Core Insights - The article discusses a new research paper by Yann LeCun's team that reveals the hidden capability of the self-supervised model JEPAs (Joint Embedding Predictive Architecture) to learn data "density" [2][5][6] - This finding challenges the long-held belief that JEPAs only excel at feature extraction and are unrelated to data density [7] Group 1: Key Findings - JEPAs can autonomously learn the commonality of data samples during training, allowing them to assess the typicality of a sample without additional modifications [6][11] - The core discovery is that the anti-collapse mechanism enables precise learning of data density, which was previously underestimated [11][12] - The research highlights that when JEPAs output Gaussian embeddings, they must perceive data density through the Jacobian matrix, making the learning of data density an inherent result of the training process [11] Group 2: Practical Applications - The team introduced a key tool called JEPA-SCORE, which quantifies data density and scores the commonality of samples [14][15] - JEPA-SCORE is versatile and can be applied across various datasets and JEPAs architectures without requiring additional training [16][17] - Experiments demonstrated that JEPA-SCORE effectively identifies typical and rare samples across different datasets, confirming its reliability and general applicability [18] Group 3: Research Team - The research was a collaborative effort involving four core researchers from Meta's FAIR, including Randall Balestriero, Nicolas Ballas, and Michael Rabbat, each with significant backgrounds in AI and deep learning [26][28][30][32][34][36]

Core Insights - Ant Group has officially released its flagship model, Ling-1T, which boasts one trillion parameters, surpassing both open-source models like DeepSeek-V3.1-Terminus and closed-source models such as GPT-5-main [1][56] - Ling-1T demonstrates state-of-the-art (SOTA) performance in various complex reasoning benchmarks, including code generation and mathematical reasoning [1][3] - The model exhibits impressive reasoning speed, initiating thought processes almost instantaneously upon input [4][60] Performance and Capabilities - Ling-1T achieved optimal performance on the AIME 25 competition mathematics leaderboard, outperforming numerous models [3] - The model can efficiently handle complex logical deductions and generate lengthy texts with smooth output [4][60] - In practical tests, Ling-1T effectively solved a spatial geometry optimization problem by proposing four distinct solutions, each with detailed steps and applicable scenarios [8][9] Technical Innovations - The model's architecture is based on Ling 2.0, with a total parameter count expanded to one trillion, allowing for enhanced information storage and expression [38][41] - The training process involved over 20 trillion tokens of high-quality, reasoning-focused data, supporting a maximum context window of 128K tokens [39][40] - A novel "mid-training + post-training" approach was employed, enhancing the model's reasoning capabilities and efficiency [40][59] Training Methodology - The training was divided into three phases: initial knowledge acquisition, reasoning skill development, and mid-training to prepare for post-training [45][44] - A new learning rate strategy, WSM (Warmup-Stable and Merge), was introduced to optimize training without traditional decay, resulting in improved performance across tasks [49][48] - The LPO (Linguistics-Unit Policy Optimization) method was innovatively applied, allowing for more precise training by using sentences as the optimization unit [52][54] Market Context - The release of Ling-1T positions Ant Group among the leading players in the trillion-parameter open-source model space, alongside Qwen and Kimi [61] - The ongoing trend of rapid advancements in China's open-source model landscape is highlighted, with multiple significant releases from various companies [62][56] - The competitive landscape suggests that further innovations and surprises in the large model sector are likely to emerge from China [63]

△ 对比DeepScientist与人类专家的研究进展 在AI文本检测任务中，DeepScientist仅用两周时间就实施和验证了超过 1000种 不同的假设，在此期间取得了相当于人类三年的进展。 在RAID数据集测试中，DeepScientist设计的方法实现了 7.9% 的AUROC提升，成功 超越了人类现有SOTA方案 。 另外DeepScientist还在智能体失败归因、LLM推理加速等任务上也分别达成了新的SOTA。 DeepScientist团队 投稿 量子位 | 公众号 QbitAI 人类科学家三年的工作量，如今AI两周就能轻松搞定！ 最近，来自西湖大学的自然语言处理实验室发布了 DeepScientist 系统，这也是 首个 具有完整科研能力，且在无人工干预下，展现出目标 导向、持续迭代、渐进式超越人类研究者最先进研究成果的AI科学家系统。 下面是更多详细内容介绍。 从"科研助理"到"首席科学家"：AI科研模式的变革 过去的AI Scientist系统，如果不给定一个清晰明了的科研目标，就很容易陷入对现有知识的机械组合与无效试探的窠臼中，最终形成的科研 产出在人类专家看来缺乏焦点，科学价值不高 ...

