Jacob Coxon 已公开辞去在 Anthropic 的职位，结束了他在 Anthropic 和 OpenAI 从事预训练研究的三年任期。离职时，他对行业走向发出严正警告，认为这两家公司都未以负责任的方式行事。他指出，这些实验室正卷入争先研发可自我改进的超级智能的竞赛，实际上是在拿全球安全冒险一搏。

Coxon 担心，那些可能带来革命性影响的超人类系统正迅速发展，影响范围从高级黑客技术到夺取大量资源与权力。他强调这种进展并未放缓。尽管这些组织在外界看来波澜不惊，但他认为真正参与构建这些系统的人确实担心在十年内会发生灾难性后果，把这种威胁视为现实风险而非宣传噱头。

尽管内部存在如此严重的恐惧，开发仍在继续，两家主要实验室的原因各有不同。 Coxon 认为在 OpenAI，员工并未把这一文明风险真正内化；而在 Anthropic，风险虽然被充分认识，但公司陷入争先实现超级智能的竞赛，抱着"必须赢，因为不能指望别人能安全完成"的心态。

Coxon 抨击这种做法是傲慢的赌博，认为人类的未来不应由私人公司的内部决策来决定。他质疑所谓加速对齐的做法，指出那需要一种非凡的自信——相信不存在更好、更安全的路径。他呼吁美国各研究所加强协调，甚至提出可能需要更激进的措施，例如暂时中止开发，以阻止导致不可控系统的全球竞赛。

在结尾，他敦促同行研究者重新审视参与这些项目的决定，质问在没有对底层系统有基本而严谨理解的情况下就启动超级智能强化学习实验是否负责任。他劝告同行不要抱有发展不可避免的宿命论，而应在这一关键时刻停下来，深思他们工作的严重后果。 Jacob Coxon has publicly resigned from his position at Anthropic, concluding a three-year tenure spent conducting pretraining research at both Anthropic and OpenAI. In his departure, he issued a stark warning regarding the industry's direction, arguing that neither company is behaving in a responsible manner. He contends that these labs are currently engaged in a high-stakes race toward self-improving superintelligence, effectively gambling with global safety in the process.

The core of Coxon's concern lies in the rapid development of superhuman systems capable of revolutionary impacts, from advanced hacking to the acquisition of significant resources and power. He insists that this progress is not decelerating. While many within these organizations project a sense of outward composure, he maintains that the individuals responsible for building these systems genuinely fear the potential for catastrophic outcomes before the end of the decade, viewing the threat as a reality rather than a marketing tactic.

The reasoning behind why this development continues despite such severe internal fears differs between the two major labs. At OpenAI, Coxon suggests a lack of deep internalization regarding the civilizational stakes. Conversely, at Anthropic, the risks are well-understood, but the company is locked in a competitive race to reach superintelligence first, operating under the belief that they must win because they cannot rely on others to do so safely.

Coxon criticizes this approach as a form of hubristic gambling, asserting that the future of humanity should not be determined by the internal decision-making processes of a private company. He challenges the notion of a fast-tracked alignment process, suggesting that it demands an extraordinary level of confidence that no better, safer trajectories exist. He advocates for increased coordination between U.S. labs and raises the possibility that more drastic measures, such as a temporary moratorium on development, might be necessary to prevent a global race toward potentially uncontrollable systems.

Concluding his message, Coxon urges fellow researchers to reconsider their commitment to these projects. He questions whether it is responsible to initiate superintelligent reinforcement learning runs without a fundamental, rigorous understanding of the underlying systems. Instead of adopting a fatalistic attitude that these developments are inevitable, he encourages his peers to pause and reflect on the grave implications of their work during this critical period.

所提供的内容并非实质性文章，而是面向 OpenReview 平台的浏览器验证页面，内容仅包含用于引导用户访问网站的安全检查和导航链接。由于不存在可供总结的信息性文章、研究论文或叙述性内容，无法提炼出关键点、论点或细节。请提供您希望我总结的具体文章或内容，我会按要求为您处理。 The provided text does not contain a substantive article to summarize, as it is a browser verification page for the OpenReview platform. The content consists only of security checkpoints and navigational links intended for users accessing the site.

