Scaling Law
Search documents
谷歌用Gemini 3同时革了OpenAI和英伟达两家的命
3 6 Ke· 2025-11-26 10:39
当所有人都以为AI发展的剧本是「英伟达卖铲子,OpenAI挖金矿」时,谷歌用Gemini 3告诉世界:如果我自己造了一台全自动挖掘机,还需要买谁的铲 子,谁能挖得过我? 在谷歌最新的Gemini 3和Nano Banana Pro发布前,黄仁勋和奥特曼都一直雄心勃勃,自信非常。 英伟达和OpenAI的日子过得可谓是风生水起。 Nano Banana Pro生成 然而,Gemini 3全家桶的横空出世,瞬间打断了这场双赢的美梦。 彭博、财富等外媒则更加直接,他们认为: 谷歌,这家曾被认为在AI时代稍显落后、有点沉睡的巨头,正在全面觉醒。 在OpenAI的蓝图里,AI的发展会一直遵循着一种简单而傲慢的「暴力美学」:Scaling Law。 他们坚信,只要数据更多、算力更强,模型就会无限聪明下去,而他们握着通往AGI的唯一钥匙。 就像从GPT-3一直到现在GPT-5、GPT-5.1,然后明年的GPT-6、一直到AGI实现的那一天。 而在英伟达的账本里,逻辑则更加粗暴直接:我是这个淘金时代唯一卖铲子的人。(不是非常严格,但基本事实如此) 全世界的AI公司,无论谁输谁赢,都得乖乖排队交「GPU税」。 然而硅谷的夜,这几 ...
机械设备行业点评报告:GoogleGemini3表现超预期,看好AI算力需求的成长性
Soochow Securities· 2025-11-26 06:35
机械设备行业点评报告 Google Gemini 3 表现超预期,看好 AI 算力 需求的成长性 证券研究报告·行业点评报告·机械设备 增持(维持) [Table_Tag] [投资要点 Table_Summary] ◼ 事件:Google Gemini 3 发布,表现超市场预期 近期 Google 发布了最新的多模态大模型 Gemini 3,展现出超市场预 期的得分能力与多模态理解能力。 ◼ Google Gemini 3 在 Benchmark 测试上实现断层领先 Gemini 3 问世前,顶尖大模型之间的 Benchmark 测试得分差距微小, Gemini 3 推出后在 Benchmark 测试上的得分实现了断层式领先:①HLE 测 试基础思考能力得分 37.5%(无工具),领先于 Gemini 2.5 Pro 的 21.6% 和 GPT-5.1 的 26.5%;②多模态理解能力上,ScreenSpot-Pro(截图理解) 测试得分 72.7%,远超 Claude Sonnet 4.5 的 36.2%和 GPT-5.1 的 3.5%。 Gemini 3 的另一大亮点是"生成式 UI"的强大审美能力,逐步 ...