允中 发自 凹非寺 量子位 | 公众号 QbitAI 光轮智能 × NVIDIA 重磅直播即将开启！ 双方将携手揭秘如何利用SimReady与AI打通Sim2Real （仿真到现实） 。 直播核心看点 Sim2Real技术突破 深度解析双方如何基于SimReady与AI，实现从虚拟仿真到物理世界的无缝迁移，攻克机器人开发落地中的关键挑战。 合作进展独家披露 光轮智能与NVIDIA在技术研发、场景应用等方面的最新合作成果与规划。 大咖实战视角 两位专家将结合实践经验，分享机器人、AI领域的技术趋势与商业化路径。 主讲嘉宾 直播时间 对机器人及AI领域感兴趣的朋友，欢迎扫码预约，锁定直播席位！ *本文系量子位获授权刊载，观点仅为原作者所有。 一键三连 「点赞」「转发」「小心心」 欢迎在评论区留下你的想法！ 更多详情，请戳文末「阅读原文」。 Steve Xie，光轮智能创始人兼CEO Madison Huang，NVIDIA产品营销高级总监 北京时间 ： 10月9日 凌晨0:00 太平洋时间 ： 10月8日 上午9:00 点亮星标 — 完 — 科技前沿进展每日见 ...

Jay 发自 凹非寺 量子位 | 公众号 QbitAI 什么AI应用公司和方向是OpenAI看好的？ 这不，OpenAI公布了30家Tokens消耗破万亿的"大金主"。 | Number | Name | Company | Role | Number | Name | Company | Role | | --- | --- | --- | --- | --- | --- | --- | --- | | 1 | Isaac Andersen | Duolingo | Senior SWE | 16 | Praty Sharma | HubSpot / Dashworks | Al / CoFounder | | 2 | Alex Atallah | OpenRouter | CEO and CoFounder | 17 | Denis Shiryayev | JetBrains | Group Product Manager | | 3 | Chris Colon | Indeed | Director, Al Platforms | 18 | Sam Spelsberg | Delphi | Co-fou ...

Core Insights - The article discusses the recent transition of Shunyu Yao, a prominent AI researcher, from Anthropic to Google DeepMind, highlighting his background and motivations for the move [1][4][41]. Group 1: Background and Career Transition - Shunyu Yao, a distinguished alumnus of Tsinghua University, recently joined Google DeepMind as a Senior Research Scientist after leaving Anthropic, where he contributed to the Claude AI model [1][41]. - Yao's departure from Anthropic was influenced by a fundamental disagreement in values, which he stated accounted for 40% of his decision, while the remaining 60% involved internal details he chose not to disclose [21][24]. - His experience at Anthropic was marked by a high workload, which he described as "super busy," preventing him from reflecting on his transition from physics to AI research until after his departure [7][8][18]. Group 2: Insights on AI Research - Yao expressed that the field of AI research, particularly in large models, is currently in a chaotic state, akin to the early days of thermodynamics, where foundational principles are not yet fully understood [14][15][16]. - He noted the rapid evolution of AI, with the Claude model progressing from version 3.7 to 4.5 within a year, emphasizing the fast-paced nature of advancements in the field [27]. - Yao's background in theoretical physics provided him with a unique perspective on AI research, allowing him to appreciate the ability to identify patterns without fully understanding the underlying principles [16][18]. Group 3: Academic Achievements - During his undergraduate studies, Yao made significant contributions to condensed matter physics, publishing groundbreaking work in the prestigious journal Physical Review Letters [30][31]. - His research achievements include the introduction of new physical concepts and theories related to non-Hermitian systems, which have been recognized as substantial contributions to the field [32][33]. - After completing his PhD at Stanford University, Yao's work continued to focus on cutting-edge topics in quantum mechanics, further establishing his reputation as a leading researcher [35].

