梦晨 发自 凹非寺 量子位 | 公众号 QbitAI 点赞者就2000多，用实际行动退订的也不少。 退订者中有最高价 20倍Max套餐 的重度用户。 原本在开发者社区口碑甚好，甚至 Claude Code单产品年化收入估算达到5亿美元 的Anthropic，到底因何犯了众怒？ 工程师Ahmad Osman细数几大罪状： 就这甚至还没列完，可想而知这位开发者有多愤怒了。 Claude出现大危机，不是因为最近的某些骚操作，而是 产品本身就出了问题 。 已经有AI工程师带头呼吁大家退订（这里PoS指Piece of Shit，也就是一坨 ）。 评论区有人补充，最糟糕的是模型悄悄变差，而你白白浪费了一小时才能意识到，没有哪个专业的开发环境是不能固定版本的。 好，现在骂也骂了退也退了，活还是得干，总不能退回到古法手工写代码吧。 那么以后用啥？ 有很多人集中转投去了隔壁 OpenAI Codex ，甚至惊动了奥特曼本曼。 OpenAI Codex强势崛起 如果你在前几天打开美国贴吧Reddit的Claude Code吧，就会发现怎么全是讨论OpenAI Codex的，都要怀疑是不是走错门了。 在白天高峰时段，用到的是缩水 ...

按库克的说法，这波新品一切都以设计为核心。 效果上看，iPhone系列的镜头模组也确实基本告别了过去的"浴霸"模式。 克雷西 鱼羊 发自 凹非寺 量子位 | 公众号 QbitAI 标准版iPhone终于也用上高刷了！ 刚刚结束的苹果春晚上，iPhone、AirPods Pro和Apple Watch相继登台亮相。 | ... | 0000 | 00000 | | --- | --- | --- | | 新款 | 新款 | 新款 | | iPhone 17 Pro | iPhone Air | iPhone 17 | | 创新设计,打造巅峰性能和超 | 迄今最薄 iPhone，身藏高 | 拉高好感度,再添耐用性。 | | 长续航。 | 能内核。 | | 当然最令果粉激动的，还是这次iPhone全系都安排了高刷。 确实，型号基础，刷新率就不基础。 除了iPhone，耳机和手表也迎来重要升级，牙膏直接挤爆。 比如AirPods Pro也变身智能穿戴，不仅能够进行同声传译，还支持心率检测。 还有Apple Watch也支持了5G通信，还新增了重磅健康功能。 不得不说这一波库克的刀法是真的变温柔了 （希望英伟达的老黄也 ...

Core Viewpoint - The article discusses Ashish Vaswani's shift towards open-source AI research, criticizing closed-source companies for hindering scientific exploration and innovation in the field of artificial intelligence [2][28]. Group 1: Background and Motivation - Ashish Vaswani, co-creator of the Transformer model, believes that companies like OpenAI are too focused on commercialization, neglecting foundational research [2][4]. - After facing challenges with the Scaling Law and internal conflicts at Adept AI, Vaswani decided to establish Essential AI to pursue open-source research [10][11][14]. - The shift to focus entirely on foundational research was supported by the board and investors, indicating a recognition of the need for more open alternatives in AI [19][20]. Group 2: Vision for Open Source - Vaswani aims to create a platform similar to DeepSeek in the Western world, emphasizing the importance of open-source AI in education and healthcare [6][26]. - He argues that AI should not only serve commercial interests but also benefit the general public, allowing access to advanced technologies in underserved areas [28][29]. - Essential AI's recent research suggests that breakthroughs in large language models can be achieved during the pre-training phase, potentially reducing training costs significantly [31][33]. Group 3: Challenges with Closed Source - The article highlights the "innovator's dilemma" faced by AI companies, where the focus on commercialization can stifle innovation and long-term research [39][42]. - Many AI firms have shifted resources away from research to prioritize immediate profitability, especially in a challenging market environment [36][38]. - Vaswani expresses concern that the best-performing models from closed-source companies may actually impede progress in AI development [35]. Group 4: Funding and Sustainability - Vaswani proposes a cross-subsidization model, where revenue from certain services can support open-source initiatives, ensuring the sustainability of Essential AI [53][57]. - He believes that open-source models can be more profitable in the long run, as they foster a collaborative ecosystem that drives innovation [61].