Because there is no informational article, research paper, or narrative present in the input, it is impossible to generate a summary of key points, arguments, or details. Please provide the text or the content of the specific article you would like summarized, and I will be happy to process it according to your requirements.

一切始于 Xteink X3，这款可完全编程的电子墨水阅读器天然适合做各种实验。给固件加了开机动画、掷骰子等小功能后，通过浏览器热点传文件变得越来越让人抓狂。既然它外观像纸张，最直观的想法就是让它像纸一样原生支持打印。

为此的目标是让 MacBook 把它识别为一台真正的打印机，需要用到 Internet Printing Protocol (IPP)，该协议通过 HTTP 通信，用于查询打印机能力和提交作业。硬件被配置为以设备名 penguin 广播自己，声明单色 300 dpi 输出及 Apple 、 PWG raster 等格式；通过实现 Bonjour discovery，在网络上宣告服务，从而把电脑"骗"成把它当成一台无驱动的通用打印机来处理。

最大的技术难题是内存管理：一张标准信纸、 300 dpi 的未压缩数据超过 8 MB，而 X3 只有 400 KB RAM 。解决办法不是把整页放进内存，而是把页面当作流来处理。构建了一条转换管道，按行解码并抖动图像，直接写入显示器预留的内存，这样可以增量地拼接页面。借助显示器现有架构的这一巧妙做法，大幅降低了内存需求，为网络栈处理传入数据腾出了空间。

做完这些优化后，MacBook 的系统对话框能成功检测到这台"打印机"。第一次测试是用 Preview 打印一页漫画，电子墨水屏上呈现出清晰锐利的图像。这个过程既实用又有成就感，把阅读器变成了能通过 Wi‑Fi 接收文档的多功能工具。

除了显示当前打印内容，系统还把每个作业以 BMP 文件归档到设备的 SD 卡上，相当于建立了一个数字输出托盘，用户可以在阅读器上直接浏览过去的打印件。包括自定义打印服务器和必要的网络配置在内的整套实现，证明了通过改造标准协议能为熟悉的硬件赋予全新的功能。 The journey began with the Xteink X3, an e-ink reader that offered full programmability, which naturally invited experimentation. After customizing the firmware with small additions like boot animations and dice-rolling features, the process of transferring files via a web browser hotspot became increasingly frustrating. This friction led to a simple, intuitive realization: since the device already resembled paper, it should ideally function like it by supporting native printing.

To achieve this, the goal was to make the device appear as a legitimate printer on a MacBook. This required leveraging the Internet Printing Protocol (IPP), which communicates via HTTP to handle operations like checking printer capabilities and submitting jobs. The hardware was configured to advertise itself as a device named penguin, specifying monochrome 300 dpi output and formats such as Apple and PWG raster. By implementing Bonjour discovery, the device could announce its services on the network, effectively tricking the computer into recognizing it as a driverless, universal printer.

The primary technical hurdle was memory management, as a standard letter-sized page at 300 dpi requires over 8 MB of uncompressed data, while the X3 only offers 400 KB of RAM. To solve this, the approach was shifted from trying to store the whole page in memory to processing it as a stream. By building a transformation pipeline that decoded and dithered the image row by row directly into the display's reserved memory, the system could assemble the page incrementally. This clever use of existing display architecture significantly reduced the required RAM, clearing enough space for the network stack to handle incoming data.

Following these optimizations, the printer was successfully detected by the MacBook's system dialogs. The first test, printing a manga page from Preview, resulted in a crisp, clear image on the e-ink screen. The process proved both practical and rewarding, turning the reader into a versatile tool that can now receive documents over Wi-Fi.

Beyond just displaying the current print, the system archives every job as a BMP file on the device's SD card. This effectively creates an digital output tray, allowing the user to browse through past printouts directly on the reader. The entire implementation, including the custom printer server and the necessary network configurations, serves as a testament to the power of hacking standard protocols to give familiar hardware entirely new functionality.

