高分模型未必懂科学，有的只是在「死记硬背」！MIT揭秘：模型越聪明，对物质的理解就越趋同。既然真理路径已清晰，我们何必再深陷昂贵的算力竞 赛？ 现在的AI for Science，就像一场「多国峰会」，大家用不同的语言描述同一件事。 有人让AI读SMILES字符串，有人给AI看原子的3D坐标，大在不同的赛道上比谁预测得准。 但有一个问题：这些AI是在「找规律」，还是真的理解了背后的物理真相？ 在MIT的一项研究中，研究员把59个「出身」不同的模型凑在一起，观察它们在理解物质时，隐藏层表达是否相同 。 论文链接：https://arxiv.org/abs/2512.03750 结果非常惊人：虽然这些模型看数据的方式天差地别，但只要它们变得足够强大，它们对物质的理解就会变得极度相似 。 更神奇的是，一个读文字的代码模型，竟然能和一个算受力的物理模型在「认知」上高度对齐 。 它们沿着不同的路，爬到了同一座山峰的顶端，开始共同描绘物理与现实的「终极地图」。 真理的汇合：为什么顶尖模型越长越像？ 为了验证这些模型是否真的在靠近真理，研究者引入了一个关键指标：表征对齐度。 简单来说，就是看两个模型在处理同一个分子时，它们 ...

当地时间1月7日，谷歌母公司Alphabet迎来其资本市场的一个重要时刻，其股价逆市上涨2.43%，收报 321.98美元，市值增至3.88万亿美元，自2019年以来首次超越苹果公司，并拿下美股市值第二名的宝 座。 2025年以来，得益于在AI领域的卓越表现，Alphabet股价累计飙升逾60%，创下2009年金融危机以来的 最佳年度表现，并领跑美股七大科技巨头。反观苹果，去年其股价累计涨幅不足10%，在七巨头中仅略 高于亚马逊，其搭载AI功能的新一代Siri助手迟迟不能如期发布，加之硬件更新预期提前透支，导致市 场对其增长想象力的预期持续下降。 谷歌与苹果之间的的市值沉浮，显示出的不仅是市场对两家企业AI布局进展的"用脚投票"，更是AI时代 的到来，对科技巨头发展前景与产业话语权的深刻重构。当应用开发平台从Windows系统和个人电脑向 AI迁移，底层计算体系从以CPU为核心转向以GPU为核心，那些真正掌握AI核心能力的企业，才更有 可能在未来的科技变革中占得先机。 近日，谷歌母公司Alphabet以高达47.5亿美元对价，全资收购数据中心能源供应商Intersect的举措，即受 到业界广泛关注。作为大型科 ...

出走5年，估值翻倍！曾被嘲笑「太保守」的Anthropic，正凭3500亿美元身价硬刚OpenAI。看理想主义者如何靠极致安全与Coding神技，在 ARR激增的复仇路上，终结Sam Altman的霸权！ 2026开年最震撼的消息！Anthropic计划融资100亿美金，仅半年，估值就从1830亿涨至3500亿。 这不仅是数字的狂飙，更是一场筹谋5年的「完美反击」。 当OpenAI深陷人才流失与烧钱的泥沼时，曾被称为「叛逃者」的Amodei兄妹，正带着Claude向旧主的王座发起冲锋。 历史回溯：「叛逃者」的初心 这场价值3500亿美元的逆袭，早在5年前就已埋下伏笔。 2021年，随着微软的巨额注资，OpenAI逐渐从非盈利转向「利润上限」模式。 由于理念分歧，Dario Amodei和Daniela Amodei兄妹带着7名核心成员毅然离职。 他们一直担心，当AI进化的速度超越了人类的治理能力，安全是否会被商业利益献祭？ 当时，硅谷将他们视为「理念偏执者」，认为在算力竞争中，空谈安全无异于自缚手脚。 之后的5年，两家公司几乎在所有关键决策上，选择了完全相反的方向。 Sam Altman选择了一条更激进的扩张 ...

