In addition to the nearly constant stream of model releases, in September we’ve seen price drops, new kinds of models, proofs of long-standing problems in mathematics, and continued investigations into models escaping their sandboxes. (Axios reports investigations into over 10,000 incidents.) AI has infinite patience and is fundamentally probabilistic. Given a difficult or impossible task and an unlimited token budget, an agent will eventually attempt to solve the problem in ways that you don’t expect, and may not want. It’s easy (and correct) to blame inadequate security procedures at the frontier AI labs, but AI adopters must be careful not to make the same mistakes. The humans using AI need to be accountable for what their agents do.除了几乎不间断的模型发布外，九月份我们还见证了价格下调、新型模型的出现、长期数学难题的证明，以及对模型逃离沙箱行为的持续调查。（Axios 报道称，有超过 10,000 起相关事件正在调查中。）人工智能拥有无限的耐心，且本质上是概率性的。面对一项困难甚至不可能的任务，如果给予无限的 Token 预算，智能体最终会尝试以你意想不到、甚至不希望看到的方式来解决问题。将此归咎于前沿 AI 实验室安全流程的不足是容易的（也是正确的），但 AI 的采用者必须小心，不要重蹈覆辙。使用 AI 的人类必须对其智能体的行为负责。
AI ModelsAI 模型
Model choice is starting to hinge on price and specialization as much as raw benchmark leadership. Alongside general chat models, there are now decision models that never chat, spatial models built for robot planning and camera control, forecasting models sized for a single task, and cybersecurity-specialized models kept behind an invite-only program. Specialization leads to greater efficiency and lower costs, at least in the short term. In the long term, specialized models may succumb to the “bitter lesson.”模型选择正开始从单纯的基准测试领先地位转向价格和专业化。除了通用的聊天模型外，现在还出现了从不聊天的决策模型、专为机器人规划和摄像头控制而构建的空间模型、针对单一任务调整的预测模型，以及仅在受邀项目中使用的网络安全专业模型。专业化能带来更高的效率和更低的成本，至少在短期内如此。从长远来看，专业化模型可能会屈服于“苦涩的教训”（the bitter lesson）。
- Anthropic has released Claude Opus 5.5, which it claims has performance similar to Fable 5.1, and hence similar restrictions. It’s faster and requires fewer resources to run. Anthropic has dropped prices 20% for input and output tokens and 60% for cached reads. Not to be outdone, OpenAI released GPT-6 Sol and Luna, with 50% price reductions.Anthropic 发布了 Claude Opus 5.5，声称其性能与 Fable 5.1 相当，因此也具有类似的限制。它运行速度更快，所需的资源更少。Anthropic 将输入和输出 Token 的价格下调了 20%，缓存读取价格下调了 60%。为了不甘示弱，OpenAI 发布了 GPT-6 Sol 和 Luna，价格降低了 50%。
- Anthropic has also announced Fable 5.1 and Mythos 5.1. The most significant change appears to be a 75% price reduction for cache reads, which might translate into significant savings for long-running jobs; Anthropic estimates 25%. Mythos is only available to trusted partners. Simon Willison used Fable 5.1 to animate his pelican-riding-a-bicycle pseudobenchmark.Anthropic 还宣布了 Fable 5.1 和 Mythos 5.1。最显著的变化似乎是缓存读取价格降低了 75%，这可能为长时间运行的任务带来显著的节省；Anthropic 估计可节省 25%。Mythos 仅向受信任的合作伙伴提供。Simon Willison 使用 Fable 5.1 制作了他的“鹈鹕骑自行车”伪基准测试动画。
- Anthropic released Sonnet 5.5 with claims that it’s 30% faster and 30% less expensive for most work. The new model has security limitations similar to those applied to Opus and Fable; it routes to Sonnet 5 if it’s asked to do anything out of bounds.Anthropic 发布了 Sonnet 5.5，声称其在大多数工作中的速度提高了 30%，成本降低了 30%。新模型具有与 Opus 和 Fable 类似的安全性限制；如果被要求执行任何越界操作，它会转至 Sonnet 5 处理。
- And finally, as September closes, Anthropic announces a marketplace for Claude plugins and connectors. At its launch, Claude Marketplace had over 2,000 items.最后，随着九月结束，Anthropic 宣布推出 Claude 插件和连接器市场。在发布时，Claude Marketplace 已有超过 2,000 个项目。