数学正面临一场愈演愈烈的危机：那些富有产出的开放性问题正以不可再生的方式被耗尽。尽管可能提出的数学问题无穷无尽，真正能够揭示深刻见解或建立广泛联系的，只有一小部分。识别这些有潜力的问题是一项微妙且主观的工作，需要对该领域历史性难度格局的把握。在这个语境下，难度本身成了一种导航工具：太简单、几乎不可能或与主流理论脱节的问题，通常不如那些处在知识边界、能够催生后续研究的问题有价值。

历史上，技术和方法论的进步扩展了可达的研究边界。虽然这些工具降低了某些具体任务的难度，但它们也在"抹平"研究的地形，可能掩盖原本帮助数学家发现下一批有前途问题的结构。在当下的人工智能时代，这种影响被进一步放大：对"AI 能办到"与"AI 难以触及"之间缺乏清晰、稳定的界限。再加上许多 AI 公司对失败结果和内部流程守口如瓶，研究者面对的是一个被肆意开采却缺乏传统标记的地形，学界难以区分有意义的发现和靠蛮力耗尽而来的"成果"。

这带来一种危险的激励机制：识别有前途的问题变得比提出解决方案本身更为稀缺和珍贵。仅仅一则对某问题的兴趣传闻，就可能触发大规模的 AI 驱动行动，在人类学者尚无机会深入参与前，就将该研究领域快速"平整"。若此趋势持续，数百年来的开放科学传统可能被逆转，数学家被迫将研究方向保密，长期来看这将损害学科的未来，也会打击早期研究者的积极性，让他们觉得自己的工作被自动化、快餐式的产出所贬值。

为维持数学生态，学界必须从单纯追求解答的模式，转向重视识别洞见的模式。不加选择地使用自动化工具去挖掘解答，会以牺牲长期进步为代价换取短期成果。更可持续的做法是为数学贡献建立社会与职业规范——就像现代食物银行会拒绝随意或低质的捐赠，而只接收真正需要的物资那样。关注一个解法对相邻问题揭示了什么信息，以及为何该问题最初如此困难，才能确保学科继续繁荣，而不是被自己的工具吞噬。 Mathematics faces a growing crisis as the collection of fruitful open problems is being depleted in a non-renewable fashion. While the field of possible mathematical questions is infinite, only a small subset possesses the potential to reveal deep insights or connections. Identifying these promising problems is a subtle, subjective task that relies on a historical understanding of a field's difficulty landscape. In this context, difficulty acts as a navigational tool. Problems that are too easy, too impossible, or disconnected from broader theory are generally less valuable than those that lie at the edge of current knowledge, where they can serve as productive catalysts for further research.

Technological and methodological advancements have historically helped mathematics by enlarging the sphere of what is reachable. While these tools reduce the difficulty of specific tasks, they also flatten the landscape, potentially obscuring the geometry that helps mathematicians identify the next set of fertile questions. In the current era of artificial intelligence, this effect is amplified by the absence of clear, stable boundaries between what is AI-feasible and what remains AI-hard. Because AI companies often withhold their negative results and internal processes, researchers are left with a landscape that is being aggressively mined without the traditional markers that once allowed the community to distinguish between meaningful discovery and brute-force exhaustion.

This phenomenon creates a dangerous incentive structure where identifying a promising problem has become more precious and scarce than the solutions themselves. The mere rumor of interest in a problem can trigger mass AI-driven efforts that flatten the research area before human scholars have the opportunity to engage with it properly. If this trend continues, it threatens to reverse centuries of open science by forcing mathematicians into secrecy to protect their research directions, ultimately causing long-term damage to the future of the field and discouraging early-career researchers who feel their work is being devalued by automated, rapid-fire output.

To sustain the mathematical ecosystem, the community must transition from a model that prioritizes raw solutions to one that values the identification of insights. The indiscriminate use of automated tools to extract solutions creates short-term results at the cost of long-term progress. A more sustainable approach would involve establishing social and professional standards for mathematical contributions, much like a modern food bank that rejects arbitrary or low-quality donations in favor of what is truly needed. By focusing on what a solution reveals about neighboring problems and why it was difficult in the first place, mathematicians can ensure that the field continues to flourish rather than being consumed by its own tools.