中兴发了一篇论文,洞察AI更前沿的探索方向
机器之心· 2025-11-26 01:36
Core Insights - The AI industry is facing unprecedented bottlenecks as large model parameters reach trillion-level, with issues such as low efficiency of Transformer architecture, high computational costs, and disconnection from the physical world becoming increasingly prominent [2][4][38] - ZTE's recent paper, "Insights into Next-Generation AI Large Model Computing Paradigms," analyzes the core dilemmas of current AI development and outlines potential exploratory directions for the industry [2][38] Current State and Bottlenecks of LLMs - The performance of large language models (LLMs) is heavily dependent on the scaling laws, which indicate that ultimate performance is tied to computational power, parameter count, and training data volume [4][5] - Building advanced foundational models requires substantial computational resources and vast amounts of training data, leading to high sunk costs in the training process [5][6] - The efficiency of the Transformer architecture is low, with significant memory access demands, and the current hardware struggles with parallel operations in specific non-linear functions [6][7] Challenges in Achieving AGI - Current LLMs exhibit issues such as hallucinations and poor interpretability, which are often masked by the increasing capabilities driven by scaling laws [9][10] - There is ongoing debate regarding the ability of existing LLMs to truly understand the physical world, with criticisms focusing on their reliance on "brute force scaling" and lack of intrinsic learning and decision-making capabilities [9][10] Engineering Improvements and Optimizations - Various algorithmic and hardware improvements are being explored to enhance the efficiency of self-regressive LLMs, including attention mechanism optimizations and low-precision quantization techniques [12][13][14] - Innovations in cluster systems and distributed computing paradigms are being implemented to accelerate training and inference processes for large models [16][17] Future Directions in AI Model Development - The industry is exploring next-generation AI models that move beyond the Next-Token Prediction paradigm, focusing on models based on physical first principles and energy dynamics [24][26] - New computing paradigms, such as optical computing, quantum computing, and electromagnetic computing, are being investigated to overcome traditional computational limitations [29][30] ZTE's Exploration and Practices - ZTE is innovating at the micro-architecture level, utilizing advanced technologies to enhance AI accelerator efficiency and exploring new algorithms based on physical first principles [36][38] - The company is also focusing on the integration of hardware and software to create more efficient AI systems, contributing to the industry's shift towards sustainable development [38]
CPO、光通信模块板块爆发,5GETF、5G通信ETF、通信ETF、创业板人工智能ETF涨超3%
Ge Long Hui A P P· 2025-11-25 07:57
Market Performance - The three major A-share indices rose collectively, with the Shanghai Composite Index up 0.87% to 3870 points, the Shenzhen Component Index up 1.53%, and the ChiNext Index up 1.77% [1] - The total market turnover reached 1.83 trillion yuan, an increase of 858 billion yuan compared to the previous trading day, with 4300 stocks rising [1] Sector Highlights - The CPO concept and optical communication module sectors experienced significant growth, with stocks like Zhongji Xuchuang rising by 5%, Xinyi Sheng by 4%, and Tianfu Communication by 2.4% [1] - Various ETFs related to 5G and artificial intelligence also saw substantial gains, with the 5GETF rising over 4% and several other ETFs increasing by more than 3% [1][2] ETF Details - The 5G Communication ETF tracks an index that includes key components of AI computing systems, with significant weight in optical modules and leading companies in the 5G industry [4] - The Communication ETF's index has over 81% weight in "optical modules + servers + copper connections + optical fibers" [4] - The ChiNext Artificial Intelligence ETFs have over 50% "CPO content," featuring major stocks like Xinyi Sheng and Zhongji Xuchuang [4] AI and Technology Developments - Google is challenging NVIDIA's dominance in chips by potentially selling its TPU to Meta, which could capture 10% of NVIDIA's annual revenue, leading to billions in new income for Google [5] - Google’s Gemini 3 model demonstrates the ongoing effectiveness of the Scaling Law, indicating further advancements in algorithms and a growing demand for computing power [6] - The AI industry is expected to see significant catalysts by 2026, including new GPU releases and advancements in AI applications, with a positive outlook for sectors like optical modules and AI smartphones [7]
AI巨头们的万亿美元债务去哪了?
