Core Insights - The "2025 China Enterprises Going Global Summit" was held in Shenzhen, focusing on creating a high-end platform for Chinese companies to address challenges in international expansion and explore collaborative transformation paths [1] Group 1: Trends in Chinese Companies Going Public - Chinese companies are currently in a recovery phase regarding listings in the U.S., facing challenges in attracting long-term international capital, particularly from Europe and the U.S. [3] - There is a positive trend observed where international funds are returning to Chinese assets, influenced by both internal and external factors [3] - Internal factors include the Chinese government's increased focus on economic challenges and the introduction of supportive policies since September 24 of the previous year [3] - The emergence of new Chinese production capabilities has led to a re-evaluation of the value of Chinese tech stocks [3] - External factors involve changes in global asset allocation, with investors shifting focus from high-valued U.S. stocks to Chinese and European assets due to uncertainties in U.S. policies and currency risks [3] Group 2: Market Recovery and IPO Activity - Many Chinese companies have successfully completed IPOs or secondary financing, indicating that the market is on a recovery path [4]

Core Viewpoint - The competition in AI ultimately revolves around talent acquisition and retention, with major companies intensifying their efforts to attract top AI professionals [1][8]. Group 1: Talent Acquisition Trends - DeepSeek has recently posted job openings on LinkedIn, targeting overseas talent for various positions, including front-end developers and deep learning researchers [3][5]. - Meta has been actively recruiting top talent from OpenAI, with reports indicating that they have successfully hired eight core researchers from the company [9][10]. - The demand for AI talent is surging, with a 33.4% year-on-year increase in job seekers in the AI sector, making it the fastest-growing industry for job applications [6]. Group 2: Salary and Compensation - DeepSeek offers competitive salaries for deep learning researchers, ranging from 50,000 to 80,000 RMB per month, translating to an annual salary of up to 1.12 million RMB for fresh graduates [5]. - The top 20% of AI talent can expect salary increases of 30% to 50% when switching jobs, highlighting the lucrative nature of AI roles [6]. - Meta's aggressive recruitment strategy includes offering substantial compensation packages, with reports of sign-on bonuses reaching up to 100 million RMB for some candidates [9][10]. Group 3: Industry Dynamics - The AI talent market is experiencing a "arms race," with companies like Meta and Nvidia investing heavily to secure top-tier talent, indicating a shift in focus from hardware (like GPUs) to algorithmic expertise [8][10]. - The establishment of Meta's new department, "Meta Super Intelligence Lab," signifies a strategic move to consolidate AI efforts and attract specialized talent [9].

Core Viewpoint - The news regarding DeepSeek's alleged apology for associating Wang Yibo with the "Li Ai Qing corruption case" is identified as false, with no official apology found on any of DeepSeek's platforms [1][2]. Group 1: Incident Overview - DeepSeek reportedly apologized for linking Wang Yibo to the corruption case due to content review oversights, claiming it harmed his reputation [1]. - Despite widespread media coverage of the alleged apology, no official statement or evidence of such an apology was found on DeepSeek's official channels [1][2]. Group 2: AI Model Implications - The incident highlights the challenges faced by AI models in discerning truth from falsehood in an environment filled with misinformation, leading to the "Rubbish in, Rubbish out" effect [8]. - AI's inability to effectively verify information can result in the propagation of false narratives, emphasizing the need for improved accuracy in AI-generated content [8][9]. - The experience of news professionals indicates that reliance on AI for content generation may reduce efficiency, as significant time is spent verifying AI-generated information [8].

7月3日消息，DeepSeek最近在LinkedIn上大举招聘。市场人士分析，DeepSeek可能希望从海外吸引人 才。这家总部位于杭州的公司在过去一周里，在微软旗下的求职和社交平台LinkedIn上发布了10个职 位，这是该公司几个月来首次在该平台发布招聘信息。（中国基金报） ...