Core Viewpoint - The article discusses the limitations of AI models in understanding distorted text, highlighting that while humans can easily comprehend such text, AI struggles significantly, indicating a fundamental gap in AI's text recognition capabilities [2][23]. Group 1: AI Performance in Text Recognition - A research team from various institutions conducted experiments showing that leading AI models like OpenAI's GPT-5 and Google's Gemini perform poorly when faced with distorted Chinese characters and English words [2][4]. - In tests involving 100 four-character Chinese idioms, AI models achieved a maximum average matching rate of only 12.1%, while humans scored 100% [7]. - For 100 eight-letter English words, AI models also failed to provide correct answers, demonstrating a consistent inability to recognize and interpret distorted text [10][20]. Group 2: Human vs. AI Understanding - Humans can easily read distorted text due to their advanced visual processing and understanding of language structure, while AI relies on pattern matching without grasping the underlying text structure [9][24]. - The inability of AI to separate and combine symbols leads to failures in recognizing even slightly altered text, which humans can still understand [26][28]. - The research emphasizes that human reading comprehension is a multi-modal process, relying on various sensory inputs and reasoning abilities, unlike AI's current capabilities [29]. Group 3: Implications for AI Development - The findings suggest that to enhance AI's resilience in text recognition, there is a need to rethink how vision-language models (VLMs) integrate visual and textual information, potentially requiring new training data and structural insights [28]. - The results also highlight the challenges AI faces in practical applications, such as education and security, where non-standard text recognition is crucial [30].

西风 发自 凹非寺 量子位 | 公众号 QbitAI 刚刚，百度深度思考模型升级上线了！ 升级后的文心 大模型X1 .1 ，在 事实性、指令遵循、智能体 等能力上均有显著提升。 官方展示了其在智能客服场景复杂长程任务中的应用，在System Prompt中输入用户的问题后，文心X1.1借助模型本身智能体能力，即可自 动拆分复杂任务，调用不同工具逐步规划执行，且严格遵循服务流程和业务规则。 再用它编写python脚本，让25个彩色粒子在真空圆柱形容器里弹跳、留轨迹，还要带容器旋转和场景缩放。 效果丝滑，粒子全程守规矩没出界： 用HTML动 画整活归并排序，排序过程动态可视化，算法步骤一目了然： 最新开源思考模型ERNIE-4.5-21B-A3B-Thinking 发布，该模型在ERNIE-4.5-21B-A3B基础上训练而来，在内容创作、逻辑推理、数学计 算、代码生成与工具调用等多个任务中表现卓越。 此外，百度发布了 ERNIEKit文心大模型开发套件 ，提供更加便捷的模型后训练方案，仅 需 4张GPU即可对ERNIE-4.5-300B-A47B模型进 行高效调优 ，进一步降低开发者将模型 落地到实际应用的门槛 ...

不圆 发自 凹非寺 量子位 | 公众号 QbitAI 最近AI生图赛道简直卷疯了！ 从Nano Banana的爆火，到即梦AI 4.0，豆包4.0接连上线，一直专注于视频大模型的Vidu也按捺不住了： Vidu Q1 参考生图堂堂登场！同时支持 7张 参考。 主体一致性比起谷歌Nano Banana也毫不逊色。 （Nano Banana最多支持3张参考图） 量子位抢先实测了这款模型，它的表现相当不错——能够自由引用的7张参考图，带来了极高的可操作性。 用简单的自然语言描述即可。 或者是直接生成时尚大片，现场拍摄啥的都省了。 我们探索了很多有趣的玩法，提示词、图片都放在下面了，一起来看一下！ 7张参考图，能怎么玩？ 我们实测了几种玩法，比如让各种违和的元素凑成一张和谐的画面、或者是制作时尚大片…… 可以说，只要有创意，万物皆可合成。 万物皆可合成 无论是让秦始皇骑北极熊在上海喝柠檬水： 还是让李白坐火箭成功登月： 参考图一放，就看Vidu Q1的结果是否符合想象。 潮流单品秒变OOTD 既然有那么多参考，岂不是可以直接全套换装？ 所有单品一键上身，是时候展现搭配之力了（摩拳擦掌）。 用这套提示词，不管是地中海还 ...

Core Viewpoint - The article discusses the innovative RoboBallet project, which combines Graph Neural Networks (GNN) with Reinforcement Learning to enhance multi-robot collaboration in complex environments, showcasing significant advancements in robotic motion planning and task allocation [5][9][24]. Summary by Sections Introduction - RoboBallet is a collaborative robotic system that allows multiple robotic arms to work together efficiently in a shared space without collisions [1][2]. Technical Innovation - The project utilizes GNNs for strategy networks and state-action value estimation, enabling the control of up to 8 robotic arms and managing 56 degrees of freedom [6][9]. - It addresses three complex sub-problems: task allocation, task scheduling, and motion planning, which are traditionally challenging for existing algorithms [10][12]. Methodology - The environment is modeled as a graph structure, where nodes represent robots, tasks, and obstacles, and edges denote relationships among them [11][14]. - The GNN processes dynamic graph sizes and generates joint velocity commands for the robotic arms based on observed states [14][15]. Performance Metrics - RoboBallet was tested in a simulated environment with 4 to 8 robots, 40 tasks, and 30 obstacles, demonstrating superior performance compared to traditional methods [18][19]. - The planning speed is remarkable, with each planning step taking approximately 0.3 milliseconds on an NVIDIA A100 GPU, achieving over 300 times real-time planning speed [21]. - The average execution time decreased by about 60% as the number of robots increased from 4 to 8 [22]. Generalization and Applications - The model exhibits zero-shot transfer capabilities, allowing it to adapt to new environments without additional training [24]. - RoboBallet's efficiency can optimize work unit layouts, reduce task execution time by 33%, and enhance fault-tolerant planning [24].