AI亦难逃"流量魔咒"，市场正呈现与互联网时代惊人相似的推广逻辑。 争夺春晚流量盛宴，就是最新具象的表现。作为年终最大的流量池，抢占春晚舞台，已经成为争抢超级市场应用的快速入口。此前，无论是微信摇一 摇，"集五福"抢红包，甚至京东、抖快和B站，都曾站上春晚的舞台。 而今年，这个位置，AI上桌了。12月24日，有报道表示，字节跳动旗下火山引擎将成为2026年中央广播电视总台春晚独家AI云合作伙伴，其智能助手豆 包也将配合上线多种互动玩法。 种种消息表明，豆包距离一款国民级应用的距离，越来越近。而凭借AI时代的超级入口，字节也正形成"流量—产品—流量"的闭环。 实际上，不仅仅是字节，放眼过去，无论是OpenAI的GPT系列、谷歌的Gemini，还是DeepSeek以及其他AI应用，其市场策略无不遵循着"争夺流量入 口"的原则。 这一现象，在互联网大厂尤为明显。数据显示，2025年第一季度，全球前十大AI应用的用户获取成本中，流量渠道占比平均达到68%，而产品差异化功能 投入仅占22%。这一比例与2015年移动互联网应用的投入结构几乎一致。 -01- "得流量者得天下"？ 剥开技术炫酷的外衣，审视这场席卷全球的AI ...

OpenAI的十年，是从理想主义的乌托邦，跌入商业现实的修罗场，再奋力攀登技术与商业双重巅峰的传奇历程。它不仅仅是一家公司的故事，更是这个 时代技术狂热、资本博弈、伦理困境和未来憧憬的集中表现，进而可预见的未来，OpenAI最有可能走向三种截然不同的命运： 风暴的前夜 2010年代中期的硅谷，AI人才争夺战已进入白热化阶段。谷歌在2014年以6.5亿美元收购DeepMind，此举震撼了整个行业。与此同时，Facebook（现 Meta）也在不遗余力地网罗AI专家。谷歌与DeepMind的联盟在强化学习等领域展现出强大统治力，而Facebook则深耕于计算机视觉和社交网络AI应 用。 这种"双寡头"格局引起了硅谷其他力量的不安。埃隆·马斯克多次公开表达对人工智能失控风险的担忧。在他看来，将关乎人类命运的技术集中于少数几 家以利润为导向的商业公司手中，是极其危险的。与此同时，时任著名创业孵化器Y Combinator（YC）总裁的山姆·奥特曼正凭借对技术趋势的敏锐嗅 觉和强大的资源整合能力，在硅谷精英网络中建立起独特的影响力——他既看到了AI的巨大潜力，也洞悉了现有格局的弊端。 一场旨在打破垄断、以更安全方式引领 ...

文 | 划重点KeyPoints，作者 | 李越 12月18日，2025腾讯ConTech大会暨腾讯科技Hi Tech Day正式播出，中国工程院院士、知名专家和学 者、头部科技企业创始人及知名投资人齐聚一堂，共同探讨智能时代的机遇与挑战。 原本能够带领我们通往AGI的Transformer，是否已经触碰到了天花板？ 只会做题的优等生 在2017年之前，AI自然语言处理（NLP）的主流方式还是RNN（循环神经网络）和LSTM（长短期记忆 网络）。它们处理信息的方式像一个勤恳的阅读者，必须按顺序一个字一个字地读，效率低下且难以捕 捉长距离的语义关联。 2017年，Google论文《Attention Is All You Need》横空出世，彻底改变了这一切。 Transformer架构抛弃了循环，引入了"自注意力机制"。它不再按顺序阅读，而是能同时关注句子中的所 有词，并计算它们之间的关联权重。 在圆桌论坛环节，当主持人把话筒递给阶跃星辰首席科学家张祥雨，询问关于模型架构未来时，这位学 术大牛抛出了一枚"深水炸弹"：现有的Transformer架构无法支撑下一代Agent。 而就在不久前，斯坦福大学教授、"A ...

