AI正将一门依赖试错的实验科学，转变为一门可预测、可编程的数据科学。 作者 | 申思琦 编辑 | 硬 AI "未来，所有的生物学都将以'计算机模拟'（in silico）为起点，并以'计算机模拟'为很大程度的终点。"（Almost everything will largely start in silico, largely end in silico） 当英伟达（NVIDIA）创始人兼CEO黄仁勋在J.P. Morgan医疗健康大会和GTC大会的聚光灯下反复提及这一论断时，生物学正在从一门依赖试错的实验科学，转变 为一门可预测、可编程的数据科学。 而英矽智能（Insilico），正是这个被黄仁勋多次 点名 的"第一性原理"践行者。 2025年12月30日，英矽智能（3696.HK）在香港联合交易所主板正式挂牌， 开盘价报35港元，较发行价升45%，市值达到195亿港元 。 | 10 901 日 周 月 8 年 1分 5分 15分 30分 60分 8 80期 10图 | 角 九期 酮自选 ▼ 0007200 | | 英矽智能-B 03696 | | 60 95 | 35.000 9567 | 38.347 ...

Meta以数十亿美元收购AI初创公司Manus，整个收购谈判在极短时间内完成，前后不过十余天。Manus今年早些时候的年度经常性收入已达1.25亿美元，通过订阅服务向企业销售 AI代理。这项收购将为Meta在AI领域的巨额投资提供更直接的收入回报。 | 硬·AI | 作者 | 赵 | 颖 | | --- | --- | --- | --- | | 编辑 | 硬 | AI | | Meta以数十亿美元收购AI初创公司Manus，这是这家社交媒体巨头成立以来第三大收购交易，仅次于WhatsApp和Scale AI。这笔交易标志着Meta在AI领域的激 进投资策略进入新阶段，也为中国创业者在全球AI竞赛中树立了新标杆。 周二，据《晚点LatePost》，Manus母公司蝴蝶效应在被收购前正以20亿美元估值进行新一轮融资。整个收购谈判在极短时间内完成，前后不过十余天。收购完 成后，蝴蝶效应将保持独立运作，创始人肖弘将出任Meta副总裁。 随后，Meta首席人工智能官 Alexandr Wang发推文称，欢迎Manus AI的加入。 CEO Red 肖弘也发声称，"这不仅仅是一次收购。它验证了我们一直以来努力构建的未来 ...

低空经济新样本。 硬·AI 作者 | 硬 AI 编辑 | 硬 AI 一场关于物流效率的 革命， 正在城市拥堵的街道上空 展开 。 近日， 广州海珠区联合海珠城发、迅蚁及淘宝闪购正式启动 "城市微空港"项目，首批试点位于保利广场与华新中心。 该项目采用 "无人机干线+骑手末端"的接驳模式，实测将5.7公里路程的配送时间从20分钟缩短至9分钟。 不同于传统的单一人力配送， "城市微空港"采用了"无人机干线+骑手末端"的接驳模式。用户下单后，商品由无人机完成主要路段的空中运输，抵达接驳站后由骑 手完成最后配送。据项目方披露，该服务目前不产生额外配送费用。 在基础设施方面，项目强调 "坪效"与空间复用。新建的城市微空港不仅是无人机起降平台和空地物流接驳站，还集成了24小时机器人咖啡馆及微型城市展厅功 能，通过"一地多用"提升设施的综合商业价值。 02 配送时效缩短超50% 基于保利广场至华新中心航线的实测数据显示，两地路面距离 5.7公里，途经15个红绿灯，传统骑行需耗时约20分钟。 相比之下，无人机空运仅需 9分钟，且单次飞行可挂载多个订单，显著提升了特定场景下的物流效率。 为解决运力瓶颈，新型微空港采用了 "双通 ...