不动嘴不动手，只靠意念就能对手机发号施令？ 请看VCR： 两个人语言不通，现在也可通过意念说话，然后直接翻译成对方的语言，并通过骨传导耳机输出。 克雷西 发自 凹非寺 量子位 | 公众号 QbitAI MIT初创团队推出了一款非侵入式穿戴设备，能够让人类实现用意念"说话"。 这个穿戴可以让人类随时随地实现无动作书写、创作、交流，甚至帮助有特殊言语障碍的人恢复声音。 研发团队表示，打造这款穿戴设备是为了延伸人类思维，让每个人都能轻松探索自己的世界。 智能穿戴学会"读心术" 这款穿戴名字叫AlterEgo，来自于拉丁语，意思是"另一个自我"。 AlterEgo是一种可穿戴的静默语音交互平台，允许用户在没有声音或明显动作的情况下与计算设备进行双向交互。 这意味着用户可以像自言自语一样在心里"说话"，而系统能够理解并处理这些"静默"输入，词汇准确率可达92%。 它还能通过骨传导耳机将反馈提供给用户，通过不干扰外部环境的方式直接传输到用户的耳朵里，提供完整的输入-输出交互体验。 AlterEgo支持用户通过静默语音控制各种应用程序。例如，用户可以通过内心默念数学计算式，设备会进行计算并反馈结果。此外，用户也可 以设置提 ...

Core Viewpoint - The use of AI in both resume writing and screening has not improved job search efficiency, leading to a cycle where applicants and HR are increasingly frustrated with the process [20][41]. Group 1: Job Market Dynamics - Many job seekers are using AI tools to create resumes, yet this has not translated into better job opportunities [20]. - Companies are also employing AI to filter resumes, resulting in a high volume of applications being screened out without human review [28][41]. - The job market has become a closed loop where both applicants and HR are caught in a cycle of dissatisfaction [29][30]. Group 2: Applicant Experiences - A case study of a recent graduate, Harris, illustrates the struggle; he applied for 200 positions and received no offers, with many applications going unanswered [13][14]. - Another example is Martine, a lawyer's assistant with 10 years of experience, who faced similar challenges despite multiple interviews [17][18]. - Applicants report various reasons for rejection, including lack of qualifications, keyword mismatches, and even being overqualified [4][38]. Group 3: HR Perspectives - HR professionals express frustration over the influx of similar AI-generated resumes, which they find unoriginal and lacking personal touch [33][35]. - There is a sentiment among HR that AI-generated applications do not reflect genuine effort, making them easier to dismiss [35]. - The use of AI in the hiring process has led to a disconnect between applicants and employers, complicating the recruitment landscape [41].

梦晨 发自 凹非寺 量子位 | 公众号 QbitAI 奥特曼点名表扬了两个波兰人。 没有他们，OpenAI就不是今天的样子。 他们在OpenAI的贡献从Dota项目大规模扩展了强化学习，到领导了GPT-4的预训练，还与 Ilya和Lukasz共同推动了导致推理突破的最初想 法。 当然，奥特曼对他们如此高评价或许还有另一个原因： 在2023年OpenAI内乱事件中， 他俩也是带头站出来宣布辞职 ，要追随奥特曼离开的。 从高中同窗到OpenAI重聚 故事还要从波兰的一所学校说起， 格丁尼亚第三高中。 他们是OpenAI首席科学家 Jakub Pachocki 以及头衔为"Technical Fellow"的 Szymon Sidor 。 △ 左：Jakub Pachocki，右：Szymon Sidor 两人不仅是 波兰老乡 ，而且是 高中同学 ，读博时分别选择了计算机科学和机器人，后来又在OpenAI重聚。 在ChatGPT风靡全球、每天服务数亿用户的今天，奥特曼感慨大多数人永远不会想到背后那些付出心血的人，这两位波兰科学家，正是其中 的关键角色。 先说Pachocki（以下简称 帕哥 ）这边。 15岁的时候， ...