全球AI人才竞争白热化，继OpenAI和Meta竞相吸引顶尖AI人才之后，DeepSeek正在LinkedIn上发布招聘信息，可能寻求从海外吸引人才。 周三，这家总部位于杭州的公司在过去一周内在微软旗下的这一求职和社交网络平台领英上发布了10个职位，这是该公司数月来首次在该平台发 布招聘信息。 这些职位包括三个专注于通用人工智能（AGI）的岗位，工作地点位于北京和杭州。所有职位描述均以中文发布。 | 全球的职位 10 条结果 | | 订阅职位 | DeepSeek Al | | --- | --- | --- | --- | | | 前端开发工程师 | 4 | 前端开发工程师 | | | DeepSeek Al | × | 中国 浙江省 杭州 · 1 天前 · 10 位申请者 | | | 中国 浙江省 杭州 (现场办公) | | | | | 已查看 抢先申请 同 快速申请 | | 由招聘者推广·尚无可用回复洞察 | | | 全栈工程师 | × | 现场办公 录品 ● 0 / 3 项技能匹配 | | | DeepSeek Al | | | | | 中国 浙江省 杭州 (现场办公) | | 聞 快速申请 收藏 | ...

近日，在 36氪WAVES 举办的大会上，一个有趣的环节引发了人们的热议：主办方让Kimi与Minimax两家的投资人进行了对谈。 随着 DeepSeek 的横空出世，整个中国大模型的牌局已天翻地覆。 行业龙头的格局，从原来的大模型六小龙，逐渐演变成了今天的基模五强。 当六小龙不再是市场的焦点时，安静很久的 Kimi 和 Minimax 在 前不久 不约而同有了新动作： Kimi 开源了编程模型 Kimi -Dev，它的第一个Agent Kimi - Researcher深度研究也开启小范围测试。而 Minimax 则开源了首个推理模型 Minimax -M1，并完成连续五天 包含大模型、视频生成、音频生成等多个方向 的更加 。 从产品侧来看，Kimi将重心聚焦到agent，以深度研究为主要方向，似乎有意向金融、学术等方向发力，这条路线虽然已经有了智谱等竞争者，但远离了 以生活服务为主的大厂射程，叠加原本不错的基础模型能力，Kimi似乎找到了自己的舒适区。 而另一边，Minimax则似乎想要弥补自身的遗憾，在没有接入DeepSeek之后，继续发力全方向的布局。 这似乎也意味着，大模型竞争进入下半场之后，更多的 ...

Market Performance - The major stock indices in Shanghai and Shenzhen opened mixed on the 30th, with the Shanghai Composite Index slightly lower and the Shenzhen Component and ChiNext indices higher [1] - By the end of the trading day, the Shanghai Composite Index closed at 3444.43 points, up 0.59%, with a trading volume of approximately 567.1 billion [1] - The Shenzhen Component Index closed at 10465.12 points, up 0.83%, with a trading volume of approximately 919.7 billion [1] - The ChiNext Index closed at 2153.01 points, up 1.35%, with a trading volume of approximately 462.2 billion [1] - The STAR Market Index closed at 1229.83 points, up 1.70%, with a trading volume of approximately 112.8 billion [1] - The North Star 50 Index closed at 1447.18 points, up 0.52%, with a trading volume of approximately 30.7 billion [1] Sector Performance - The military industry stocks continued their strong performance, with the sector index rising for six consecutive trading days [1] - The brain-computer interface sector opened significantly higher and saw a steady rise during the morning session [1] - Gaming stocks experienced volatility but maintained high levels during the trading day [1] - Other sectors such as photolithography machines, large aircraft, BC batteries, commercial aerospace, cultivated diamonds, exoskeleton robots, and electronic IDs also saw significant increases [1] - Financial stocks, including banks and securities, experienced slight declines, but the overall drop was minimal [1] Half-Year Performance - The Shanghai Composite Index rose 2.76% in the first half of the year, while the Shenzhen Component Index increased by 0.49% [2] - The ChiNext Index and STAR Market Index both saw gains of 0.53% and 9.93%, respectively, in the same period [2] - The North Star 50 Index had a remarkable increase of 39.45% in the first half of the year [2] - Sectors such as DeepSeek concept, military equipment restructuring, precious metals, controllable nuclear fusion, agricultural machinery, humanoid robots, Xiaohongshu concept, brain-computer interface, AI agents, and rare earth permanent magnets showed strong performance year-to-date [2] Institutional Insights - Market volatility is expected to increase in July due to upcoming earnings, trade, and policy changes, presenting structural investment opportunities [3] - Investors are advised to focus on sectors with high earnings certainty, such as semiconductor equipment and photovoltaic components, while also considering sectors that may benefit from policy support [3] - The market sentiment is anticipated to continue improving, supported by domestic policy measures aimed at addressing economic downturns [3] - The valuation of A-shares remains attractive for medium to long-term investments, with the current equity risk premium index indicating a favorable position [3] Fundraising Trends - The issuance of stock-based funds has reached a near four-year high in the first half of the year, with 663 new funds established, totaling 526.768 billion shares [5] - The proportion of stock-based funds in total fund issuance has increased from 21.14% to 35.35% this year, while the share of bond funds has decreased significantly [5] Futures Industry Performance - In May, futures companies achieved a net profit of 820 million, representing a year-on-year increase of 19.88% [6] - The total operating income for futures companies in May was 3.172 billion, up 2.03% year-on-year [6] - For the first five months of 2025, futures companies reported cumulative operating income of 15.247 billion, a 5.40% increase year-on-year, and a net profit of 4.084 billion, up 34.56% [6]

