GPT-5.2也发布了有几天了。 在发布之后，大家又吵的满头开花，不过实测完之后，确实有一点失望。 不过，这篇文章不是来撩GPT-5.2的实测的。 而是聊GPT-5.2发布之前的一个很有意思的神奇的事情，就是，在模型发布之前，总能有人，提前预测市场。 很多关注AI资讯的朋友可能都知道，GPT-5.2这玩意，从9号就开始各种小道消息，然后又传13号凌晨2点发。 然后又有一群人说不发。 结果，在一片吵闹中，最后，OpenAI还是凌晨2点把这玩意掏出来了。 不过，一点也不意外。 因为我当时几乎百分百肯定GPT-5.2肯定会发出来的。 因为在各种各样的消息里，有一个平台却显得贼冷静，就跟预言机一样。 这是我11号夜里截的图，里面的每根线，都代表着概率曲线。 橙色的最上面的线，就是预测OpenAI会在12月11号（美国时间）发布新模型，浅蓝色的线是预测会在12号发，天蓝色是预测13号发，黄色则是15号发。 可以看到，在11号前几天，预测GPT-5.2会在11号（美国时间）发布的概率就居高不下，一直维持在80%以上，除了当天凌晨11点到12点的区间中，有一 波巨大的震幅，发布的概率一下子变低了，其他时间的概率一直都维持在 ...

在各种小道消息，各种预测之后。 终于，在OpenAI十周年的这一天。 也就是今天的凌晨2点，GPT-5.2终于跟大家见面了。 这是Gemini 3 Pro爆火，第一次让OpenAI没有领先优势，奥特曼在内部官宣红色警戒状态之后，他们掏出的第一款模型。 也是OpenAI的十周年献礼。 而这款模型的特点也非常有意思。 OpenAI的原话是： We are introducing GPT‑5.2, the most capable model series yet for professional knowledge work.（我们正式发布 GPT-5.2，这是迄今 为止在专业知识工作方面能力最强的一代模型系列。 ） 专业知识工作，记住这个关键词，后面要考。 我们先从各种跑分上看，其实能看到，一些跑分其实没有质的飞跃，有一种数码厂开始挤牙膏的感觉。。。 | | OpenAl | Run with maximum available reasoning effort. | Anthropic | Google | | --- | --- | --- | --- | --- | | | GPT-5.2 | GPT-5 ...

是我关注的一个博主，Howie.Serious发的。 昨天刷到了一条非常有意思的推特。 他发了一个很有趣的点，就是即使是世界上现在最牛逼的NanoBananaPro，在世界知识如此屌爆的情况下，AI，还是没有办法生成左手写字的图片。 这事特别有意思。 我立马用Gemini上的NanoBananPro试了下。 果然翻车了，而且是非常稳定的翻车。 我又直接用Lovart跑了十几种张图，只对了2次，其他的，全错。 我又去试了其他的大模型，包括chatgpt、seedream，grok，也在这个小小的提示词上全军覆没。 刷刷刷给我生成了一堆右手，让我都有点混乱了，我那一瞬间都在怀疑是不是我自己分不清左右了。。。 我又尝试了一些进阶版。 比如，右手拿着苹果左手写字。 这个已经非常明确了吧，我已经给他做限制了。 还是会生成右手写字左手拿苹果的图。。。 GPT直接给我玩鬼畜了。 非常的倔强。。。 在好奇之下，我又试了一些其他的case。 比如，让一个人左手拿着橘子右手拿着苹果。 甭管是谁，就算是蜘蛛侠来了也没用，也得用右手。。。 左手拎着一只鸡，右手拎着大高达，翻车。 翻车。 穿个不同颜色的写字，翻车。。。 左手举起魔法棒 ...

Core Viewpoint - The article discusses the open-sourcing of AutoGLM by Zhipu, highlighting its significance in the context of mobile AI agents and the potential for innovation in this space [2][5][11]. Group 1: Open-Sourcing of AutoGLM - Zhipu has released the AutoGLM mobile agent framework and the AutoGLM-Phone-9B model as open-source, marking a significant development in mobile AI technology [2][6]. - The open-sourcing comes at a time when the Doubao mobile assistant has been banned, positioning AutoGLM as a viable alternative in the mobile AI landscape [5][13]. - The article draws parallels between the open-sourcing of AutoGLM and historical tech movements, suggesting that it could lead to a proliferation of applications similar to what happened with Stable Diffusion [13][19]. Group 2: Deployment Modes and Privacy - AutoGLM offers three deployment modes: local deployment, cloud deployment, and hybrid deployment, each with varying levels of privacy and performance [6][9]. - Local deployment ensures maximum privacy as all data processing occurs on the device, while cloud deployment requires careful handling of data transmission [6][9]. - The article emphasizes the importance of privacy in AI applications, suggesting that future advancements in mobile chip technology will enable more powerful local processing [6][19]. Group 3: Implications for the Future - The open-source nature of AutoGLM could democratize access to mobile AI agents, allowing individuals to create personalized assistants that run locally on their devices [19][21]. - The article reflects on the potential societal changes that could arise from widespread adoption of personal AI agents, including shifts in how individuals interact with technology [25][29]. - It suggests that the evolution of mobile AI agents could lead to a new era of user empowerment, where individuals have greater control over their digital interactions [19][29].

