晓静 腾讯科技《AI未来指北》特约作者 2025 年 6 月 6 日，第七届北京智源大会在北京正式开幕，强化学习奠基人、2025年图灵奖得主、加拿 大计算机科学家Richard S. Sutton以"欢迎来到经验时代"为题发表主旨演讲，称我们正处在人工智能史上 从"人类数据时代"迈向"经验时代"的关键拐点。 Sutton指出，当今所有大型语言模型依赖互联网文本和人工标注等"二手经验"训练，但高质量人类数据 已被快速消耗殆尽，新增语料的边际价值正急剧下降；近期多家研究也观察到模型规模继续膨胀却收效 递减的"规模壁垒"现象，以及大量科技公司开始转向合成数据。 以下为演讲全文： 当前大型模型已逼近"人类数据"边界，唯有让智能体通过与环境实时交互来生成可随能力指数级扩 张的原生数据，AI 才能迈入"经验时代" 。 真正的智能应像婴儿或运动员那样在感知-行动循环中凭第一人称经验自我学习 。 强化学习范例（如 AlphaGo、AlphaZero）已证明从模拟经验到现实经验的演进路径，未来智能体 将依靠自生奖励和世界模型实现持续自我提升 。 基于恐惧的"中心化控制"会扼杀创新，多主体维持差异化目标并通过去中心化合作实现双赢 ...

腾讯广告技术 . 腾讯广告技术官方阵地，分享团队最新前沿成果及广告技术应用。 聚焦全模态生成式推荐 向技术天才发出 邀请函 今年4月，腾讯宣布启动史上最大就业计划，三年内将新增28000个实习岗位并加大转化录用，其中2025 年将迎来10000名校招实习生，有六成面向技术人才开放。腾讯广告其业务形态因对算法实时性和复杂 度的极致要求，一直是前沿AI技术的重要应用阵地。我们也期待通过腾讯广告算法大赛，为高校学子提 供兼具前瞻性与实战价值的演练场，角逐AI技术之巅。 今日，2025 腾讯广告算法大赛正式启动！这场以 「智 AI，荐未来」 为主题的全球性赛事，聚焦 "全模 态生成式推荐"（All-Modality Generative Recommendation） 这一前沿课题，不仅是学术与产业碰撞 的舞台，更是技术人才直通腾讯核心业务、以生成式推荐新范式重构行业生态、抢占职业发展先机的黄 金机会。 标题 三大核心亮点 与顶尖资源同行，为技术理想而战 院士领衔评审天团，技术视野与产业洞察双重加持 以下文章来源于腾讯广告技术 ，作者腾讯广告算法大赛 大赛特邀中国科学院胡事民院士、香港中文大学金国庆教授、北京大学崔斌 ...

Group 1: Models - OpenAI's o3-pro and 4o thinking model are highlighted as significant advancements in AI modeling [2] - Meta's V-JEPA 2 world model and Mistral AI's Magistral reasoning model are also noted for their contributions to the field [2] - MiniCPM 4.0 from 面壁智能 and the open-source dots.llm1 from 小红书 are mentioned as key developments in AI models [2] Group 2: Applications - OpenAI's advanced voice personification and AI math genius applications are recognized for their innovative use of AI technology [2] - ByteDance's 豆包大模型1.6 and 即梦图片3.0 are significant applications in the AI landscape [2] - Other notable applications include Google's Veo 3 Fast version and ElevenLabs' Eleven v3, showcasing the diversity of AI applications [2] Group 3: Technology - Figure AI's labor system and advancements in robotics by 理想汽车 and 荣耀 are discussed as part of the technological progress in AI [3] - NVIDIA's quantum CUDA-Q and Apple's six major OS updates reflect ongoing technological innovations [3] - The启蒙系统 from 中科院 is also mentioned as a significant technological development [3] Group 4: Perspectives - Altman discusses the timeline for AGI technology, while Ilya Sutskever emphasizes AI's potential to accomplish everything [3] - OpenAI raises concerns about human dependency on AI, and Sergey Levine engages in a discussion about the essence of large models [3] - Richard Sutton introduces the concept of an experience era, indicating a shift in how AI is perceived and utilized [3] Group 5: Capital and Events - Meta's investment in Scale AI and the establishment of a superintelligence reconstruction group are significant events in the AI investment landscape [3][4] - The copyright lawsuit involving Midjourney and a large-scale nuclear power agreement by Meta are also noteworthy events [4]

