时代踩下油门和刹车，远比想象中来得更快。 2019年，阿浪在湖北省一个县级市加盟了第一家蜜雪冰城，全年营业额干到150多万，"抛去成本也有的 赚"。彼时，新茶饮市场正以每年超30%的增速狂飙，蜜雪冰城的门店刚突破7000家，尚处于拓荒年 代。特别是在下沉市场，仍有大片空白。 2020年，武汉疫情解封，阿浪看准空置门店和低租金的窗口期，逆势连开几家蜜雪冰城，完成原始积 累。他前后加盟20多家茶饮门店，成了茶饮"圈地运动"的受益者。 2024年初，他又陆续加盟了茶百道、古茗和塔斯汀等，但发现游戏规则已变。此时，全国茶饮门店突破 50万家，市场从"增量竞争"彻底转为"存量厮杀"；蜜雪冰城门店数已达3.3万家，古茗突破9000家，茶 百道也冲刺到近8000家。 "2025年下半年，依然没有一家回本，开始亏钱。"阿浪告诉亿邦动力，"古茗投了60万，好的时候月净 利润4万多。但是受季节影响大，冬天基本不赚钱。熬不住，回本周期太长，最后30多万兑出去了。" "投资回报比不值得，不想干了。"2026年，他倾向于把店面全部兑出去。硬币的另一面，现在仍有许多 人排队加盟蜜雪冰城，或者在找店。但在阿浪眼里，那已是别人想象中的红利。 ...

Core Viewpoint - The spring transportation season has seen a decline in the popularity of "marinated snacks," particularly the three major duck brands, which are facing performance downturns and store closures due to market saturation and changing consumer preferences [2][3][8]. Group 1: Company Performance - The three major duck brands, Zhou Hei Ya, Jue Wei, and Huang Shang Huang, have experienced significant declines in performance, with Jue Wei projected to lose between 160 million to 220 million yuan in 2025, marking its first annual loss [6][19]. - The market capitalization of these brands, once in the hundreds of billions, has now dwindled to tens of billions, and over 5,300 stores were closed in the first half of 2025 alone [6][19]. - Zhou Hei Ya, despite its high-end positioning, saw its net profit drop by over 94% in 2022, while Huang Shang Huang's store count fell below 2,898 by mid-2025, lower than in 2019 [19][21]. Group 2: Market Dynamics - The marinated snack market is experiencing a slowdown, with the market size expected to reach 157.3 billion yuan in 2024, reflecting a drastic decline in growth rate to 3.7% compared to over 15% in previous years [21][23]. - Consumer behavior has shifted, with a preference for lower-priced options; the most common spending range for marinated snacks is 20-30 yuan, while the major brands often exceed this price point [23][25]. - The competitive landscape has evolved, with local brands and snack giants entering the marinated snack market, offering similar or better value propositions, thus diluting the unique selling points of the major brands [25][27]. Group 3: Strategic Changes - To regain market share, the major brands are attempting various marketing strategies, including collaborations with popular culture and introducing new product lines [30][32]. - There is a push towards redefining their product offerings, with Zhou Hei Ya and Jue Wei introducing lower-priced items and expanding into hot marinated dishes to attract younger consumers [34][36]. - Operational efficiency is becoming crucial, with brands needing to close underperforming stores and focus resources on profitable locations to improve overall performance [38][41].