组委会 发自 凹非寺 量子位｜公众号 QbitAI 为了让更多从业者感受智能浪潮的跃迁，也为了给予更多同行同路人掌声与鼓舞，我们将正式启动 「2025人工智能年度榜单」评选报名 。 这是量子位人工智能年度榜单的 第8年 。八年来，我们见证了技术的突破与落地，产业的融合与重塑，也见证了一批又一批推动时代前行的 企业、人物与产品。 在人工智能重新定义一切的时代里，智能技术已不再是单一工具，而是产业与社会协同进化的驱动力。我们期待通过这场年度评选，去发现并 致敬那些真正引领变革、开拓边界的探索者与实践者。 产品榜 人物榜 2025 人工智能年度 焦点人物 详细评选标准及报名方式如下。 2025 人工智能年度领航企业 本次评选将从 企业 、 产品 、 人物 三大维度，设立五类奖项。欢迎企业踊跃报名！ 让我们共同见证年度之星，点亮未来的方向。 企业榜 2025 人工智能年度 领航企业 2025 人工智能年度 潜力创业公司 2025 人工智能年度 杰出产品 2025 人工智能年度 杰出解决方案 将面向中国人工智能领域，评选出最具综合实力的企业， 参选条件 ： 评选标准 ： 2025 人工智能年度潜力创业公司 聚焦于中国人 ...

西风 闻乐 发自 凹非寺 量子位 | 公众号 QbitAI 刚刚，诺贝尔物理学奖揭晓！ 今年颁给了量子力学领域的三位科学家 Joh n Clarke 、Michel H. Devoret和John M. Martinis ，以表彰他们： 在电路中发现宏观量子力学隧穿效应和能量量子化现象。 其中John M. Martinis曾是谷歌AI量子实验室的量子硬件首席科学家，与团队在《Nature》曾发表划时代论文，首次通过一台拥有53个量 子比特的处理器实现了"量子霸权"。 John Clarke John Clarke的研究方向主要涉及超导性和超导电子学，特别是低温物理和超导电子学领域。 他最为人知的贡献是 发明和 改进了 超导量子干涉仪 （SQUID） ，这是一种极其灵敏的磁通量-电压转换器，可应用于凝聚态物理、地球 物理学、天体物理学、宇宙学、医学物理等领域，被誉为" 磁学领域的游标卡尺 "。 John Clarke1942年出生于英国剑桥，1964年、1968年分别获得剑桥大学基督学院和达尔文学院的物理学学士、硕士和博士学位，2003年 获得剑桥大学理学博士学位。 1968年，他以博士后身份进入加州大学伯克 ...

西风 发自 凹非寺 量子位 | 公众号 QbitAI OpenAI开发者日2025，新品发布密度远超往年。 奥特曼带着一系列最新内容来了—— 现在，大伙儿可以 在ChatG PT中和各种" App"对话 ，只需 在提示词中呼唤应 用 名 即可触发调用。 在日常对话中，ChatGPT也会根据用户的需求，主动推荐用户使用某款App： 另 外，开发者构建Agent有新工具了—— AgentKit 。 OpenAI研究员现场演示，在 8分钟内 用AgentKit为其开发者日官方网站构建Agent。 AgentKit是一个Agent工具包，包含诸多模块，可以帮助开发者更高效设计工作流。 例如通过Agent Builder模块，开发者无需从代码开始，只需添加并设置一系列节点即可用 可视化直观的方式快速构建Agent ： 这还没完，OpenAI还宣布 AI编程神器Codex正式发布 ，并且推出了三项新功能。 OpenAI研究员Steven Heidel在X上透露，最新发布的Agent Builder不到6周就完成了端到端构建，而 Codex编写了80% 的P R 。 比如直接在ChatGPT 5的对话框中输入"course ...

组委会 发自 凹非寺 量子位｜公众号 QbitAI 为了让更多从业者感受智能浪潮的跃迁，也为了给予更多同行同路人掌声与鼓舞，我们将正式启动 「2025人工智能年度榜单」评选报名 。 这是量子位人工智能年度榜单的 第8年 。八年来，我们见证了技术的突破与落地，产业的融合与重塑，也见证了一批又一批推动时代前行 的企业、人物与产品。 在人工智能重新定义一切的时代里，智能技术已不再是单一工具，而是产业与社会协同进化的驱动力。我们期待通过这场年度评选，去发现 并致敬那些真正引领变革、开拓边界的探索者与实践者。 本次评选将从 企业 、 产品 、 人物 三大维度，设立五类奖项。欢迎企业踊跃报名！ 让我们共同见证年度之星，点亮未来的方向。 企业榜 产品榜 人物榜 将面向中国人工智能领域，评选出最具综合实力的企业， 参选条件 ： 评选标准 ： 2025 人工智能年度潜力创业公司 聚焦于中国人工智能领域创新创业力量，将评选出最具投资价值和发展潜力的AI创业公司， 参选条件 ： 评选标准 ： 2025 人工智能年度 焦点人物 详细评选标准及报名方式如下。 2025 人工智能年度领航企业 2025 人工智能年度 领航企业 2025 ...