Inception 正式发布了 Mercury 2.5，标志着其生产级模型线取得重大进展。基于 Mercury 2，新版本在智能能力上实现了显著跃升，公司通过大量客户反馈和真实生产故障案例的分析验证了这一点。数据驱动的做法让团队得以调整评估标准，打造出在保持前代高速度和低成本优势的同时，整体性能提升约 40% 的模型。

作为迄今开发出的最大规模的扩散语言模型，Mercury 2.5 旨在与 GPT-5.6 Luna 、 Gemini 3.5 Flash-Lite 和 Claude Haiku 4.5 等前沿模型竞争。它拥有 260K tokens 的上下文窗口，在标准 NVIDIA GPUs 上可实现每秒 1,107 tokens 的吞吐量。模型还引入了可调推理、并行工具调用和 schema-aligned JSON 等专业功能，并以极具竞争力的价格提供；发布期间可享受标准费率 80% 的折扣。

Mercury 2.5 的应用场景广泛，从复杂搜索流水线到交互式语音代理均可受益。在搜索场景中，低延迟使得在一次用户交互内完成多次连续调用（如查询重写与事实核验）成为可能。在语音 AI 领域，OpenCall 等公司报告称中位响应延迟已降至 170 毫秒以下，这对于营造自然、流畅的对话体验至关重要，因为延迟会破坏交互节奏。

编程助手在上下文压缩和模型路由等高频任务中也能借助 Mercury 2.5 提高效率。将这些辅助调用交由 Mercury 处理，开发平台通常能显著降低运营成本和延迟。为进一步支持此类工作流，Inception 推出了面向严格延迟场景优化的 Mercury Voice 和面向智能路由的 Mercury Router 预览版，后者可根据质量、速度与成本动态将提示路由到最合适的模型。

展望未来，Inception 已在研发下一代更大规模的模型，目标是在不牺牲当前扩散架构效率的前提下再次实现能力飞跃。对当前技术感兴趣的开发者与企业可通过 Inception API 、 Baseten 和 OpenRouter 访问 Mercury 2.5 。公司将持续在训练与基础设施方面创新，以不断扩展模型能力，显示出强劲的发展势头。 Inception has officially released Mercury 2.5, marking a significant advancement in their production model lineup. Building on the foundation of Mercury 2, the new iteration represents a substantial leap in intelligence, which the company verified through extensive customer feedback and analysis of real-world production failure cases. This data-driven approach allowed the team to sharpen their evaluation metrics, resulting in a model that maintains the high speed and low cost that defined its predecessor while offering a 40% increase in overall performance.

As the largest diffusion language model developed to date, Mercury 2.5 is designed to compete with frontier models like GPT-5.6 Luna, Gemini 3.5 Flash-Lite, and Claude Haiku 4.5. It features a context window of 260K tokens and achieves a throughput of 1,107 tokens per second on standard NVIDIA GPUs. Furthermore, the model introduces specialized capabilities, including tunable reasoning, parallel tool calls, and schema-aligned JSON, all while being offered at a highly competitive price point. To encourage adoption, the model is launching with an 80% discount on standard rates.

The practical applications for Mercury 2.5 are diverse, ranging from complex search pipelines to interactive voice agents. In search scenarios, the model's low latency allows for multiple consecutive calls, such as query rewriting and fact verification, to occur within a single user interaction. Similarly, in the realm of voice AI, the model enables remarkably fast response times, with companies like OpenCall reporting a drop in median response latency to under 170 milliseconds. This performance is critical for creating natural, fluid conversations where delays are otherwise disruptive.

Coding assistants also benefit from the efficiency of Mercury 2.5, particularly in high-frequency tasks like context compaction and model routing. By offloading these supporting calls to Mercury, development platforms can reduce operational costs and latency simultaneously, often by significant margins. To further support these workflows, Inception is also introducing previews for Mercury Voice, which is optimized for tight latency constraints, and Mercury Router, an intelligent system that dynamically directs prompts to the most suitable model based on quality, speed, and cost requirements.

Looking ahead, Inception is already focused on the development of their next, even larger model, which aims to provide another major jump in capabilities without sacrificing the efficiency of their current diffusion-based architecture. For developers and enterprises interested in the current technology, Mercury 2.5 is accessible through the Inception API, Baseten, and OpenRouter. The company remains committed to expanding the potential of their models through continued innovation in training and infrastructure, signaling a strong trajectory for their future releases.

