闻乐 发自 凹非寺 量子位 | 公众号 QbitAI 不是拼凑知识点，AI这次是真搞研究。 一个叫 Virtuous Machines 的AI系统，花了17小时、114美元，找了288个真人做实验，写了一篇30页的学术论文。 而且还是 从选题到成稿 全自动化速通！？ 来看看这个AI都写了点啥。 AI自动化做科研：从灵光一现到可发表论文 AI自主完成的这个论文属于 认知心理学领域 ，具体聚焦于 人类视觉认知 相关的研究方向。 而且它可不是瞎写，而是靠人类的科研套路来。 先是基于认知心理学理论提出研究问题，比如"视觉工作记忆与心理旋转能力有没有关系"、"心理意象清晰度对视觉认知任务表现有什么影 响"等。 （视觉工作记忆是指人类维持并处理视觉信息的能力，涉及信息存储、操作和提取过程；心理旋转是指通过心理操作实现空间客体旋 转以完成知觉匹配的认知过程） 像人类一样搞科研 接着设计实验方案，考虑到了样本量计算、控制变量，还用VVIQ2量表测量被试（对象）的心理意象清晰度； 在确定好实验方案后，它还通过在线平台Prolific招募了288名被试对象，等277份有效数据（部分被试未完成实验，被AI筛除了）收上来，它 又连续 ...

Core Insights - The article highlights how AI, specifically GPT-5, has significantly aided mathematician Terence Tao in solving complex mathematical problems, reducing the time and effort required for manual calculations and coding [1][2][3]. Group 1: AI's Role in Mathematics - Terence Tao expressed that without AI assistance, completing similar tasks would take several hours, primarily due to manual coding and debugging [1]. - Tao utilized GPT-5 to tackle a problem on MathOverflow regarding the relationship between the least common multiple sequence and highly abundant numbers, which required extensive numerical searches [7][10]. - The AI's ability to assist in this mathematical inquiry marks a new era of collaboration between humans and machines in exploring complex problems [5][29]. Group 2: Problem-Solving Process - Initially, Tao attempted to have GPT-5 generate a Python program to search for counterexample parameters but faced issues with long execution times and improper initial parameters [19][20]. - He then shifted to a step-by-step dialogue with GPT-5, breaking down the larger problem into smaller, manageable parts, which ultimately led to the successful generation of the required parameters [21][22]. - The final solution involved a concise 29-line Python script generated by GPT-5, which Tao used for independent verification, confirming the results aligned with his heuristic predictions [23][24]. Group 3: Broader Implications of AI in Research - This instance is not the first time Tao has employed AI for mathematical problem-solving; he has previously used AI for various projects, demonstrating its potential as a mediator in mathematical proofs [27][28]. - The article suggests that while AI may not achieve accolades like the Fields Medal in the short term, it can significantly enhance the efficiency and effectiveness of mathematical research [28][29].

Core Viewpoint - OpenAI has responded strongly to the lawsuit filed by xAI, denying all allegations of corporate espionage and asserting that the lawsuit is an attempt to intimidate its employees [2][3][10]. Group 1: Allegations by xAI - xAI has made three main allegations against OpenAI: violation of federal trade secret laws, intentional interference with xAI's economic relationships with its employees, and violation of California's unfair competition laws [11]. - Specific incidents cited include the alleged theft of proprietary information by former xAI engineers Xuechen Li and Jimmy Fraiture, who are accused of transferring sensitive data to OpenAI [12][14][15]. - xAI also claims that a former senior finance executive left without signing a confidentiality agreement and took critical strategic information to OpenAI [19][20]. Group 2: OpenAI's Defense - OpenAI has categorically denied the allegations, stating that Xuechen Li never officially joined the company and did not transfer any proprietary information [27][29]. - Regarding Jimmy Fraiture, OpenAI asserts that any actions taken during his "garden leave" were personal and not directed by OpenAI, and that no confidential information was received [31][32]. - OpenAI emphasizes that the unnamed finance executive's departure was unrelated to any alleged poaching and was due to refusing to engage in improper financial practices at xAI [33][34]. Group 3: Legal Proceedings - OpenAI has filed a motion to dismiss xAI's lawsuit, arguing that the claims lack merit and that the inclusion of names of former employees not accused of wrongdoing is an act of intimidation [37]. - A hearing for this motion is scheduled for November 18, 2025, which will address procedural matters rather than the substantive issues of the case [38].