Core Insights - Google has made significant advancements in the field of AI with the release of T5Gemma 2 and FunctionGemma, focusing on small models that can operate efficiently on edge devices [1][3][37] Group 1: T5Gemma 2 Overview - T5Gemma 2 is part of the Gemma 3 family and emphasizes architectural efficiency and multimodal capabilities, distinguishing itself from larger models like Gemini [3][4] - The model is available in three sizes: 270M, 1B, and 4B parameters, showcasing its versatility [5] - T5Gemma 2 outperforms corresponding models in the Gemma 3 series across various benchmarks, particularly in code, reasoning, and multilingual tasks [9][11] Group 2: FunctionGemma Overview - FunctionGemma is designed for function calling optimization, allowing it to run on mobile devices and browsers, making it suitable for applications like voice assistants and home automation [7][40] - The model has 270M parameters and is optimized for specific tasks, demonstrating that smaller models can achieve high performance in targeted areas [44][46] - FunctionGemma aims to transition AI from a conversational interface to an active agent capable of executing tasks and interacting with software interfaces [43][56] Group 3: Architectural Innovations - T5Gemma 2 represents a return to the encoder-decoder architecture, which is seen as a modernized revival of classical Transformer models, contrasting with the dominant decoder-only models like GPT [14][30] - The model's architecture allows for better handling of "hallucination" issues and provides inherent advantages in multimodal tasks [32][34] - Google employs a technique called "model adaptation" to efficiently train T5Gemma 2, leveraging existing models to reduce computational costs [36] Group 4: Strategic Implications - The release of these models reflects Google's strategic positioning in the AI landscape, particularly in mobile computing and edge AI, as it seeks to maintain control over the Android ecosystem [52][64] - FunctionGemma's design philosophy aims to democratize AI capabilities across various applications, making advanced functionalities accessible to developers without significant infrastructure costs [64] - By establishing a standard protocol for AI interactions with applications, Google is enhancing its competitive edge in the mobile AI market [57][58]

加州大学伯克利分校（UC Berkeley）刚刚发布了一份重磅论文：《Measuring Agents in Production》。 （论文地址：https://arxiv.org/pdf/2512.04123） 这份论文，基于来自全球的真实请求：306名从业者深度调研，20个企业级部署案例，覆盖 26 个行业。 这是AI Agent 领域，迄今最大规模的实证研究。 最核心的三个信息： 这份报告信息非常多，容我慢慢道来。 73%为生产力买单，金融成Agent 第一战场 先说一个数字： 73%的从业者表示，部署Agent的首要目的是"提高生产力"。 其中，金融与银行业是第一大战场，占比39.1% 其次是科技（24.6%）和企业服务（23.2%） 。 除了这些，Agent 还在很多意想不到的地方落地： 保险理赔流程自动化：代理人负责处理从保单查询到风险识别的序列排序流程。 生物医学工作流自动化：在科学发现领域，Agent 用于自动化执行复杂的实验和数据分析流程。 企业内部运营支持：涵盖人力资源信息搜索、站点故障事件诊断等多个方面。 这些跨行业的成功案例证明，AI Agent已经具备解决真实世界复杂问题的能力，并 ...