Core Viewpoint - The article discusses the recent experiences and challenges faced by users of the Doubao mobile assistant, highlighting its initial appeal and subsequent issues with major apps like WeChat and Alipay, which led to user restrictions and account bans [1][2][19][25]. Group 1: Product Experience - Doubao mobile assistant has gained popularity, with Nubia phones equipped with it selling out quickly, indicating strong market interest [2]. - The initial user experience was positive, with features like task automation and integration with apps being well-received [3][5][7]. - However, after a live demonstration, users faced significant issues, including account restrictions from major platforms, severely impacting usability [19][25][26]. Group 2: Industry Dynamics - The article draws parallels between the current situation and historical battles for control over digital entry points, emphasizing that the competition is shifting from traditional platforms to AI assistants [29][30][61]. - Major platforms view the emergence of AI assistants as a threat to their business models, leading to aggressive actions against such technologies [28][46]. - The narrative suggests that the rise of AI assistants could disrupt existing power structures in the app ecosystem, potentially benefiting users but threatening the survival of established platforms [41][46][55]. Group 3: Future Outlook - The author expresses optimism about the technological advancements in AI, suggesting that improvements in processing power will eventually address current limitations and privacy concerns [63][64]. - There is a cautionary note about the unpredictability of how AI will manifest in the future, urging users to be careful with sensitive information until more robust solutions are available [67]. - The article concludes with a reflection on the chaotic nature of emerging technologies, suggesting that while current experiences may be frustrating, they are part of a larger evolution towards a more integrated AI-driven future [70][74].

Core Viewpoint - The article emphasizes the transformative capabilities of Lovart's new text editing feature, which significantly enhances design efficiency and creativity, potentially replacing traditional design tools like Photoshop [8][41][101]. Group 1: Lovart Membership and Features - The author purchased a premium annual membership for Lovart during a promotional event, highlighting the value of the included tools like NanoBanana Pro and other AI applications [2][3][4]. - The membership provides access to advanced features such as text editing, which is seen as revolutionary for designers [8][9]. Group 2: Text Editing Functionality - Lovart's new text editing feature allows users to modify text within images easily, addressing a common pain point in design work where text cannot be edited after image generation [19][20]. - The process involves uploading an image, extracting text, and editing it directly in a user-friendly interface, which is a significant improvement over traditional methods [30][32]. Group 3: Enhanced Design Capabilities - The combination of text editing and other features like Touch Edit allows for seamless modifications of both text and styles, increasing overall design efficiency [75][66]. - The Mockup feature enables designers to apply their work onto various templates, streamlining the process of creating presentation-ready designs [76][78]. Group 4: Industry Impact - The article suggests that Lovart's capabilities represent a shift in the design industry, moving away from traditional tools and methods towards more intuitive, AI-driven solutions [41][100]. - The author reflects on the historical challenges faced by designers and contrasts them with the ease of use provided by Lovart, indicating a significant evolution in design practices [90][96].

Core Viewpoint - The article discusses the recent updates to Ant Group's Lingguang platform, highlighting its new features that allow users to create mini-games and applications easily, emphasizing the accessibility of technology for ordinary users [3][28][67]. Group 1: Lingguang Platform Features - Lingguang now supports not only mini-applications but also mini-games, expanding its functionality [28]. - The platform has received overwhelmingly positive feedback from users, indicating its effectiveness and user-friendliness [5][7]. - Users can create applications by simply stating their needs, without requiring any coding knowledge, which lowers the technical barrier [26][67]. Group 2: User Experience and Feedback - Users have expressed excitement about the platform, with many noting its intuitive design and strong performance compared to other AI products [6][7]. - The platform's ability to generate applications quickly has been highlighted, with examples of educational games being created in seconds [41][60]. - The article mentions specific user experiences, such as a history teacher creating a game based on the Three Kingdoms to make learning more engaging [35][41]. Group 3: Future Potential and Vision - The article envisions a future where AI can generate artistic materials for games, further enhancing the creative possibilities for users [59]. - It emphasizes the importance of making technology invisible, allowing users to focus on their ideas rather than the technical complexities behind them [64][67]. - The ultimate goal is for individuals to transform their ideas into reality effortlessly, marking a significant revolution in how technology is utilized [68][70].