Core Insights - Nano Banana has officially opened its API, allowing developers to integrate it into their products and enabling large-scale content production for enterprises [9][10] - The API pricing is set at approximately $0.039 per image output, translating to about 0.28 yuan, with a cost of $30 for every 1 million image output tokens [2][15][16] - Google has introduced two new features: customizable aspect ratios and a pure image generation mode, enhancing its utility for content creators [3][8] Pricing and Cost Structure - Each image generated costs about $0.039 (approximately 0.28 yuan), with the maximum image size being 1024x1024 pixels, consuming around 1290 tokens [16] - The pricing for image generation is 12 times higher than the Gemini 2.5 Flash text mode [17] New Features - The first new feature allows users to customize aspect ratios, offering over ten options including 16:9, 9:16, 4:3, and 3:2, catering to various visual content needs [4][18] - The second feature supports pure image output mode, which returns only images without additional text, saving tokens and reducing contextual interference, ideal for real-time previews and e-commerce displays [7][8] Application and Usability - Users can create their own applications directly in Google AI Studio by inputting prompts, making it accessible for non-developers [13][14] - The new features are designed to meet the practical needs of content creators, positioning Nano Banana as a more practical tool [8]

这主要源于现有模型在物体属性理解上的不足，以及细粒度感知能力的局限。 为缓解上述问题，华中科技大学团队和金山办公团队联合提出了两个核心模块： 语义增强特征提取器 （SEFE） 和 交错局部视觉耦合 （ILVC） 。 前者融合语义特征与像素级特征，提升物体属性推理能力，从而获得更精确的分割结果。 后者基于分割掩码提取局部特征后，自回归生成局部描述，为模型提供细粒度监督，从而有效减少理解幻觉。 最终，研究团队 构建了在分割和理解两项任务上均取得SOTA的多模态大模型LIRA 。 LIRA团队 投稿 量子位 | 公众号 QbitAI 多模态大模型需要干的活，已经从最初的文生图，扩展到了像素级任务 （图像分割） 。 不过，无论是OMG-LLaVA，还是提出了embedding-as-mask范式的LISA （CVPR 2024） ，都还存在分割结果不够精确，以及理解过程中 出现幻觉两大痛点。 与InternVL2相比，LIRA在保持理解性能的同时，额外支持图像分割任务；与OMG-LLaVA相比，LIRA在图像分割任务上平均提升8.5%，在 MMBench上提升33.2%。 目前，LIRA项目已被ICCV 2025录用 ...

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 2025 人工智能年度潜力创业公司 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 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 ： 聚焦于中国人 ...

衡宇 发自 麦蒿寺 量子位 | 公众号 QbitAI 我惊！ 图灵奖得主、AI三巨头之一的LeCun在Meta待得是如坐针毡。 对于Meta来说，这是一次"对齐公司战略"的制度微调；但 对LeCun及FAIR团队而言，这几乎是对学术自由、研究自主性的正面挑衅 。 而且 这不是一次孤立的骚操作 。 准确来说，最近几个月，Meta内部一系列混乱、缺乏透明度的AI战略重组，都让LeCun非常不满。 Yann LeCun已经直接跟同事表示，自己可能会辞去FAIR首席科学家的职务 。 LeCun可是FAIR的联合创始人之一，这么多年一直驻扎FAIR，起着学术研究和前瞻洞悉的引领作用。 知情人士透露，这不是LeCun的一时冲动，而是LeCun对Meta近几个月在AI部门组织调整等骚操作不满，长久积压，情绪终于爆发了。 △ 图片由ChatGPT生成 一个月前Meta内部开始施行的策略，大概是压死骆驼的最后一根稻草： 即日起，FAIR若是要对外发表论文，必须先经过TBD实验室的额外审核。 如果审核发现论文价值大，论文就先不往外发，论文作者还得帮助论文成果在Meta产品中落地，才能继续自己的日常研究工作。 过去几个月，FAIR ...