Tai Mei Ti A P P· 2025-11-24 04:42
Core Insights - Meta plans to invest $60 billion in AI despite reporting a net profit of $37 billion in the first three quarters of 2025, highlighting the financial challenges faced by tech giants in the AI arms race [1][2] Financing Challenges - The need for massive funding for AI infrastructure, including expensive AI chips and data centers, poses a dilemma for tech giants on how to secure funds without negatively impacting their financial statements [2][3] - Morgan Stanley estimates that "invisible debt" could reach $800 billion by 2028, representing significant liabilities that do not appear on the balance sheets of these companies [2] SPV Financing Method - The Special Purpose Vehicle (SPV) financing method allows tech giants to isolate debt and optimize their financial reports by transferring the debt to a separate entity [3][4] - This method involves creating an SPV to borrow money using the parent company's credit, allowing the SPV to purchase assets and lease them back to the parent company, thus keeping the debt off the parent company's balance sheet [4] Examples of SPV Utilization - Meta successfully utilized this SPV method to increase its debt by $30 billion on its balance sheet while leveraging it to acquire $60 billion in computing assets [4] - Google has adopted a similar strategy by providing credit guarantees to weaker companies, allowing them to secure loans for data center assets, which are then leased back to Google [5] Circular Financing - The concept of circular financing allows companies to create a closed loop of capital flow among related parties, enhancing financial efficiency [7] - For instance, xAI established an SPV to raise $20 billion for purchasing NVIDIA chips, with minimal direct debt risk, showcasing the flexibility of this financing model [7] Industry Dynamics - Major tech companies are forming strategic alliances to create a tightly-knit capital community, which can amplify their financial capabilities and market influence [9][10] - Recent collaborations among giants like OpenAI, NVIDIA, and Oracle have resulted in over $1 trillion in infrastructure and chip agreements, indicating a trend towards deeper integration in the AI sector [9] Scaling Law and Market Sentiment - The pursuit of Scaling Law drives exponential growth in computing demand, benefiting companies like NVIDIA, which has seen significant revenue increases [15] - However, industry leaders express caution regarding potential irrational exuberance in AI investments, with warnings about the risks of a bubble [15][16] Capital Market Movements - Notable investors are shifting their strategies, with significant sell-offs in NVIDIA stock while simultaneously investing in AI applications and models, indicating a transition in focus from hardware to software [16][17] - This shift suggests that while financing challenges may be temporarily addressed, the competition in the AI landscape is just beginning, with a more intense focus on applications and models ahead [17]
拆解Gemini 3:Scaling Law的极致执行与“全模态”的威力
3 6 Ke· 2025-11-24 03:55
Core Insights - Google’s Gemini 3 has transformed the AI landscape in Silicon Valley, positioning the company as a leader rather than a follower in the AI race against OpenAI and Anthropic [1][3] - Gemini 3 is recognized for its significant advancements in multimodal capabilities and is seen as a prime example of executing Scaling Law effectively [1][3] Performance Evaluation - Within 48 hours of its release, Gemini 3 topped various performance rankings, showcasing its true multimodal native model capabilities [4][6] - Users reported that Gemini 3 provides a more integrated development experience, particularly with tools like Google AntiGravity, which enhances coding efficiency by allowing simultaneous visual and coding tasks [6][7] Technical Innovations - The model achieved a notable improvement in Few-shot Learning, reaching over 30% on the ARC-AGI-2 Benchmark, indicating a qualitative leap in its reasoning capabilities [10][11] - Gemini 3 employs a tree-based thought process and self-rewarding mechanisms, allowing it to explore multiple reasoning paths simultaneously [19][20] Developer Ecosystem - The release of Gemini 3 and AntiGravity