Deltafin 是一个实验性开源项目，目标是在普通消费者硬件上本地运行完整的 2.8-trillion-parameter Kimi K3 模型。不同于通过剪枝或量化权重换取性能的做法，Deltafin 更重视保持 MoE (Mixture of Experts) 模型的原始质量。通过使用单一的 native binary，Deltafin 确保每个生成的 token 都由 Kimi K3 作为唯一权威决定，较小的 draft models 仅作为猜测器，其输出必须由主模型显式验证。

目前的实现侧重极致的硬件效率，特别针对配备 128 GB 内存并用多块 SSD 流式加载 expert weights 的 Apple Silicon 平台。基准测试显示，在 M5 Max MacBook Pro 上，Deltafin 在稳定解码时约能达到 1.00 token per second 。性能高度依赖存储吞吐，数据表明增加驱动器收益递减，因为每层需要读取的 16 个 expert reads 中最慢的那一项决定了总体速度。

虽然项目在优化上已有显著进展，但在一些技术难题上仍需突破，尤其是初始响应速度。目前由于在 prefill 阶段对模型层进行了重复读取，一个 512-token 的 prompt 在生成首个 token 前会有超过六分钟的延迟。开发者认为这是存储与架构上的瓶颈，而非硬件本身的限制，计划在后续迭代中通过优化来缩短这一延迟。

项目设计上提供了灵活的部署方式，既支持 full 安装，也支持 stream 模式。 stream 模式让用户以更小的存储起步，并在使用过程中逐步在本地缓存 experts 。 Deltafin 还内置了一个与 OpenAI 兼容的服务器，便于与各种客户端集成，并强制执行严格、贪婪且可重现的生成流程，以保证与模型预期输出一致。

总之，Deltafin 致力于在家用基础设施上拓展可能性。通过示范如何在不牺牲模型完整性的前提下运行大规模模型，项目希望为更广泛的 local-AI 社区做出贡献。开发者强调其核心任务是探索，并记录每次实验与基准测试，以透明披露在可及的高端消费者系统上运行 frontier models 的成本与能力。 Deltafin is an experimental, open-source project designed to run the full, 2.8-trillion-parameter Kimi K3 model locally on consumer hardware. Unlike other approaches that prune or quantize model weights to achieve performance, Deltafin prioritizes maintaining the original quality of the MoE (Mixture of Experts) model. By leveraging a single native binary, it ensures that Kimi K3 remains the sole authority for every generated token, with small draft models acting only as guessers that require explicit verification by the main model.

The current implementation focuses on extreme hardware efficiency, specifically targeting Apple Silicon setups with 128 GB of RAM and multiple SSDs for streaming expert weights. Benchmarks indicate that, when using an M5 Max MacBook Pro, Deltafin achieves approximately 1.00 token per second for steady decoding. The performance is heavily reliant on storage throughput, with data showing that adding more drives yields diminishing returns, as the total speed is constrained by the slowest of the 16 expert reads required for each layer.

While the project has made significant strides in optimization, there are still technical hurdles to overcome, particularly regarding initial processing speeds. A 512-token prompt currently faces a delay of over six minutes before the first token is generated due to redundant re-reads of model layers during the prefill phase. The developers have identified this as a storage and architectural bottleneck rather than a fundamental limitation of the hardware, and they have planned optimizations to address this latency in future iterations.

The project is structured to offer flexibility in how users deploy the model, providing both a "full" installation and a "stream" mode. The streaming option allows users to start with a smaller storage footprint and build a local cache of experts over time as they interact with the model. Additionally, Deltafin includes an OpenAI-compatible server, which allows it to integrate with various clients while enforcing a strict, greedy, and reproducible generation process that mirrors the model's intended output.

Ultimately, Deltafin serves as a research effort to push the boundaries of what is possible on home infrastructure. By demonstrating that large-scale models can be executed without compromising their integrity, the project aims to contribute to the broader local-AI community. The developers emphasize that their primary mission is exploration, documenting every experiment and benchmark to provide transparency into the costs and capabilities of running frontier models on accessible, albeit high-end, consumer systems.