CoRL 2025 投稿 量子位 | 公众号 QbitAI 手把手教机器人，直接就能让它学到真本事！ 不管是采茶沏茶做早餐，这些精细活儿都「手」到擒来。还是 灵巧手 的那种。 △ 星动纪元灵巧手星动XHAND 1 △ 星动纪元灵巧手星动XHAND 1 来自斯坦福大学、哥伦比亚大学、摩根大通AI研究院、卡耐基梅隆大学、英伟达提出了一种数据采集与策略学习框架 DexUMI —— 利用人手作为自然接口将灵巧操作技能迁移至多种灵巧手 。该框架通过硬件与软件的双重适配，最大限度缩小人手与各类灵巧手之间的具身 差异。 相信不少人看过去年他们发布的UMI，通过记录并学习人类操作，让夹爪类机器人学会洗碗，结果在行业引发不小的关注和轰动。 除了效果惊艳，更深层的原因在于他们让夹爪的操作数据采集迅速便利化，行业多家厂商迅速跟进，推出了工业化数采产品。 今年，他们将夹爪升级到更复杂更高自由度的灵巧手，让机器人学会更丰富更精细的操作任务，势必也将引发新一轮灵巧手数采革命。 正在首尔举办的CoRL 2025中，它被提名为最佳论文。 值得一提的是，其中全驱采用的灵巧手正是出自国产明星具身玩家 星动纪元 的星动XHAND 1。 手把手教机 ...

闻乐 发自 凹非寺 量子位 | 公众号 QbitAI 让它模拟"给ChatGPT发信息"，它不仅生成了画面，还来了一段有问有答的"交互"。 先是编了一个问题：Write a playful haiku about a cat staring out the window.（写一首关于猫凝视窗外的俏皮俳句。） Sora2太卷了。 居然能预测ChatGPT的输出、渲染HTML？！ 然后又以ChatGPT回答的模式给出了音频回应：Whiskers pressed to glass. Birds gossip beyond the pain. Tail flicks. Daydreams fly. （中文大意是："胡须紧贴玻璃。鸟儿在窗外叽喳。尾巴轻摇。白日梦飞扬。） 全程以ChatGPT的机械女声回答，并且俳句音节还卡得严丝合缝。 这段 视频场景+LLM推理 的实测效果让一众网友惊叹，甚至有人说"Sora2模糊了视频生成和交互式AI的边界"。 而这段代码在真实浏览器中渲染的样子be like： 实际上不仅是像这样能预测ChatGPT的推理回答，Sora2还能渲染HTML。 通过了玻璃折射测试 还有人让Sora2渲染 ...

Core Insights - Thinking Machines Lab has launched its first product, Tinker, which simplifies model fine-tuning to the level of modifying Python code [1][12] - The company has moved past the "zero product, zero revenue" valuation of $84 billion [2] Product Overview - Tinker is a flexible API designed for fine-tuning language models, allowing researchers to control algorithms and data without managing infrastructure [12][13] - The initial support for Tinker includes Qwen3 and Llama3 series models, enabling easy switching between small and large models with a simple string modification in Python code [15] - Tinker’s API automates low-level training steps while handling scheduling, scaling, and error recovery [17] Technical Features - Tinker utilizes LoRA to allow multiple training tasks to share the same GPU, reducing costs and enabling more parallel experiments [22] - The gradient update strategy for Tinker is defined as: New parameters = Original parameters + Learning rate × Advantage value × Gradient of log probability [28] Industry Reception - Tinker has garnered significant attention in the industry, with beta testers noting its excellent balance between abstraction and tunability compared to other fine-tuning tools [30] - Research teams from prestigious institutions have already achieved notable results using Tinker [30] Strategic Vision - Thinking Machines Lab aims to reinvent a version of OpenAI that emphasizes open research sharing and greater freedom for researchers [10][11] - The company’s mission aligns with making cutting-edge models more accessible for customization based on individual needs [14]