科研就是不断探索问题的边界 人们过去花了数个世纪来接纳数字"零"的存在。而今，"零"正在帮助神经科学家们理解人脑如何感知虚无。与感知和意识相关的 神经科学研究，大多聚焦于我们如何意识到事物的"存在"。然而，对"不存在"的体验也构成了我们意识体验的重要组成部分—— 我们经常能觉察到那些肉眼无法看见的事物，而揭示这背后的神经基础对充分理解意识问题同样重要。 当我观鸟时，总是遇到这样一个尴尬场景——同行的观鸟人指着树冠，让我快看叶子后面藏着的那只 鸟。而每当我举起望远镜来回搜寻时，永远只能沮丧地看见鸟的"空影"。 这类对"不存在之物"的生动体验，对于我们的内心世界而言非常常见，但大脑如何上演这出"皇帝新 衣"式的独角戏仍是个谜—— 当没有任何东西可供感知时，大脑如何产生感知体验？ 作为一个对意识问题感兴趣的神经科学家，研究"虚无"的神经基础无疑是个极其诱人而又富有挑战的课 题。幸运的是， 比起其他虚无，有一种更具体的虚无形式——0，至少0是有形的。 为此，人们不惜花 上大量精力，尝试抓住"零"这个线索——研究人脑如何感知数字"零"，或许就能够最终解开大脑迷雾重 重的"虚无主义"。 "零"在人类社会的发展中扮演了一个 ...

生成式AI 一、 Meta开源发布V-JEPA 2世界模型：开启物理推理新时代 1. Meta开源V-JEPA 2世界模型，能理解物理世界，用100万小时视频数据训练，可实现零 样本规划和机器人控制； 2. 模型仅需62小时训练即可生成规划控制模型，在行为分类和预测方面达到顶级表现，成功 率达65%-80%； 3. Meta发布三个物理理解基准测试，揭示AI与人类在物理推理能力上仍存在差距，未来将发 展分层次和多模态JEPA模型。 https://mp.weixin.qq.com/s/M1mKgpz4ecCIC3xKq50k-A 二、 小扎AGI重新 组队， 科技公司的顶尖工程师们应来尽来 1. Meta CEO扎克伯格正组建"超级智能"小组，已成功挖角谷歌DeepMind首席研究员Jack Rae和其他AI顶尖人才； 2. Jack Rae是"压缩即智能"思想代表人物，曾负责Gemini的"思考"功能，在DeepMind工作 7年参与多个重要模型开发； 3. Meta为吸引AI人才提供7-9位数薪酬方案，计划建立约50人规模的团队， 或 以 百 亿美 元收购Scale AI 及 其 团队 。 https:/ ...

Group 1 - The article emphasizes the potential of the AI identification system in addressing the challenges of misinformation, highlighting its role as a crucial front-end support in content governance [1][4] - It points out that over 20% of the 50 high-risk AI-related public opinion cases in 2024 were related to AI-generated rumors, indicating a significant issue in the current content landscape [1][3] - The article discusses the three main challenges posed by AI-generated harmful content: lower barriers to entry, the ability for mass production of false information, and the increased realism of such content [3][4] Group 2 - The introduction of a dual identification mechanism, consisting of explicit and implicit identifiers, aims to enhance the governance of AI-generated content by covering all stakeholders in the content creation and dissemination chain [5][6] - The article notes that explicit identifiers can reduce the credibility of AI-generated content, as studies show that labeled content is perceived as less accurate by audiences [6][8] - It highlights the limitations of the AI identification system, including the ease of evasion, forgery, and misjudgment, which can undermine its effectiveness [8][9] Group 3 - The article suggests that the AI identification system should be integrated into the existing content governance framework to maximize its effectiveness, focusing on preventing confusion and misinformation [11][12] - It emphasizes the need to target high-risk areas, such as rumors and false advertising, rather than attempting to cover all AI-generated content indiscriminately [13][14] - The responsibilities of content generation and dissemination platforms should be clearly defined, considering the challenges they face in accurately identifying AI-generated content [14]

岗位： 腾讯研究院 AI for Good方向实习生 岗位描述 1、 研究方向：AI for Good 2、日常工作包括：数据分析及可视化、报告撰写、创意策划等 任职要求 3、实习期内必须持有学生证（注意：大四保研后将有2个月没有学生证，不符合规定） 待遇 1、工资150元/天（税后） 2、社科或商科，或有设计背景的交叉学科 3、有扎实的实证研究功底 4、熟练使用各类AI工具，同时拥有良好的创造力 5、能熟练使用量化研究工具，有量化研究作品者优先，数据可视化能力强者优先 其他要求 1、2024年6月6日前入职 2、一周可实习4天（不含周末），延续至少4个月 1、工作态度务实、勤恳、守时、负责 2、工作地点为北京朝阳亚洲金融大厦 应聘方式 1、发送简历、既往研究作品至 simonelu@tencent.com 点个 "在看" 分享洞见 2、请尽量多地发送各种能展现个人能力的作品 3、邮件标题请写作： 姓名+学校+专业+到岗时间 ...