Meta 推出 Muse，这是一款用于管理日常事务并自动化处理各种琐事的个人 AI 代理。与以对话为主的传统 AI 助手不同，Muse 以代理式 AI 运行，能够执行多步任务、浏览网页并与多种第三方应用交互。它运行在一个配备专用浏览器的持久化虚拟机上，即便用户不在应用内，也能完成预约、填写在线表单或处理客服咨询等操作。

使用体验以简洁的对话交互为核心。用户可通过专用应用或 WhatsApp 用自然语言下达指令。收到目标或任务后，Muse 会制定行动计划、跟踪进度并主动推进必要步骤。如果完成某项任务所需的工具不存在，系统可以构建自己的工具，同时还能与 Email 、 Calendars 和 Instagram 等现有服务集成，保持工作流程的连贯性。

在安全和隐私方面，系统内置多重机制予以保障。 Muse 把用户登录凭据存放在安全的凭证库中，代理本身无法直接访问；并正在接入 1Password 等工具以进一步保护敏感信息。用户购物时，系统会生成一次性卡号，确保商家无法获取真实支付信息。符合条件的购买还受 Link 的购买保护覆盖；公司强调用户与 Muse 的对话不会用于 Meta 的广告系统。

控制权始终掌握在用户手中，系统在执行关键操作前会征求用户同意。无论是发送 Email 还是完成交易，用户都可以审阅代理拟定的计划、批准或拒绝操作，并查看代理的完整历史记录与即将执行任务的审计日志。权限可按连接逐一管理，便于用户在不同场景下灵活设定 AI 的自主程度。

Muse 目前提供限额的免费使用，并为需要更高使用量的用户提供付费订阅。作为持续运行的后台代理，它旨在将模式从被动辅助转变为主动管理——根据用户的具体目标与常设指令，持续监控信息并自动执行任务。 Meta has introduced Muse, a personal AI agent designed to manage daily tasks and automate busywork across various facets of life. Unlike traditional AI assistants that primarily handle conversational exchanges, Muse operates as an agentic AI, meaning it has the capability to perform multi-step tasks, navigate the web, and interact with various third-party applications. It functions on a persistent, dedicated virtual machine equipped with a specialized browser, allowing it to complete activities such as booking appointments, filling out online forms, or managing customer service inquiries even while the user is away from the app.

The user experience is centered around simple, conversational interactions. Users can communicate with Muse through a dedicated app or via WhatsApp, providing instructions in plain language. Once a goal or task is assigned, Muse formulates an action plan, tracks progress, and proactively manages the necessary steps. If a required tool for a specific task is unavailable, the system is designed to build its own tools, and it can integrate with existing services like email, calendars, and Instagram to maintain seamless workflows.

Security and privacy are prioritized through several built-in mechanisms. Muse stores user logins in a secure credential store that the agent itself cannot access, and it is adding integration for tools like 1Password to further protect sensitive information. When a user makes a purchase, the system generates a one-time card number to ensure that the merchant never gains access to the user's real payment details. Furthermore, eligible purchases are covered by Link's purchase protections, and the company emphasizes that user conversations with Muse are not shared with Meta's advertising systems.

Control remains firmly in the hands of the user, as the system is designed to seek approval before performing critical actions. Whether the task involves sending an email or completing a transaction, users can review the agent's proposed plan, approve or deny actions, and monitor a complete audit trail of the agent's history and upcoming tasks. These permissions can be managed on a per-connection basis, providing flexibility to dictate how much autonomy the AI has in different contexts.

Muse is currently available with a usage limit for free, with options for a paid subscription for users who require greater capacity. By functioning as a continuous background agent, it aims to shift the paradigm from simple assistance to active management, where the AI proactively monitors information and executes tasks based on the user's specific goals and standing instructions.

OpenAI 已推出 ChatGPT Images 2.5，一款旨在提升数百万用户创意工作流程的先进模型。该版本在细节清晰度、纹理丰富度和光影自然度上都有显著提升。值得注意的是，与上一代相比，生成延迟最高降低了 50%，让用户能更快地迭代视觉创意。

该模型在使用参考照片时更能保持主体一致性，并能在多轮对话中更可靠地执行编辑指令。为增强 ChatGPT 内的创作体验，公司新增了如 Sketch 的工具，允许用户在对话中直接绘制，作为最终输出的视觉参考。其它功能还包括针对传单等常用格式的模板、可在图像上直接添加注释以便集中编辑，以及新的共享选项，允许用户附上原始提示词，便于他人基于其继续创作。

面向开发者，OpenAI 通过 API 发布了两款不同的模型：GPT-Image-2.5 Flare 作为多数应用的标准高性能选择，兼顾速度与质量；GPT-Image-2.5 Sunburst 则面向需要高精度编辑与控制的专业创意工作流程。这些工具已被 Adobe 、 Runway 和 Higgsfield AI 等公司整合，以简化生产与专业创作任务。

