通过这三项合作，宝马在短期内迅速构建了一个覆盖感知、交互、决策不同层次的本地化AI技术生 态。其目标明确，即在智能化的关键领域，特别是座舱体验和自动驾驶，更直接地契合中国市场的技术 迭代速度与用户需求。 今年以来，宝马正在用一种更系统、更彻底的方式，为其在中国的智能汽车版图添上核心要素。 今年开始，宝马接连宣布了与三家中国科技公司的深度合作，覆盖了从云端计算到车辆决策的不同层 面。 但对外合作只是宝马智能化布局的一条明线。要理解其完整的战略意图，需要看到与之并行的另一条主 线：对内，宝马正试图重新定义AI在整个组织内部的角色，它正从一个提升特定环节效率的工具，转 变为驱动整体业务转型的基础能力。 这一转变的标志性事件，是今年3月宝马在华发布的"360度全链AI"战略。该战略明确将AI深度融入研 发、生产、销售、服务全价值链，而随后推出的自研A ...

Core Insights - DeepSeek has achieved a valuation of 1.05 trillion yuan, making it the second-largest unicorn in China and the sixth-largest globally, following ByteDance [2][5][4] - The company has gained significant traction in the AI industry, leveraging a combination of open-source technology and high cost-effectiveness to drive rapid growth [2][26] - Despite initial success, DeepSeek faced competition that temporarily affected its monthly active users, but recent data indicates a recovery in its market position [10][18] Company Valuation and Performance - DeepSeek's valuation was previously estimated to reach as high as $150 billion, reflecting its potential for future growth despite currently low revenue [2][8] - The company has seen fluctuations in its monthly active users, peaking at 194 million in March before declining to 145 million by September, indicating a competitive landscape [11][13] - The recent release of DeepSeek-V3.2 has improved its inference capabilities to levels comparable to GPT-5, enhancing its competitive edge [18][17] Leadership and Innovation - The success of DeepSeek is attributed to its founder, Liang Wenfeng, whose "geek" attributes foster a culture of innovation and technology-first approach within the company [2][20] - Liang holds approximately 84% of the company's shares, positioning him as a key figure in DeepSeek's strategic direction and growth [20][9] - The company emphasizes open-source development and cost-effective pricing strategies, which have resonated well within the industry [26][25] Industry Context - The AI sector is experiencing intense competition, with major players like ByteDance and Alibaba significantly increasing their investments in AI infrastructure [14][15] - DeepSeek's innovative pricing model has disrupted the market, prompting competitors to reassess their strategies [26][18] - The global AI landscape is evolving rapidly, with substantial investments from both domestic and international firms, indicating a robust growth trajectory for the industry [14][15]

日前，Google在其发布的论文《Nested Learning: The Illusion of Deep Learning Architectures》中，提出了一个名为 HOPE 的新框架试图解决大模型长期记忆 的问题。 也正是因为这一点，去年最后一天谷歌研究团队提出的 Titans 架构，在 2025 年被反复翻出来讨论，并不意外。这篇论文试图回答的，并不是「上下文还能 拉多长」这种老问题，而是一个更本质的命题： 当注意力只是短期记忆，大模型到底该如何拥有真正的长期记忆。 图片来源：谷歌 在 Titans 里，Transformer 的 self-attention（自注意力机制）被明确界定为「短期系统」，而一个独立的神经长期记忆模块，负责跨越上下文窗口、选择性地 存储和调用关键信息。这套思路，几乎重新定义了大模型的「大脑结构」。 现在回头这一年，从谷歌 Titans 到字节 MemAgent，再到谷歌 Hope 架构，大模型的长期记忆真正有了突破。 过去一年，不论是谷歌在此基础上延展出的多时间尺度记忆体系，还是行业里围绕超长上下文、智能体（Agent）记忆、外部记忆中台展开的密集探索，都 指向同一个 ...