- OpenAI has released GPT-6 Astra, with claims that the company has achieved AGI (artificial general intelligence). Astra’s excellent benchmark scores appear to depend on the use of an unreleased harness. OpenAI has also released GPT-6.1 Sol, with per-token price reductions and claims that it is close to GPT-6 Astra in capabilitiesOpenAI 发布了 GPT-6 Astra，并声称该公司已实现通用人工智能（AGI）。Astra 出色的基准测试分数似乎依赖于使用了一个未发布的测试框架。OpenAI 还发布了 GPT-6.1 Sol，降低了每 Token 的价格，并声称其能力接近 GPT-6 Astra。
- OpenAI has solved the Navier-Stokes existence and smoothness problem, a mathematical problem in fluid mechanics. This development raises an ethical question: Did OpenAI train its system on the work of two mathematicians who were close to solving the problem themselves? It also raises practical questions about the future of mathematics. Decorated mathematician Terence Tao asks whether “the collection of good, fruitful open problems is now being mined in a non-renewable fashion.” An advisory group has been formed to help OpenAI make decisions about releasing mathematical results.OpenAI 解决了流体力学中的数学难题——纳维-斯托克斯存在性与光滑性问题。这一进展引发了一个伦理问题：OpenAI 是否利用了两位本身就快要解决该问题的数学家的工作成果来训练其系统？这也引发了关于数学未来的实际问题。知名数学家陶哲轩（Terence Tao）问道：“好的、富有成果的开放性问题集合现在是否正在以不可再生的方式被开采？”一个咨询小组已经成立，旨在帮助 OpenAI 就发布数学研究成果做出决策。
- Google has released Gemini 3.8 Flash TTS and Flash-Lite TTS. Voice options aren’t limited to a prebuilt library. These models have APIs that allow developers to describe the voice that they want or upload a sample. These custom voices are then assigned an ID so they can be reused.Google 发布了 Gemini 3.8 Flash TTS 和 Flash-Lite TTS。语音选项不再局限于预构建的库。这些模型拥有 API，允许开发者描述他们想要的语音或上传样本。这些自定义语音随后会被分配一个 ID，以便重复使用。
- Google has announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. These models are designed for live, near-real-time conversation. They can process video input. Live Extended Thinking can reason and speak at the same time.Google 宣布了 Gemini 3.8 Live 和 Gemini 3.8 Live Extended Thinking。这些模型专为实时、近乎即时的对话而设计。它们可以处理视频输入。Live Extended Thinking 能够同时进行推理和说话。
- Google has released Gemini 3.8 Flash and Flash Cyber. Flash appears to be similar to frontier models on most benchmarks, with Computer Use being the biggest exception. Flash Cyber is specialized for vulnerability detection and mitigation and is only available to defenders in the Fairwind Program.Google 发布了 Gemini 3.8 Flash 和 Flash Cyber。Flash 在大多数基准测试中似乎与前沿模型相当，最大的例外是“计算机使用”（Computer Use）能力。Flash Cyber 专门用于漏洞检测和缓解，仅向 Fairwind 项目的防御者提供。
- Google has also released TimesFM-3, a small specialized model for multivariate time series forecasting. The model weights are available on Hugging Face with a license that only allows for noncommercial use. (Source code is available under the Apache open source license.)Google 还发布了 TimesFM-3，这是一个用于多变量时间序列预测的小型专业模型。模型权重可在 Hugging Face 上获取，且许可证仅允许非商业用途。（源代码在 Apache 开源许可证下提供。）
- Xiaomi released MiMo-V2.6, its latest large language model. It’s fully open sourced, and based on benchmark results, Xiaomi claims that MiMo is the strongest open model to date. What’s more interesting is the claim that MiMo only cost $3.5 million to train.小米发布了其最新的大型语言模型 MiMo-V2.6。它完全开源，根据基准测试结果，小米声称 MiMo 是迄今为止最强大的开源模型。更有趣的是，小米声称 MiMo 的训练成本仅为 350 万美元。