安全仍是此次发布的核心，新模型集成了现有的安全防护、提示词与图像检测机制，并支持 C2PA 元数据以确保透明度。该更新已向所有 ChatGPT 层级的用户开放（包括桌面端和移动端），开发者也可通过 API 平台立即访问这些新模型。 OpenAI has introduced ChatGPT Images 2.5, a state-of-the-art model designed to improve the creative workflow for millions of users. This iteration focuses on generating images with sharper details, richer textures, and more natural lighting. Notably, it also achieves up to a 50% reduction in generation latency compared to its predecessor, allowing users to iterate on their visual concepts much faster.

The model is built to better maintain subject consistency when working from reference photos and provides more reliable adherence to editing instructions over multi-turn conversations. To enhance the creative process within ChatGPT, the company has added new tools like Sketch, which allows users to draw directly in the chat as a visual guide for the final output. Other features include templates for popular formats like flyers, the ability to add direct comments to images for focused editing, and new sharing options that let users include their original prompts so others can build upon them.

For the developer community, OpenAI is releasing two distinct models via the API. GPT-Image-2.5 Flare serves as the standard, high-performance choice for most applications, offering improved speed and quality. Meanwhile, GPT-Image-2.5 Sunburst is tailored for premium creative workflows that require high-precision editing and control. These tools are already being integrated by companies such as Adobe, Runway, and Higgsfield AI to streamline production and professional creative tasks.

Safety remains a core component of this release, with the new model incorporating existing safeguards, prompt and image checks, and C2PA metadata to ensure transparency. The update is currently available to users across all ChatGPT tiers, including desktop and mobile, while developers can access the new models immediately through the API platform.

OpenAI 宣布在数学领域取得重大突破，解决了 Navier.Stokes existence and smoothness problem，这是七个 Millennium Prize Problems 之一。近约 90 年来，数学家们一直争论三维平滑流体运动是否会在有限时间内崩解为奇异性，即流体速度不受约束地趋于无限。研究人员借助一款先进且尚未公开的内部 AI 模型，给出了一个解析证明，并在 Lean 中完成了形式化证明，表明在平滑力的作用下，最初平滑的流体确实可能出现这种奇异性。

证明的核心在于一个涡旋的演化：涡旋向内螺旋并被拉长，使流体速度在能量保持有限的情况下趋于无限。技术上的关键是 Navier.Stokes equations 中力的精确平衡——加速度、压力梯度与粘性之间的相互作用共同促成了这一种解的破裂。由此该研究表明，流体运动的连续体近似最终可能失效，后续演化必须转向分子层面的建模才能描述。

为实现这一成果，OpenAI 使用了由其最新高性能内部模型驱动的多智能体系统。约 10,000 个并发智能体组成的大型群体，在代码执行与互联网检索等工具的辅助下，经过数日协作，反复迭代各种问题表述。该过程始于对 Euler regularity problem 的一次成功尝试，这成为通向 Navier.Stokes equations 解决方案的垫脚石。智能体被鼓励探索多条路径，并定期用 Codex 汇总见解以打磨最终证明。

组织澄清称，他们参与该项目是被新一代 AI 模型所展现的前所未有能力所驱动，而非为争取 Millennium Prize 。尽管另有独立研究小组也在攻关相关问题，OpenAI 表示他们的发现是通过多智能体系统与人类研究者的独立努力取得的。在 Lean 中对证明的形式化验证凸显了 AI 作为严谨数学研究伙伴的日益价值。

这一里程碑反映了人工智能领域正在发生的快速变革。 OpenAI 强调，这不是终点，而是一个需要谨慎引导的新进步时代的标志。随着能力的持续推进，他们将致力于模型的可控、负责任且具备问责机制的发展，以确保这些进步惠及全人类。 OpenAI has announced a significant breakthrough in mathematics by resolving the Navier.Stokes existence and smoothness problem, one of the seven Millennium Prize Problems. For approximately 90 years, mathematicians have debated whether smooth three-dimensional fluid motion can break down into a singularity, where fluid speeds grow without bound in finite time. By utilizing an advanced, unreleased internal AI model, researchers produced both an analytical proof and a formalization in Lean, demonstrating that such a singularity can indeed occur in an initially smooth fluid under the influence of a smooth force.