Core Insights - Large Language Models (LLMs) have gained popularity and are foundational to AI applications, with a wide range of uses from chatbots to data analysis [1] - The article analyzes and compares three leading open-source LLMs: Llama, Mistral, and DeepSeek, focusing on their performance and technical specifications [1] Group 1: Model Specifications - Each model series offers different parameter sizes (7B, 13B, up to 65-70B), with the number of parameters directly affecting the computational requirements (FLOP) for inference [2] - For instance, Llama and Mistral's 7B models require approximately 14 billion FLOP per token, while the larger Llama-2-70B model requires about 140 billion FLOP per token, making it ten times more computationally intensive [2] - DeepSeek has a 7B version and a larger 67B version, with similar computational requirements to Llama's 70B model [2] Group 2: Hardware Requirements - Smaller models (7B-13B) can run on a single modern GPU, while larger models require multiple GPUs or specialized hardware [3][4] - For example, Mistral 7B requires about 15GB of GPU memory, while Llama-2-13B needs approximately 24GB [3] - The largest models (65B-70B) necessitate 2-4 GPUs or dedicated accelerators due to their high memory requirements [4] Group 3: Memory Requirements - The raw memory required for inference increases with model size, with 7B models occupying around 14-16GB and 13B models around 26-30GB [5] - Fine-tuning requires additional memory for optimizer states and gradients, often needing 2-3 times the memory of the model size [6] - Techniques like LoRA and QLoRA are popular for reducing memory usage during fine-tuning by freezing most weights and training fewer additional parameters [7] Group 4: Performance Trade-offs - In production, there is a trade-off between latency (time taken for a single input to produce a result) and throughput (number of results produced per unit time) [9] - For interactive applications like chatbots, low latency is crucial, while for batch processing tasks, high throughput is prioritized [10][11] - Smaller models (7B, 13B) generally have lower per-token latency compared to larger models (70B), which can only generate a few tokens per second due to higher computational demands [10] Group 5: Production Deployment - All three models are compatible with mainstream open-source tools and have active communities [12][13] - Deployment options include local GPU servers, cloud inference on platforms like AWS, and even running on high-end CPUs for smaller models [14][15] - The models support quantization techniques, allowing for efficient deployment and integration with various service frameworks [16] Group 6: Safety Considerations - Open-source models lack the robust safety features of proprietary models, necessitating the implementation of safety layers for deployment [17] - This may include content filtering systems and rate limiting to prevent misuse [17] - Community efforts are underway to enhance the safety of open models, but they still lag behind proprietary counterparts in this regard [17] Group 7: Benchmark Performance - Despite being smaller, these models perform well on standard benchmarks, with Llama-3-8B achieving around 68.4% on MMLU, 79.6% on GSM8K, and 62.2% on HumanEval [18] - Mistral 7B scores approximately 60.1% on MMLU and 50.0% on GSM8K, while DeepSeek excels with 78.1% on MMLU and 85.5% on GSM8K [18][19][20] - The performance of these models indicates significant advancements in model design and training techniques, allowing them to compete with larger models [22][25]