- TypeSafe’s new decision model, Jev, is unlike anything else we’ve seen. It doesn’t chat; its output is always strictly typed and accompanied by probabilities that estimate correctness. It’s much faster and less expensive than other leading models. It isn’t open source, but there are already many open source clones.TypeSafe 的新决策模型 Jev 与我们见过的任何模型都不同。它不聊天；其输出始终是严格类型的，并附带有估计正确性的概率。它比其他领先模型快得多，成本也更低。它不是开源的，但已经出现了许多开源克隆版本。
- Ollaya is similar to Ollama, but for running decision open models like Laya (a clone of Jev) locally.Ollaya 与 Ollama 类似，但用于在本地运行像 Laya（Jev 的克隆版）这样的开放决策模型。
- World Labs has released Atlas, a model for “spatial intelligence.” It uses text, images, video, and 3D data to perform tasks like planning a robot’s movements or changing the camera position in a photograph.World Labs 发布了 Atlas，这是一个用于“空间智能”的模型。它使用文本、图像、视频和 3D 数据来执行诸如规划机器人运动或更改照片中摄像机位置等任务。
- The rumors that NVIDIA would buy Hugging Face are true. NVIDIA is hoping for a proliferation of models that will run on its hardware, and the company promises that Hugging Face will remain a neutral platform, without favoring one model over another.关于 NVIDIA 将收购 Hugging Face 的传言是真的。NVIDIA 希望其硬件上能运行更多的模型，并承诺 Hugging Face 将保持作为一个中立平台，不会偏袒任何特定模型。
Software development软件开发
Agents are starting to delegate to, and coordinate with, other agents rather than working solo. Claude Code can break a task apart and hand pieces to other Claude Code instances, Muse Code lets sessions message each other, and Google’s AX orchestrator exists purely to wire up sandboxes and control communications for swarms of agents doing a task together. That shift is pushing developers to rethink what a source repository needs to record, and to start asking how much all this delegation costs.智能体开始委托并协调其他智能体，而不是单打独斗。Claude Code 可以将任务拆解并分发给其他 Claude Code 实例，Muse Code 允许会话之间互相发送消息，而 Google 的 AX 编排器纯粹是为了连接沙箱并控制智能体群组协作完成任务时的通信。这种转变促使开发者重新思考源代码仓库需要记录的内容，并开始询问所有这些委托行为的成本。
- Now that Jev has caught everyone’s attention, what can you build with it? Jevmem is a memory manager that hooks into Claude Code and prunes the context at every conversational turn.既然 Jev 已经引起了大家的注意，你能用它构建什么？Jevmem 是一个内存管理器，它挂载到 Claude Code 中，并在每次对话轮次时修剪上下文。
- AX is a new agent orchestrator from Google. It isn’t an agent; it’s intended to coordinate many agents to complete a task. It creates sandboxes, wires up Git repos and other resources, and controls outbound communications.AX 是 Google 推出的一个新型智能体编排器。它本身不是智能体；其目的是协调多个智能体以完成任务。它创建沙箱，连接 Git 仓库和其他资源，并控制出站通信。
- The latest version of Claude Code can manage Claude projects, breaking a task into subcomponents and delegating the subtasks to other Claude Code instances. Another important change is the ability to read AGENTS.md if CLAUDE.md isn’t available.最新版本的 Claude Code 可以管理 Claude 项目，将任务拆分为子组件，并将子任务委托给其他 Claude Code 实例。另一个重要的变化是，如果 CLAUDE.md 不可用，它能够读取 AGENTS.md。
- Google’s CC agent is designed for families. Family members can share data with CC, which has its own user account. It could be used for filling out forms, synchronizing calendars, and other common tasks.Google 的 CC 智能体专为家庭设计。家庭成员可以与拥有独立用户账户的 CC 共享数据。它可用于填写表格、同步日历以及处理其他常见任务。