Core Viewpoint - The year 2025 marks a transformative period for Chinese business, driven by technological advancements and strategic market maneuvers, including DeepSeek's cost paradigm shift in AI, the establishment of a "stabilization fund" by state-owned enterprises, and fierce competition in various sectors like food delivery and consumer products [2][3][4]. Group 1: AI and Technology - DeepSeek's R1 model demonstrated a significant cost advantage, achieving comparable performance to OpenAI's models at a fraction of the cost, leading to a reevaluation of AI asset values globally [10]. - The Chinese stock market reacted positively to the implications of DeepSeek's success, with the Nasdaq China Golden Dragon Index rising over 4% shortly after [10]. - The launch of L3 autonomous driving vehicles in China signifies a major milestone in the commercialization of advanced driving technologies, with expectations of a market size exceeding 1 trillion yuan by 2030 [49][51]. Group 2: Market Stability Measures - In response to external economic pressures, the "national team" intervened in the stock market by establishing a "stabilization fund," which included significant investments from state-owned enterprises to restore market confidence [12][14][18]. - The People's Bank of China supported these efforts by promising sufficient liquidity to stabilize the market, reinforcing the government's commitment to maintaining financial security [14][18]. Group 3: Consumer and Service Sector Developments - JD.com entered the food delivery market, intensifying competition with Alibaba and Meituan, leading to aggressive pricing strategies and significant order volume growth [26][30]. - Pop Mart's Labubu character achieved global popularity, resulting in a revenue surge of 170%-175% in Q1 2025, with notable growth in international markets [20][22]. - The competition in the food delivery sector is characterized by substantial subsidies and promotional offers, indicating a shift towards efficiency and market share acquisition among major players [28][30]. Group 4: Capital Market Movements - The collective IPO efforts of China's "four little dragons" in the GPU sector highlight a significant moment for domestic chip manufacturers, with substantial market valuations and growth expectations [52][54]. - The stock prices of Pop Mart surged over 200% in the first half of 2025, reflecting strong market interest and future growth potential, despite a subsequent correction [22][25]. Group 5: Breakthroughs in Energy and Aerospace - China achieved significant milestones in nuclear fusion research, with advancements in plasma physics and the development of the next-generation fusion energy experimental device [58][59]. - The successful test flights of reusable rockets by both private and state-owned enterprises mark a new era in China's commercial space industry, aiming for cost reductions and increased launch frequency [60][63].

Core Insights - Jim Fan criticizes the current state of the robotics industry, highlighting significant advancements in hardware but ongoing chaos in software iteration, standardization, and technology direction [3] - He emphasizes that the mainstream Visual-Language-Action (VLA) model is fundamentally misaligned with the actual needs of robotics, advocating for a shift towards video world models as a more suitable alternative [3][11] Group 1: Hardware Reliability - Hardware reliability is identified as the biggest obstacle to software iteration, with advanced robots like Optimus and e-Atlas facing limitations due to issues such as overheating and motor failures [7] - The inability of robots to self-repair exacerbates the problem, leading to high human resource costs and low iteration efficiency in development [7] Group 2: Lack of Industry Standards - The benchmarking situation in the robotics field is described as a "catastrophe," lacking unified standards for hardware platforms, task definitions, and evaluation criteria [9] - Companies often create their own benchmarks for public announcements, leading to misleading claims of achieving "state-of-the-art" performance [9] Group 3: Fundamental Questions on Technology Direction - The VLA model is fundamentally questioned, as its parameters are primarily optimized for language and knowledge rather than physical applications, which is critical for robotics [11] - Jim Fan argues that the pre-training objectives of VLM do not align with the requirements of robotics, suggesting that increasing VLM parameters will not enhance VLA performance [11]