昨晚，AI视频领域，终于来了一点新东西。 可灵，掏出了一个全新的多模态视频大模型，可灵 O1 。 弹窗上，居然显示连发5天，后面除了可灵O1，还有新货。 而这也是第一次， 在AI视频领域，有人把参考生视频、文生视频、首尾帧生视频、视频内容修改、风格重绘、镜头延展等等多种能力，融合到了 这个大一统的可灵 O1 模型之中。 而我，也做了一个小片子，来给大家展示一下，它的能力。 可以说，这就是AI视频领域的Nona Banana。 目前，已经正式上线，所有的人也都可以玩到。 可灵 O1 的这个O，就是Omni的缩写，跟GPT-4o的那个o意思一致，这单词来自拉丁语前缀，意思就是"all，所有、一切"。 基本现在大模型圈已经有了一个心照不宣的默契，就是谁在名字里加个 Omni，基本就是在对外说。 我是一个多模态大一统的基座模型。 进去以后，就可以看到这么一个界面。 可以上传图片和视频，也可以用主体。 作为可灵超创，我在上周其实就已经拿到了内测资格，在体验了几天，花了2万多积分，做了2个小片子以后，说实话，很多的玩法让我很惊喜。 所以，我也想来跟大家，真实的聊一聊，我对可灵 O1 的评价。 话不多说，正式开始。 首先， ...

就在刚刚，豆包的手机助手，终于发布了。 然后我就拿到了一台个非常有趣的东西，豆包手机助手，不过还是技术预览版。 载体是一个跟中兴合作的工程样机。 为了让我们体验豆包手机助手，直接现搓的。 快憋死我了。 上周其实豆包的朋友，就跟我说说有个很有意思的新东西，想不想测试一下。 我说那必须要啊。 她就神奇的问我： 我为了演示，给你们录了下载的全过程。 十几年前乔布斯心中的siri，在这一刻，我觉得才真正的具象化了出来。 先给大家看看，这个豆包手机助手，在手机里，能干出什么花活。 比如，下载手游的时候，它不仅能够帮我下完游戏，还能把游戏内部的安装包也一并给我下了。 就像这样。 视频我快放了一下，整个过程大概花了七八分钟。 我当场献上我的膝盖。。。 在拿到以后，我就， 把我的备用机上的所有的数据和微信都移过去了，在深度使用一周，我想说，这玩意，真的没有辜负我的预期。 这就是一个，基于大模型能力的，真正的AI手机助手。 苹果的apple intelligence还是个饼，但豆包真正意义上的先来了。 但实际用的时候，豆包手机助手，就一个超级牛逼的一点。 就是它每一次执行任务，全部是后台运行的，不会抢占你的手机操作界面，运行状 ...

Core Insights - DeepSeek has launched a new model, DeepSeekMath-V2, which emphasizes self-verifiable mathematical reasoning, addressing limitations in previous AI models that focused solely on final answers [1][8][30]. Group 1: Model Capabilities - DeepSeekMath-V2 can not only provide answers but also self-check its problem-solving steps, allowing it to identify and correct its own mistakes [3][49]. - The model has achieved performance levels comparable to Olympic gold medalists, excelling in competitions such as IMO 2025 and Putnam 2024 [5][6][50]. Group 2: Philosophical Context - The model's development responds to concerns raised by AI experts about the gap between AI performance in assessments and real-world problem-solving capabilities [12][26]. - The approach taken by DeepSeekMath-V2 reflects a shift from external validation to internal self-assessment, promoting a deeper understanding of mathematical reasoning [50]. Group 3: Methodology - DeepSeekMath-V2 employs a dual-structure system with a Generator that creates solutions and a Verifier that critically evaluates these solutions for logical consistency and accuracy [46][49]. - The introduction of a Meta-Verifier ensures that the evaluation process is fair and accurate, enhancing the overall reliability of the model [49]. Group 4: Performance Metrics - In the IMO competition, DeepSeekMath-V2 solved 5 out of 6 problems, demonstrating its high-level capabilities [50]. - In the Putnam Competition, it scored 118 out of 120, showcasing its ability to tackle extremely challenging mathematical problems [50].