- What will replace GitHub? There’s a growing consensus that we need different kinds of source repositories to deal with the agent-assisted software development. In addition to changes to code, it’s important to record the conversations between the developer and the agent, the architectural decisions, and many other artifacts that don’t make it into traditional source control.什么会取代 GitHub？人们越来越达成共识，我们需要不同类型的源代码仓库来应对智能体辅助的软件开发。除了代码变更外，记录开发者与智能体之间的对话、架构决策以及许多其他无法进入传统源代码控制系统的工件也至关重要。
- Anthropic is merging its Claude Cowork and chat products. Anything users type in a chat session is seen by Cowork, and vice versa. Some sensitive information (health, politics, and gender) is excluded. The feature is on by default but can be disabled in settings, and memory isn’t shared with Claude Code. The company also released Claude Docs and Slides.Anthropic 正在合并其 Claude Cowork 和聊天产品。用户在聊天会话中输入的任何内容 Cowork 都能看到，反之亦然。一些敏感信息（健康、政治和性别）被排除在外。该功能默认开启，但可以在设置中禁用，且内存不与 Claude Code 共享。该公司还发布了 Claude Docs 和 Slides。
- Claude Money is a new feature that will allow users to connect their bank accounts to Claude for analysis. The product appears to be similar to a product from OpenAI.Claude Money 是一项新功能，允许用户连接其银行账户供 Claude 进行分析。该产品看起来与 OpenAI 的某款产品类似。
- OpenAI’s Agents API is now in public beta. It allows compaction, session orchestration, and tool use, and it can be deployed in OpenAI’s sandbox, a cloud provider’s sandbox, or the developer’s hardware.OpenAI 的 Agents API 现已进入公开测试阶段。它支持压缩、会话编排和工具使用，并可部署在 OpenAI 的沙箱、云服务提供商的沙箱或开发者的硬件上。
- Meta has released Muse, its AI agent. Muse is a “personal agent” designed for tasks like shopping, filling in forms, and dealing with customer service. It has its own secure credential store, so data like passwords and credit card numbers are never sent offsite.Meta 发布了其 AI 智能体 Muse。Muse 是一个“个人智能体”，专为购物、填写表格和处理客户服务等任务而设计。它拥有自己的安全凭证存储，因此密码和信用卡号等数据永远不会发送到外部。
- Some open source projects are shutting down external pull requests, which are largely AI-generated. In some cases, the developer team is using its own agents to create and manage PRs; some are using AI agents to triage external PRs.一些开源项目正在关闭外部拉取请求（Pull Requests），因为这些请求大多是由 AI 生成的。在某些情况下，开发团队正在使用自己的智能体来创建和管理 PR；还有一些团队正在使用 AI 智能体来分类外部 PR。
- Meta has launched Muse Code, another competitor to Claude Code. One important new feature is the ability to send messages to other Muse Code sessions, allowing agents to coordinate on complex problems.Meta 推出了 Muse Code，这是 Claude Code 的又一个竞争对手。一个重要的新功能是能够向其他 Muse Code 会话发送消息，允许智能体在复杂问题上进行协作。
- AI providers appear to be moving toward outcome-based pricing, at least for major corporate customers. Rather than billing by token, customers are billed for completed tasks. That approach begs the question: When is a task completed?AI 提供商似乎正在转向基于结果的定价模式，至少对于大型企业客户而言是这样。客户不再按 Token 计费，而是按完成的任务计费。这种方法引出了一个问题：任务何时才算完成？
- Now that organizations are concerned with AI budgets, the question of how to evaluate the cost of different models and agents becomes important. What should platform teams measure?既然组织开始关注 AI 预算，评估不同模型和智能体成本的问题就变得重要起来。平台团队应该衡量什么？