Core Insights - The report titled "State of AI: An Empirical 100 Trillion Token Study with OpenRouter" analyzes the usage of over 300 AI models on the OpenRouter platform from November 2024 to November 2025, focusing on real token consumption rather than benchmark scores [3][5][67] - It highlights the significant rise of open-source models, particularly from China, which saw weekly token usage share increase from 1.2% to 30%, indicating a shift towards a complementary relationship between open-source and closed-source models [2][10][74] - The report emphasizes the transition of AI models from language generation systems to reasoning and execution systems, with reasoning models becoming the new paradigm [18][83] Open-Source vs Closed-Source Models - Open-source models are no longer seen merely as alternatives to closed-source models; they have carved out unique positions and are often preferred in specific scenarios [6][70] - By the end of 2025, it is expected that open-source models will account for approximately one-third of total usage, reflecting a more integrated approach by developers who utilize both types of models [5][70] - The dominance of DeepSeek is diminishing as more open-source models enter the market, leading to a diversified landscape where no single model is expected to exceed 25% of token usage by the end of 2025 [13][77] Model Characteristics and Trends - The report identifies a shift towards medium-sized models, which are gaining market favor, while small models are losing traction [16][80] - The classification of models is as follows: large models (700 billion parameters or more), medium models (150 to 700 billion parameters), and small models (less than 150 billion parameters) [20][85] - The usage of reasoning tokens has surpassed 50%, indicating a significant evolution in how AI models are utilized for complex tasks [18][83] User Behavior and Model Utilization - AI model usage has evolved from simple tasks to more complex problem-solving, with user prompts increasing in length and complexity [27][92] - The concept of "crystal shoe effect" is introduced, where certain models lock in a core user base due to their unique capabilities, making it difficult for competitors to attract these users later [55][120] - Programming and role-playing have emerged as the primary use cases for AI models, with programming queries rising from 11% to over 50% [27][100] Market Dynamics - The report notes that the paid usage share of AI in Asia has doubled from 13% to 31%, while North America's share has fallen below 50% [129] - English remains the dominant language in AI usage at 82%, with Simplified Chinese holding nearly 5% [129] - The impact of model pricing on usage is less significant than anticipated, with a 10% price drop leading to only a 0.5%-0.7% increase in usage [129]

Core Insights - The report titled "State of AI: An Empirical 100 Trillion Token Study with OpenRouter" analyzes the usage of over 300 models on the OpenRouter platform from November 2024 to November 2025, focusing on real token consumption rather than benchmark scores [3][6][8]. Group 1: Open Source vs. Closed Source Models - Open source models (OSS) have evolved from being seen as alternatives to closed source models to finding their unique positioning, becoming the preferred choice in specific scenarios [9]. - The relationship between open source and closed source models is now more complementary, with developers often using both types simultaneously [10]. - The usage of open source models is expected to reach approximately one-third by the end of 2025, with Chinese models experiencing significant growth from 1.2% to 30% in weekly usage share [12][13]. Group 2: Market Dynamics and Model Diversity - The dominance of DeepSeek as the largest contributor to open source model usage is diminishing as more models enter the market, leading to a diversified landscape [16]. - By the end of 2025, no single model is expected to maintain over 25% of token usage, with the market likely to be shared among 5 to 7 models [17][18]. - The report indicates a shift towards medium-sized models, which are gaining market favor, while small models are losing traction [20][21]. Group 3: Evolution of Model Functionality - Language models are transitioning from dialogue systems to reasoning and execution systems, with reasoning token usage surpassing 50% [22]. - The use of model invocation tools is increasing, indicating a more competitive and diverse ecosystem [29][31]. - AI models are evolving into "intelligent agents" capable of independently completing tasks rather than just responding to queries [43]. Group 4: Usage Patterns and User Retention - The complexity of tasks assigned to AI has increased, with users now requiring models to analyze extensive documents or codebases [35]. - The average input to models has quadrupled, reflecting a growing reliance on contextual information [36]. - The "glass slipper effect" describes how certain users become highly attached to models that perfectly meet their needs upon release, leading to high retention rates [67][70]. Group 5: Regional Insights and Market Trends - The share of paid usage in Asia has doubled from 13% to 31%, indicating a shift in the global AI landscape [71]. - North America's AI market share has declined to below 50%, while English remains dominant at 82%, with Simplified Chinese holding nearly 5% [80]. - The impact of model pricing on usage is less significant than expected, with a 10% price drop resulting in only a 0.5%-0.7% increase in usage [80].