Group 1: OpenAI and Mistral AI Developments - OpenAI released the inference model o3-pro, which is marketed as having the strongest reasoning ability but the slowest speed, with input pricing at $20 per million tokens and output at $80 per million tokens [1] - User tests indicate that o3-pro excels in complex reasoning tasks and environmental awareness but is not suitable for simple problems due to its slow inference speed, targeting professional users [1] - Mistral AI launched the strong inference model Magistral, which includes an enterprise version Medium and an open-source version Small (24B parameters), showing excellent performance in multiple tests [2] - Magistral achieves a token throughput that is 10 times faster than competitors, with a pricing strategy of $2 per million tokens for input and $5 per million tokens for output [2] Group 2: Figma and Krea AI Innovations - Figma introduced the official MCP service, allowing direct import of design file variables, components, and layouts into IDEs, achieving a higher fidelity than third-party MCPs [3] - Krea AI launched its first native model Krea 1, focusing on solving issues of AI image "homogenization" and "plasticity," providing high aesthetic control and professional-grade output [4][5] - Krea 1 supports style reference and custom training, with native support for 1.5K resolution expandable to 4K, aimed at accelerating digital art creation processes [5] Group 3: ByteDance and Tolan AI Applications - ByteDance released the Doubao large model 1.6 series, which includes multiple versions supporting 256k context and multimodal reasoning, with a 63% reduction in comprehensive costs [6] - Tolan, an alien AI companion application, has achieved 5 million downloads and $4 million ARR, emphasizing a non-romantic, non-tool-like companionship experience [7] - Tolan's design integrates companionship with gamification, allowing users to customize their alien companion's appearance and develop unique planetary environments [7] Group 4: Li Auto and Figure Robotics Strategy - Li Auto established two new departments, "Space Robotics" and "Wearable Robotics," to enhance its AI strategy, focusing on creating a smart in-car experience [8] - Figure aims to provide a complete "labor force" system with humanoid robots, emphasizing fully autonomous operation and a production line capable of producing 12,000 units annually [9] - Figure plans to deliver 100,000 units over the next four years, targeting both commercial and home markets, while utilizing a shared neural network for collective learning [9] Group 5: Altman's Predictions and OpenAI Codex Insights - Altman predicts that by 2025, AI will be capable of cognitive work, with significant productivity boosts expected by 2030 as AI becomes more affordable [10] - OpenAI Codex is shifting software development from synchronous "pair programming" to asynchronous "task delegation," anticipating a transformation in developer roles by 2025 [11] - The team envisions a future where the interaction interface merges synchronous and asynchronous experiences, potentially evolving into a "TikTok"-like information flow for developers [11]

Core Viewpoint - Google's AI strategy is undergoing a significant transformation, moving towards a new phase of AI platform evolution, integrating AI deeply into advertising and content generation, which may fundamentally reshape the advertising distribution mechanism and business model [1]. Group 1: Evolution of Advertising - The evolution of Google's advertising has progressed from AdWords in 2000 to the introduction of Performance Max in 2021, which marked a shift to AI-generated content and automated multi-channel ad delivery [4][6]. - The recent I/O 2025 conference introduced AI tools like Veo 3, which can convert static images into dynamic video content, significantly lowering the barrier for high-quality video creation [5]. - The new AI capabilities are expected to accelerate the shift from resource-intensive, human-driven creative processes to highly automated, AI-driven content generation, allowing brands to reduce costs and enhance efficiency [7]. Group 2: Personalization Paradigm Shift - Advertising is transitioning from "mass personalization" to "hyper-personalization," where AI integrates directly into Google Search to provide individualized product recommendations based on user intent [9][10]. - The introduction of smart agents allows users to track prices and make purchases automatically, transforming Google from a search engine into a proactive shopping agent [10][11]. - This shift emphasizes the need for brands to adapt to a new advertising interaction model, where each ad interaction is unique and tailored to individual user experiences [11]. Group 3: Integration of Advertising and Search Experience - Google's AI search has gained 1.5 billion monthly active users, with a 10% increase in usage, indicating a shift in user behavior towards complex queries rather than simple searches [14]. - Ads are now integrated into AI-generated answers, becoming part of the useful information rather than separate bidding spaces, which fundamentally alters the advertising ecosystem [14][15]. - The development of generative AI is expected to disrupt traditional advertising value assessments, as the focus shifts from exposure metrics to conversion rates, potentially leading to a structural change in advertising pricing models [15]. Group 4: Future of Advertising Industry - Brands need to rethink their roles in the marketing value chain as AI takes over content generation and ad placement, focusing on being referenced by AI rather than just occupying search result positions [18][19]. - The blurring lines between advertising and content necessitate brands to create proprietary intelligent agents that align with their brand identity and ensure consistency in market presence [19]. - Long-term strategies should focus on achieving a balance between effective advertising conversion and brand influence, leveraging AI for precise targeting and content innovation [19].