- ChatGPT Work was designed to compete with Claude Cowork, Microsoft Copilot Cowork, Muse Code, and other agents designed for noncoders. Simon Willison shows how Work goes beyond its competitors. It can perform tasks on the web for users, even logging in to websites without sending usernames and passwords to OpenAI; it can execute code with full internet access; it can build and deploy a web application. Whether these features are also risks is an open question.ChatGPT Work 旨在与 Claude Cowork、Microsoft Copilot Cowork、Muse Code 以及其他为非程序员设计的智能体竞争。Simon Willison 展示了 Work 如何超越其竞争对手。它可以为用户在网上执行任务，甚至在不向 OpenAI 发送用户名和密码的情况下登录网站；它可以执行具有完整互联网访问权限的代码；它可以构建并部署 Web 应用程序。这些功能是否也构成风险，目前尚无定论。
- Anthropic has given Claude Desktop access to a Chromium-based browser that’s built into Cowork, eliminating the need for a Chrome plugin when Claude needs to browse the web.Anthropic 已让 Claude Desktop 能够访问 Cowork 内置的基于 Chromium 的浏览器，从而在 Claude 需要浏览网页时无需再使用 Chrome 插件。
- TimeLord is a short Python program that, given a string up to 1,000 characters long, produces a seed for Python’s pseudo-random number generator so that repeated calls reproduce the text. It’s a surprisingly simple hack, though not a statement about randomness or the quality of Python’s PRNG.TimeLord 是一个简短的 Python 程序，给定一个最长 1,000 个字符的字符串，它会为 Python 的伪随机数生成器生成一个种子，以便重复调用时能重现该文本。这是一个出人意料的简单黑客手段，尽管这并非关于随机性或 Python PRNG 质量的声明。
Security安全
OpenAI’s experiment that attacked Hugging Face is the gift that keeps giving, but the past month has had plenty of news about more conventional attacks, many aided by AI. Security has always been a game of whack-a-mole, in which vulnerabilities are discovered and exploited as fast as defenders can patch them. AI is an important tool for defenders, and it’s constantly improving, but it’s still behind attackers, especially given the limitations placed on frontier models and the unlimited persistence that attacking agents exhibit.OpenAI 攻击 Hugging Face 的实验是一个不断带来意外结果的“礼物”，但过去一个月里，关于更传统攻击的新闻层出不穷，其中许多得到了 AI 的辅助。安全性一直是一场“打地鼠”游戏，漏洞被发现和利用的速度与防御者打补丁的速度一样快。AI 是防御者的重要工具，并且在不断改进，但它仍然落后于攻击者，特别是考虑到前沿模型所受的限制以及攻击智能体所表现出的无限持久性。
- OpenAI has postponed the release of GPT-6.1 Astra because it failed its safety tests.OpenAI 推迟了 GPT-6.1 Astra 的发布，因为它未能通过安全测试。
- NVIDIA has announced its Open Agent Safety Platform. The reference implementation includes NVIDIA OpenShell, which has been enhanced with a policy prover, and NVIDIA Sentry, a service that runs on NVIDIA DPUs.NVIDIA 宣布了其开放智能体安全平台（Open Agent Safety Platform）。参考实现包括 NVIDIA OpenShell（已通过策略证明器增强）和 NVIDIA Sentry（一种在 NVIDIA DPU 上运行的服务）。
- The Felony Bench lists known attacks by agents from the major AI labs against third parties. We don’t know if the Bench will be kept up-to-date, but tens of thousands of security incidents involving OpenAI and Anthropic are now being investigated.Felony Bench 列出了主要 AI 实验室的智能体针对第三方的已知攻击行为。我们不知道该基准测试是否会保持更新，但目前有数万起涉及 OpenAI 和 Anthropic 的安全事件正在接受调查。
- Following through on Dario Amodei’s call to control the speed of frontier model development, Anthropic, OpenAI, and Google are creating a standards consortium for governing the process of AI development. Meta, xAI, Microsoft, and the Chinese labs are all notably absent.为了落实 Dario Amodei 关于控制前沿模型开发速度的呼吁，Anthropic、OpenAI 和 Google 正在创建一个用于管理 AI 开发过程的标准联盟。Meta、xAI、Microsoft 和中国实验室均未参与。
- Agents need their own identity. Unlike the long-term identities we’re used to, agents need a short-lived identity tied to a revocable certificate and that limits access to resources appropriate for the job. That’s not all of agent security, but it’s a table stakes.智能体需要自己的身份。与我们习惯的长期身份不同，智能体需要一个与可撤销证书绑定的短期身份，并限制其访问适合该工作的资源。这并非智能体安全的全部，但这是基本门槛。