Core Viewpoint - The recent regulatory documents from Chinese authorities signify a shift in the approach to digital asset regulation, moving from strict prohibition to a framework that allows for compliance and controlled utilization of tokenization of real-world assets (RWA) [1][15]. Regulatory Framework - The primary focus of the regulatory documents is on virtual currencies and RWA activities, explicitly prohibiting the tokenization of real-world assets within China and related services, which are deemed illegal financial activities [2][7]. - The documents establish a differentiated regulatory approach, allowing RWA activities under specific conditions while maintaining strict prohibitions on virtual currency operations [3][10]. Definition and Scope - RWA tokenization is defined as the use of cryptographic technology and distributed ledger to convert ownership and rights into tokens for issuance and trading [5]. - The distinction between asset digitization and tokenization is crucial, as the former does not fall under the regulatory scope, while the latter does [4][5]. Compliance Obligations - The regulations impose specific compliance obligations on various market participants, including financial institutions and technology service providers, to ensure that they do not engage in unauthorized RWA activities [9][11]. - For compliant cross-border RWA activities, technology service providers must adhere to legal frameworks, enhance risk management, and report their activities to relevant authorities [11]. Data Security and Cross-Border Concerns - The documents emphasize the importance of data security and the risks associated with cross-border data flows, particularly when domestic asset data is used in foreign financial contexts [12][13]. - Companies must ensure the legality of data transfers and classify data appropriately, especially when it pertains to sensitive information related to asset tokenization [13][14]. Implications for AIoT Companies - AIoT companies must recognize the potential implications of their data usage, especially if their outputs are utilized in RWA activities, as this could subject them to regulatory scrutiny [6][14]. - The regulatory framework necessitates that AIoT firms proactively clarify their data usage and ensure compliance with the new obligations set forth in the regulatory documents [15].

Core Insights - The article discusses the introduction of MoPPS, a new framework for model predictive prompt selection that aims to enhance the efficiency of reinforcement learning fine-tuning for large language models by accurately predicting question difficulty without the need for expensive evaluations from large models [5][26]. Group 1: Training Efficiency - MoPPS significantly reduces computational costs associated with training by minimizing the reliance on large model self-evaluations, achieving up to 78.46% reduction in rollouts compared to traditional methods [15][18]. - The framework accelerates training efficiency by 1.6x to 1.8x compared to conventional uniform sampling methods, ensuring that the most critical questions are selected for training [16][26]. Group 2: Methodology - MoPPS employs a lightweight Bayesian model to predict question difficulty, using a Beta distribution to estimate success rates for each question, which allows for efficient updates based on training feedback [8][9]. - The framework utilizes Thompson Sampling for active question selection, balancing exploration and exploitation to identify questions that are optimally challenging for the model [10][12]. Group 3: Performance Metrics - Experimental results indicate that MoPPS maintains a high correlation between predicted and actual question difficulty, demonstrating its reliability and effectiveness in training scenarios [19][22]. - The framework is compatible with various reinforcement learning algorithms and can adapt to different sampling strategies, enhancing its applicability across different training contexts [20][24]. Group 4: Industry Impact - The research has garnered attention from major industry players such as Alibaba, Tencent, and Ant Group, indicating its potential impact on the field of AI and machine learning [4]. - The MoPPS framework represents a significant advancement in the cost-effective fine-tuning of large models, potentially influencing future developments in reinforcement learning applications [26].

Summary of Key Points Core Viewpoint - On February 9, a total of 35 stocks were involved in institutional trading, with 17 stocks showing net buying and 18 stocks showing net selling by institutions [1] Institutional Net Buying - The top three stocks with the highest net buying by institutions were: - JuLi Rigging: Net buying amount of 154 million [1] - Hunan Silver: Net buying amount of 118 million [1] - FeiWo Technology: Net buying amount of 71.62 million [1] Institutional Net Selling - The top three stocks with the highest net selling by institutions were: - Tongyuan Petroleum: Net outflow amount of 168 million [1] - TuoRi New Energy: Net outflow amount of 155 million [1] - Zhongwen Online: Net outflow amount of 95.74 million [1]