Group 1: Apple Developments - Apple has unified the design of six major operating systems, introducing a new "Liquid Glass" element that significantly enhances visual effects [1] - The company has opened access to on-device large language models for all apps, integrating AI functionalities such as visual search and real-time translation [1] - Major updates to iPadOS and enhanced macOS-iPhone integration were announced, but the release of the new Siri has been delayed again [1] Group 2: Developer Tools - Apple announced Xcode 26, which integrates ChatGPT to assist developers in code writing, documentation generation, and error fixing [2] - Developers can introduce AI models from other vendors into Xcode via API keys, fostering a diverse intelligent programming ecosystem [2] - The Foundation Models framework allows developers to call local AI models with just three lines of code [2] Group 3: NoCode Tool by Meituan - Meituan launched the NoCode AI Coding Agent tool, enabling users to create websites and applications without programming [3] - NoCode combines product, design, and engineering functionalities, supporting various application scenarios such as website design and game development [3] - The tool features the ability to understand implicit needs and supports collaborative work, now fully launched and available for free [3] Group 4: Tencent's Yuanbao Upgrade - Tencent's Yuanbao desktop version has upgraded its text selection feature, adding continuous selection for automatic translation [4] - A new window pinning feature allows the translation results window to remain fixed, enhancing reading efficiency [4] - The upgraded functionality is particularly useful for browsing foreign websites and reading English documents [4] Group 5: Meta's Nuclear Power Agreement - Meta signed a 20-year nuclear power purchase agreement with Constellation Energy, with a capacity of 1,121 megawatts from the Clinton Clean Energy Center in Illinois [5] - This agreement surpasses Microsoft's previous collaboration of 835 megawatts, aimed at supporting Meta's growing energy needs for data centers and AI development [5] - The partnership will retain over 1,100 jobs and increase power generation by 30 megawatts, with supply expected to start in 2027 to support Meta's planned 1.3 million GPU scale [5] Group 6: AI Chip Design by Chinese Academy of Sciences - The Chinese Academy of Sciences launched the "Enlightenment" system, achieving fully automated design of processor chips, with performance meeting or exceeding human expert levels [6] - The system has successfully designed the RISC-V CPU "Enlightenment 2," matching the performance of ARM Cortex A53, and can automatically configure operating systems and high-performance libraries [6] - The "Enlightenment" system employs a three-layer architecture and a "three-step" technical route, potentially transforming chip design paradigms and significantly enhancing design efficiency [6] Group 7: AI Voice Interaction Insights - The founder of ElevenLabs suggests that incorporating "imperfections" in AI voice can enhance user interaction, as overly perfect voices may reduce engagement [8] - Future voice agents are expected to possess contextual awareness, transitioning from passive customer service to proactive user experience guidance [8] - As AI voice technology evolves, a new trust mechanism will emerge, focusing on verifying whether content is human-voiced rather than AI-generated [8] Group 8: Richard Sutton's Vision on AI - Richard Sutton, the father of reinforcement learning, believes AI is transitioning from the "human data era" to the "experience era," learning from real-time interactions with the environment [9] - He advocates for a decentralized cooperative model for AI development, opposing centralized control based on fear [9] - Sutton categorizes the evolution of the universe into four eras, asserting that humanity is transitioning from the third to the fourth era, with the mission to design systems capable of design [9] Group 9: Sergey Levine's Perspective on AI Learning - Professor Sergey Levine from UC Berkeley posits that large language models may merely be observers in a "Plato's cave," learning indirectly from human thought through internet text [10] - He questions why language models can learn rich knowledge from predicting the next token, while video models learn less despite containing more physical world information [10] - This perspective suggests that current AI systems may only mimic human thought rather than truly understanding the world, indicating a need for AI to learn from physical experiences [10]