Core Viewpoint - Nvidia is making a strategic move by spending approximately $20 billion to acquire technology from the startup Groq, aiming to eliminate potential threats in the efficient and low-cost AI inference chip market while integrating a top-tier team to address its technological shortcomings [2][3]. Group 1: Strategic Intent - The acquisition is not just a defensive measure against competitors but also a key strategy to build a wider moat and solidify Nvidia's absolute market leadership [2]. - Nvidia's CEO Jensen Huang emphasized the intention to integrate Groq's low-latency processors into Nvidia's AI factory architecture, expanding platform capabilities for a broader range of AI inference and real-time workloads [4][5]. Group 2: Market Dynamics - The core driver of this transaction is Nvidia's competition for the AI inference market, where its existing chips are often too large and costly for practical applications like chatbots [5]. - Groq claims its chips outperform Nvidia's in specific AI application tasks, indicating a potential threat to Nvidia's dominance as Groq's next-generation products are on the horizon [5]. Group 3: Transaction Structure - The deal is structured as a non-exclusive technology license, allowing Nvidia to hire Groq's founders and executives while Groq retains its cloud business [7][8]. - This structure is a common tactic among tech giants to circumvent regulatory scrutiny, similar to past strategies employed by Microsoft, Amazon, and Google [8]. Group 4: Competitive Landscape - Despite significant venture capital backing, challengers like Groq struggle to disrupt Nvidia's stronghold in the high-end AI chip market, as evidenced by Groq's recent revenue forecast cut [10]. - The competitive landscape is intensifying, with Google’s TPU emerging as a strong competitor to Nvidia's GPUs, and other companies like Meta and OpenAI developing their own specialized inference chips [10]. Group 5: Financial Strategy - Nvidia is leveraging its substantial cash reserves, which reached $60 billion by the end of October, to consolidate its business and pursue larger-scale technology acquisitions [12]. - The $20 billion transaction with Groq exceeds Nvidia's previous largest acquisition, indicating a willingness to invest heavily to eliminate potential threats and integrate cutting-edge technology [12].

Core Viewpoint - OpenAI is actively exploring the commercialization of its flagship product ChatGPT through advertising, aiming to monetize its large user base of nearly 900 million, which could challenge the dominance of Google and Meta in the trillion-dollar digital advertising market [1][2]. Group 1: New Advertising Model - OpenAI aims to create a new type of digital advertising that integrates seamlessly into the ChatGPT experience, utilizing detailed user interaction data to display highly relevant ads [5]. - The company is focusing on "non-intrusive" advertising methods to maintain user experience and trust, with ads potentially appearing only at specific stages of user interactions [5]. Group 2: Monetization Pressure and Market Opportunity - OpenAI faces significant monetization pressure, as only about 5% of its nearly 900 million weekly active users are paying customers, with plans to generate substantial revenue from the large free user base through advertising [7]. - The company projects that average revenue per user from free users will increase from $2 in the coming year to $15 by 2030, with total revenue from non-paying users expected to reach approximately $110 billion by 2030, achieving gross margins comparable to Meta's Facebook [7][8]. Group 3: Balancing Trust and Commercialization - Advertising has been a sensitive topic for OpenAI, with concerns that it may undermine user trust in the responses provided by ChatGPT, although the CEO has softened his stance on the feasibility of ads over time [10]. - The company is weighing the need for a sustainable business model against the goal of achieving general artificial intelligence (AGI) [10]. Group 4: E-commerce Integration and Early Stage of Advertising - OpenAI is laying the groundwork for commercialization by integrating shopping features into ChatGPT, collaborating with companies like Stripe, Shopify, Zillow, and DoorDash to enhance the user experience [12]. - Despite ongoing discussions about advertising, the advertising business is still in its early stages, with internal priorities shifting towards improving core functionalities of ChatGPT [13].