Core Insights - The article discusses a significant surge in orders for a milk tea brand, leading to system crashes during a promotional event, highlighting the challenges of scaling AI-driven marketing efforts [1][2][3] Group 1: Event Overview - On February 6, a promotional event led to over 2 million orders within two hours, causing system overload and temporary shutdowns of delivery services [2][3] - The event was characterized by a lack of clear communication to merchants and delivery personnel, resulting in confusion and operational chaos [3][4] Group 2: Technical Challenges - The system crash was attributed to insufficient server capacity to handle the high volume of concurrent requests, exacerbated by the complexity of AI processing [2][5] - The initial server capacity was only one-third of the estimated peak demand, leading to a failure in scaling up resources in time [2][5] Group 3: Marketing Strategy - The promotional strategy involved significant financial incentives, with a reported budget of 30 billion yuan for user acquisition and engagement [4][8] - The marketing approach aimed to create a "super entry point" for consumers by integrating various Alibaba services, including Taobao and Hema, into the AI platform [3][4] Group 4: Competitive Landscape - The urgency of the promotional event was partly a response to competitive pressures from other companies, such as Tencent, which had announced substantial cash incentives for users [7][8] - The article notes that the marketing tactics employed are reminiscent of traditional methods in the Chinese internet landscape, focusing on immediate user engagement rather than long-term brand loyalty [4][11] Group 5: Future Implications - The success of the promotional event raises questions about the sustainability of user engagement once the incentives are removed, as the long-term adoption of AI shopping remains uncertain [11][12] - The article suggests that while AI can enhance efficiency in specific scenarios, it still struggles to fully understand and predict consumer behavior, which may limit its effectiveness as a shopping assistant [12][13]

Core Insights - Luckin Coffee is advancing towards a secondary listing, while Nova Coffee and Lucky Coffee have moved from the industry fringe to the forefront, indicating a significant shift in the coffee sector by 2026 [1][3] - The recent actions of these two major coffee brands signal key trends in the industry, particularly in terms of competition and market dynamics [3][4] Industry Dynamics - Nova Coffee completed a multi-billion C round financing in January, attracting several prominent investment firms, marking the largest financing in China's catering industry over the past year [1] - The end of the "9.9 yuan unlimited drinks" promotion by Kudi Coffee on February 1 is seen as a sign of the retreat from price wars, although the impact on consumer pricing strategies remains significant [3][11] - The coffee industry in China underwent a paradigm shift driven by the delivery wars, with growth factors transitioning from brand premium to cost-effectiveness and convenience [3][11] Market Expansion - The "10,000 store club" expansion in 2025 is a key indicator of industry evolution, with Nova Coffee and Lucky Coffee joining this elite group, emphasizing the necessity of scale for top-tier brands [4][6] - Luckin Coffee has solidified its market leadership with a total of 29,214 stores globally by Q3 2025, having opened 3,008 new stores in a single quarter [6][10] - Kudi Coffee's aggressive expansion strategy, primarily through a franchise model, has led to over 18,000 stores by December 2025, despite not reaching its target of 50,000 stores [6][9] Competitive Landscape - The competitive landscape is characterized by a divergence in business models, with Kudi Coffee and Nova Coffee adopting lighter operational models for rapid growth [6][9] - Starbucks has entered a phase of strategic contraction in China, highlighted by its partnership with Boyu Capital, which allows Boyu to hold up to 60% equity and control [10][24] - The delivery wars have significantly altered pricing strategies, with Kudi leveraging platform subsidies to offer extremely low prices, impacting overall market dynamics [11][12] Financial Performance - Starbucks reported a 5% revenue growth in its China segment for the fiscal year 2025, but faced a 7% decline in average transaction value, indicating challenges in maintaining profitability amid competitive pressures [12][13] - The cost of delivery has surged for brands like Luckin, with delivery expenses rising to 28.9 billion yuan, a 211% increase year-on-year [13] Strategic Adjustments - Kudi Coffee's shift away from its aggressive pricing strategy marks a transition towards more rational competition, signaling the end of the price war era [14][16] - The trend of coffee brands diversifying into other food categories, such as Kudi's foray into fast food, reflects a broader strategy to capture more consumer spending [18][19] - The coffee sector is increasingly focusing on non-coffee products, with brands like Luckin and Lucky Coffee expanding their tea and juice offerings to adapt to changing consumer preferences [21][24] Future Outlook - The coffee industry is expected to continue evolving, with brands exploring international markets and lower-tier cities as growth avenues, while also addressing the challenges posed by price wars and delivery costs [24][25]