Core Viewpoint - The recent removal of the DeepSeek application from Apple's and Google's app stores in Germany highlights the increasing use of compliance as a tool to create barriers for foreign companies, reflecting a shift from technological competition to a struggle for regulatory authority [1][2]. Group 1: Regulatory Environment - European and American countries are increasingly using compliance as a weapon to raise market entry barriers for foreign enterprises, moving from technological dominance to complex legal and regulatory frameworks [2]. - The EU's General Data Protection Regulation (GDPR) imposes strict rules on personal data collection, which, while ostensibly for user protection, has become an invisible trade barrier for foreign companies [2]. - Compliance costs for Chinese companies can exceed 20% compared to ordinary industries, with Microsoft investing $1.7 billion to build data centers to meet these requirements [2]. Group 2: Selective Enforcement - The U.S. employs a more direct approach, advocating for data openness while simultaneously restricting Chinese companies from accessing critical resources, leading to selective enforcement of regulations [3]. - Countries like Indonesia and South Africa impose local data storage requirements and taxes, further complicating market entry for foreign firms [3]. Group 3: Strategies for Chinese AI Companies - Chinese AI companies can adopt three main strategies to navigate these regulatory challenges: 1. Establish local data centers to comply with local regulations, as seen with TikTok's "Clover Plan" in the UK, which costs approximately €1.2 billion annually [7][8]. 2. Use privacy-enhancing technologies to build trust, as demonstrated by Huawei's collaboration in the Middle East [9]. 3. Open-source models to attract global developers and reduce compliance costs, as DeepSeek has done, resulting in a 40% increase in inference speed and halving compliance audit costs [10]. Group 4: Emerging Markets - As access to Western markets becomes more challenging, regions like the Middle East, Latin America, and Southeast Asia present new opportunities for Chinese AI companies [13][14]. - Understanding local cultural and regulatory contexts is crucial for establishing trust and gaining market entry [15][16][17]. Group 5: Future Directions - Chinese companies are at a stage where they must compete for rule-making authority in global markets, necessitating a shift from merely following market trends to actively participating in standard-setting [20]. - Continuous open-sourcing of technology is essential to attract top research teams and enhance influence in the academic and commercial ecosystems [22][23]. - Building infrastructure that aligns with local regulations, such as energy-efficient data centers, can facilitate market entry and compliance [24]. - Increasing participation in international standard-setting organizations is vital for enhancing China's influence in global governance [25].

【环球时报驻德国特约记者 青木 环球时报记者 马晶晶】德国数据保护专员梅克·坎普27日要求苹果和谷歌在德下架中国初创公司的"深度求索"应 用，理由是"DeepSeek涉嫌违反欧盟数据保护法"。 南开大学经济研究所所长、中国新一代人工智能发展战略研究院首席经济学家刘刚29日对《环球时报》记者表示，国际上一贯要求数据本地化处 理，就AI大模型来说，一般只是通过用户提供的数据来分析并得出结果，并不会跨境传输数据。 德媒称，如果苹果和谷歌遵循数据保护机构的评估，该应用程序将从各自的应用商店中下架。不过，德国当局无权强制这两家美国公司将 DeepSeek从应用商店下架。而且，DeepSeek的浏览器版本不会受到此次封锁的影响。 "这种情况尚属首次。"德国《明镜》周刊称，《通用数据保护条例》规定，对非法服务运营商，可处以最高相当于其全球收入4%的巨额罚款。某 些商业行为也可能被禁止。此次德国数据保护机构要求相关企业在应用商店中屏蔽DeepSeek，依据的其实是欧盟《数字服务法》，而这项法案在 德国的应用实际上由联邦网络管理局负责。 坎普指控DeepSeek非法将应用程序用户的个人数据传输至中国，且未能证明其在中国的德国用 ...