Core Viewpoint - The article discusses the transformative impact of AI on power supply systems, highlighting predictions from Infineon regarding future power demands and technological advancements in the next decade [3][4]. Group 1: Key Predictions - Prediction 1: Vertical power supply will dominate future processor architectures, with current load currents expected to reach 10,000 amperes, ten times the current level [6]. - Prediction 2: High-voltage direct current (DC) power supply architectures will replace the existing 48V ecosystem as power requirements exceed 1 megawatt [12]. - Prediction 3: AI server rack power consumption will enter the megawatt era, with total power expected to exceed 1 megawatt as more GPUs are integrated into single racks [16]. - Prediction 4: Power supply rack power levels will surpass 100 kilowatts, laying the groundwork for 1 megawatt IT racks [19]. - Prediction 5: New data center power demands will approach gigawatt levels, with dedicated "AI factories" expected to consume several gigawatts of power [22]. - Prediction 6: Distribution systems will shift towards DC microgrids to accommodate gigawatt-level power demands [24]. - Prediction 7: Renewable energy will become a critical constraint for AI development, with nearly 50% of power used by global data centers expected to come from renewable sources [27][28]. Group 2: Technological Innovations - Infineon is developing a complete product portfolio from discrete power levels to four-phase vertical power modules, achieving unprecedented power density in vertical power supply solutions [9]. - The company’s 6 kW 800V to 12V demonstration board has a power density exceeding 2300W/in³, with peak efficiency reaching 97.4% [14]. - Solid-state transformer (SST) technology will play a key role in the new distribution infrastructure, capable of receiving energy directly from medium-voltage AC grids [24].

Core Argument - The over-reliance on Reinforcement Learning (RL) in AI development may be leading the industry away from achieving Artificial General Intelligence (AGI), as current models lack the ability to learn autonomously from experience like humans do [3][4]. Group 1: Skills Preconditioning Paradox - Current AI models depend on "pre-baked" skills, such as using Excel or browsing the web, which highlights their lack of general learning capabilities, indicating that AGI is not imminent [5]. - The approach of embedding specific skills into models contradicts the essence of human-like learning, which does not require extensive pre-training for every task [4][17]. Group 2: Insights from Robotics - The challenges in robotics stem from algorithmic issues rather than hardware limitations; if AI had human-like learning capabilities, robots would already be widely adopted without the need for repetitive training [6][13]. Group 3: Economic Implications of AI - The argument that "technology diffusion takes time" is seen as a self-comforting excuse; if models truly possessed human-like intelligence, they would be rapidly adopted by businesses due to lower risks and no training requirements [7][19]. - The disparity between the value created by global knowledge workers, amounting to trillions of dollars, and the significantly lower revenue generated by AI models indicates that these models have not yet reached the threshold to replace human workers [8][22]. Group 4: The Importance of Continual Learning - The key bottleneck for achieving AGI lies in the ability for "Continual Learning," rather than merely stacking RL computational power; true AGI may take another 10 to 20 years to realize [9][25]. - The process of solving the continual learning problem is expected to be gradual, similar to the evolution of context learning capabilities, and may not yield immediate breakthroughs [29][30].

Core Viewpoint - ByteDance plans to increase its capital expenditure to 160 billion RMB (approximately 23 billion USD) in 2026, focusing on AI infrastructure development, particularly in advanced semiconductor chip procurement for AI model and application development [2][3]. Group 1: Investment Plans - ByteDance's capital expenditure for 2026 is set to rise from 150 billion RMB this year to 160 billion RMB, with a significant portion allocated for AI infrastructure [2]. - Approximately half of the planned investment will be directed towards purchasing advanced semiconductor chips, with an estimated 85 billion RMB earmarked for AI processors in the upcoming year [2]. Group 2: Competitive Landscape - ByteDance is positioning itself as a leading player in the global AI competition, with its Doubao model dominating the consumer-facing AI application sector in China [4]. - The company has seen a substantial increase in daily token usage, exceeding 30 trillion times in October, compared to Google's 43 trillion times during the same period [4]. - ByteDance's flexibility as a non-public company allows it to make aggressive investments and long-term strategic moves in the AI sector, distinguishing it from other major Chinese tech firms like Alibaba and Tencent [4]. Group 3: Market Context - The planned investment reflects the proactive stance of Chinese tech companies in the AI competition, although it remains significantly lower than the combined 300 billion USD spent by major US tech giants like Microsoft, Alphabet, Amazon, and Meta on AI data center construction this year [3].