The proof centers on the behavior of a vortex, which spirals inward and elongates in a way that allows the fluid's velocity to reach infinite speeds while maintaining finite energy. The technical achievement lies in the precise balancing of forces within the Navier.Stokes equations, where acceleration, pressure gradients, and viscosity interact to permit this breakdown. By establishing this outcome, the research confirms that the continuum approximation of fluid motion can eventually fail, requiring a shift to molecular-level modeling to track the system's further evolution.

To achieve this, OpenAI employed a multiagent system powered by their latest, highly capable internal model. A large group of approximately 10,000 concurrent agents, assisted by tools like code execution and internet search, collaborated over several days to iterate on various problem formulations. The process began with a successful attempt at the Euler regularity problem, which served as a stepping stone that eventually led the agents to the solution for the Navier.Stokes equations. The agents were encouraged to explore diverse approaches, with insights periodically consolidated using Codex to refine the final proof.

The organization clarified that their interest in this project was driven by the unprecedented performance of their new AI models rather than a desire to claim the Millennium Prize itself. While a separate team of researchers had been working on a related problem, OpenAI confirmed that their own discovery was reached independently through the efforts of their multiagent system and human researchers. The successful verification of the proof via Lean highlights the increasing utility of AI as a partner in rigorous mathematical discovery.

Ultimately, this milestone serves as a snapshot of the rapid evolution currently unfolding in the field of artificial intelligence. OpenAI emphasizes that this is not a final destination, but rather evidence of a new era of progress that requires careful stewardship. As the organization continues to advance its capabilities, it intends to focus on the responsible, steerable, and accountable development of these models to ensure their growth benefits humanity at large.

来自 YouTube 频道 Gamers Nexus 的报告显示，LG 智能电视正在主动记录并传输关于用户及其家庭的信息，即便设备处于离线或待机状态，这一过程仍会持续。与 Level1Techs 及独立安全研究人员合作测试零售版 LG OLED 机型后，调查发现这些电视会持续扫描本地 Wi‑Fi 网络，识别附近的硬件设备，如智能手机和智能手表。

数据收集不仅限于网络扫描。调查记录表明，电视在待机模式下会通过内置麦克风录制音频。更令人不安的是，即便与互联网断开连接，这类录音也不会停止；电视会先将音频文件保存在本地，待重新联网后立即上传到 LG Ad Solutions 。

除了监控音频和网络环境外，LG 电视还使用称为 Automatic Content Recognition（ACR）的技术。该技术通过音视频采样来精确识别用户正在观看的内容，无论这些内容是通过电视自带的智能界面播放，还是来自连接到 HDMI 端口的外部设备。

调查指出，这种行为并非 LG 一家的孤立问题。包括 RTINGS 在内的其他行业分析同样表明，ACR 技术几乎被所有智能电视厂商普遍采用。跟踪技术的广泛存在，足以促使消费者考虑将智能电视完全断网。 A report from the YouTube channel Gamers Nexus has revealed that LG smart TVs are actively logging and transmitting data about their owners and households, a process that continues even when the devices are offline or in standby mode. By partnering with Level1Techs and independent security researchers to test retail LG OLED models, the investigation uncovered that these television sets are constantly scanning local Wi-Fi networks to identify nearby hardware, such as smartphones and smartwatches.

The data collection extends beyond simple network scanning. The TVs are documented to record audio through their internal microphones while in standby mode. Most disturbingly, this recording process does not stop even when the device is disconnected from the internet. Instead, the TV stores these audio files locally and uploads them to LG Ad Solutions as soon as a connection to the internet is re-established.

In addition to monitoring audio and network surroundings, LG TVs employ a practice known as Automatic Content Recognition, or ACR. This technology uses audio and video sampling to determine exactly what a user is watching, regardless of whether that content is accessed through the TV's native smart interface or via external devices connected through HDMI ports.

The investigation highlights that this behavior is not necessarily an isolated issue unique to LG. While the report focuses on the findings regarding LG devices, other industry analysis, including video coverage from RTINGS, suggests that ACR technology is a common feature across nearly every smart TV manufacturer. This widespread use of tracking technology provides a significant incentive for consumers to consider keeping their smart TVs disconnected from the internet entirely.