春节临近，卡牌赛道的竞争也日益升温。近日，卡游正式成为中央广播电视总台《2026年春节联欢晚 会》独家卡牌合作伙伴。 其他卡牌品牌也纷纷与央视达成合作：卡卡沃宣布与春晚联名，推出"小马奔奔卡"系列产品；吾流文化 则获得了央视86版《西游记》的卡牌独家授权。 与此同时，卡牌企业也集体在港交所外"排队"。随着Suplay向港交所递交招股书，掀开了2026年卡牌第 二梯队企业冲击上市的热潮。 比起盲盒潮玩赛道，冲击上市的卡牌企业背后站满了豪华资本阵容。闪魂由高瓴创投领投，高榕创投、 凯辉基金跟投。Hitcard背后站着红杉中国种子基金、泡泡玛特和千岛潮玩族等系列机构。而Suplay则因 为最大外部股东是米哈游，引起行业关注。 从营收规模来看，这些成立不久的卡牌企业与卡游还相距甚远。综合企业对外披露的数据以及招股书数 据来看，Hitcard2024年营收为4亿元，Suplay2024年营收为2.81亿元，到2025年前9个月，Suplay营收达 到2.83 亿元，超过2024年全年。相比之下，2024年卡游营收规模为100.57亿元，是卡游新贵们的四五十 倍。 但又因为如今卡牌企业依旧集体依赖授权IP，以及卡牌可拓展 ...

但在FlashLabs看来，这种路径回避了一个更根本的问题：如果底层模型本身并不适合长期运行与实时协作，那么再精巧的产品设计，也只是在放大系统 的结构性上限。 多数团队选择在既有模型能力之上加速产品化，尽快跑通应用与商业闭环；而也有少数人选择了一条更慢、风险更高的路径——回到前沿research和模型 层本身，重新审视Agent的基础假设。 FlashLabs，正是后者。 Open Claw的爆火，让AI Agent第一次被推向了真实的工程环境。 这一次，Agent不再只是Demo、插件或对话式工具，而是开始尝试进入企业内部，承担持续、复杂、可被验证的工作任务。但几乎与此同时，一个现实问 题也被清晰地暴露出来：当Agent走向长期运行的真实工作流，它所面临的挑战，远不止是提示词或工具调用，而是部署成本、交互效率，以及底层模型 是否适合"常驻运行"。 这也迫使行业直面一个更底层、却迟早必须回答的问题—— 如果Agent的目标是成为可靠的数字员工，它是否还应该继续建立在上一代模型与交互假设之上？ 在这一阶段，行业事实上已经形成了一种隐含共识：Agent的问题，应当通过更快的产品迭代来解决。 更复杂的Prompt、 ...

没人写代码，也没人看代码，软件照样交付？ 2026 年 2 月，一家专注于基础设施安全的公司 StrongDM 公开了一套"软件黑灯工厂"式的生产线成果。 在这个生产线里，人类不再直接写代码、也不承担代码审查；开发从交互式协作变成"把 spec 和场景喂给系统"。随后由 Agent 自动生成代码、运行测试 / 评测 harness，并在反馈回路里反复迭代，直到结果收敛、可以交付为止。团队把这套玩法写进章程，最重要的只有一句话——No hand-coded software。 StrongDM AI 还不寻常的开源了它们： 其中一个仓库是：https://github.com/strongdm/attractor。 这是他们"软件工厂"体系中最核心的非交互式编码 Agent。不过，这个仓库本身一行代码都没有：里面只有三份 Markdown 文件，极其细致地描述了软件 的完整规格说明（spec），以及 README 里的一句提示——把这些规格说明交给你选择的编码 Agent 去执行即可。 让他印象最深的，是这套规格说明的体量和细节程度：整套 spec 大约 6000–7000 行，覆盖了行为约束、接口语义以及系统 ...