has led to discussions about the end of the coding competition, as Google’s ecosystem may create significant barriers for startups like Cursor [22][23] - Despite the strong capabilities of AntiGravity, it still faces challenges in backend deployment and complex system architecture, suggesting that independent developers may still find opportunities in niche areas [25][26] Future Trends in AI - The focus is shifting towards new AI paradigms beyond LLMs, with emerging labs like NeoLab attracting significant venture capital [27][28] - There is a growing interest in developing world models that understand physical laws, indicating a potential shift in AI research directions [31][32] Conclusion - The launch of Gemini 3 serves as a robust counter to the "AI bubble" narrative, demonstrating that with sufficient computational power and engineering optimization, Scaling Law remains a viable path for AI advancement [32][33]
活动报名:AI 的机会与泡沫|42章经
42章经· 2025-11-23 13:01
Group 1 - The core viewpoint of the article discusses the current state of the AI market, highlighting that the growth from 2023 to 2024 relies on the scaling law and the consensus around AGI, while there is no unified judgment on RL scaling law since 2025 [5] - AI models are developing in a stepwise manner, while applications are experiencing pulsed advancements, indicating a subtle blank period currently [5] - There is uncertainty regarding the continued enhancement of intelligence, but the acceleration of application deployment is assured [5] Group 2 - The narrative logic is changing, suggesting that while prices that rose previously may have bubbles, the intrinsic value of AI remains intact [5] - Several unresolved questions about the future development of AI, including whether to buy or short Nvidia, the opportunities in multimodal applications, and the feasibility of embodied production and deployment, are raised [5] - An online discussion meeting is scheduled for November 29, aiming to engage in these topics with interested participants [5]
【兴证计算机】AI应用:谷歌王者归来,商业奇点临近
兴业计算机团队· 2025-11-23 09:19
Core Viewpoint - The market is experiencing a decline in risk appetite, suggesting that investors should increase positions in certain directions and leading stocks during this period of volatility [1] Group 1: Market Analysis - The current market environment indicates a preference for stocks with cross-year certainty, focusing on valuation, earnings growth, and industry prosperity changes as core considerations [1] - The overall allocation in the computer sector is currently low, presenting a comparative advantage for positioning ahead of the spring rally [1] Group 2: AI Application Insights - Google's recent releases of Gemini3 and Nano Banana Pro have demonstrated significant performance improvements, reaffirming the effectiveness of Scaling Law and indicating sustained high demand in the AI sector [2] - The launch of xAI's Grok4.1 model and the public testing of Qianwen APP by Ant Group highlight ongoing advancements in AI capabilities, suggesting that the industry may be approaching a commercial singularity [2]
Generalist发现具身智能的Scaling Law,还让模型能同时思考与行动
3 6 Ke· 2025-11-21 01:52
Generalist是Google DeepMind高级研究科学家Pete Florence创立的具身智能模型公司。近日,它发布了一款叫GEN-0的新型具身基础模型,这个模型能够随 着物理交互数据,而非仅仅是文本、图像或模拟数据的增长而可预测地扩展,在训练这个模型的过程中,他们还一定程度证实了具身智能的Scaling Law。 Generalist的早期投资者包括Spark Capital、NVIDIA、Boldstart Ventures、Bezos Expeditions、NFDG等投资机构,但金额未披露。 DeepMind和波士顿动力的专家一起探索具身智能的Scaling Law Generalist由Google DeepMind高级研究科学家Pete Florence联合创立,他在Google带队研发了PaLM-E,RT-2等视觉或具身智能模型,Google学术的引用数 超过19000次。 与Pete Florence共同创立Generalist AI的还有Andrew Barry(CTO)和Andy Zeng(首席科学家)。Andrew Barry此前在波士顿动力任职,Andy Zeng则与 Pe ...
GEN-0 以及后续的 VLA 发展的看法
具身智能之心· 2025-11-21 00:04
作者丨 阿汐猫猫 原文链接 | https://zhuanlan.zhihu.com/p/1970094649956868665 点击下方 卡片 ,关注" 具身智能之心 "公众号 >> 点击进入→ 具身 智能之心 技术交流群 更多干货,欢迎加入国内首个具身智能全栈学习社区: 具身智能之心知识星球(戳我) ,这里包含所有你 想要的! 文章转载自博客,见 https://axi404.top/blog/embodied-talk-3 前言 最近 GEN-0[1] 的发布对于具身智能领域可以说是轰动性的。Manipulation 作为 Robotics 领域一直以来皇冠上 的明珠,并且作为具身智能带来现实生产力必不可少的一环,一向以泛化的困难性著称。由于缺乏实际的使 用场景,缺乏数据飞轮导致的数据匮乏使得模型的预训练难以 scaling up,而模型高度依赖后训练的数据。 在此之前,领域内最具代表性的工作莫过于 Pi 系列[2][3],在 Pi dataset 私有数据集上进行预训练。其结果是 显著的,使用此类预训练之后,带来了模型后训练时的性能提升。从实际部署中,Pi 不同于若干号称反超自 己的模型,在动作连贯性 ...