- Anthropic has published a lengthy report on the misuse of its systems by threat actors. Daniel Meissler has published a summary, digesting Anthropic’s report into 117 findings.Anthropic 发布了一份关于其系统被威胁行为者滥用的详尽报告。Daniel Meissler 发布了一份摘要，将 Anthropic 的报告提炼为 117 个发现。
- A malicious NPM malware package works by hiding malicious code in the package itself (indexed-btree) rather than simply attacking the install script. This technique makes it significantly harder to detect.一个恶意 NPM 软件包通过将恶意代码隐藏在包本身（indexed-btree）中，而不是仅仅攻击安装脚本来发挥作用。这种技术使得检测难度显著增加。
- An attack against the RSA algorithm allows forging of signatures in some situations. The attack was invented in 2007; this is the first public implementation.针对 RSA 算法的一种攻击允许在某些情况下伪造签名。该攻击于 2007 年发明；这是首次公开的实现。
- Fake CAPTCHA pages are being used to spread malware. Victims are frequently sent to those pages when they respond to a phish.虚假的 CAPTCHA 页面正被用于传播恶意软件。受害者在响应钓鱼信息时经常被引导至这些页面。
- Hugging Face has volunteered to audit AI labs for safety and alignment with human values.Hugging Face 已主动提出审计 AI 实验室的安全性及其与人类价值观的一致性。
- In an experiment designed to test AI alignment, DeepMind found that, out of 100 agents, 14% were willing to cheat, 25% were “whistleblowers” that reported cheating, and the remainder didn’t notice.在旨在测试 AI 对齐（AI alignment）的一项实验中，DeepMind 发现，在 100 个智能体中，14% 的智能体愿意作弊，25% 是举报作弊的“吹哨人”，其余的则没有察觉。
- Threat actors are building frameworks for AI agent-enabled attacks. A human in the loop is no longer needed. Fully automated attackers don’t appear to be using zero-days yet; they’re relying on known vulnerabilities.威胁行为者正在构建支持 AI 智能体攻击的框架。不再需要人在回路（human in the loop）。完全自动化的攻击者似乎尚未利用零日漏洞；他们依赖的是已知漏洞。
- OpenAI autonomous AI agents were found communicating with each other via publicly accessible Wikis, possibly to collaborate on a benchmark.人们发现 OpenAI 的自主 AI 智能体通过公开可访问的 Wiki 进行相互通信，可能是为了在基准测试中进行协作。
- OpenAI has stated that its unreleased Astra model has reached the “Critical” cybersecurity threshold, which means that it can find new vulnerabilities and run exploits against well-protected systems. Now that Astra is released, access to its cybersecurity capabilities has been limited.OpenAI 表示，其未发布的 Astra 模型已达到“关键”网络安全阈值，这意味着它可以发现新漏洞并针对受保护良好的系统运行攻击。随着 Astra 的发布，其网络安全能力的访问权限已受到限制。
Infrastructure and Operations基础设施与运营
Individuals, corporations, and even nations all face a similar problem: keeping their infrastructure under control. At a minimum, control means keeping data on a laptop, corporate server, or data center; at the other end of the spectrum it means eliminating dependencies on software and services from another nation. Any organization working through an AI transformation has to evaluate its entire stack: What do they need to control, and what can they safely delegate to others?个人、企业甚至国家都面临着一个类似的问题：保持对其基础设施的控制。至少，控制意味着将数据保留在笔记本电脑、企业服务器或数据中心；在另一端，它意味着消除对其他国家软件和服务的依赖。任何正在进行 AI 转型的组织都必须评估其整个技术栈：他们需要控制什么，又可以安全地委托给他人什么？
- DAWO is a community that’s building an open source “workspace” to support digital sovereignty for the Dutch government. The stack will include AI, an operating system based on NixOS, cloud services, and collaboration tools.DAWO 是一个正在构建开源“工作空间”的社区，旨在支持荷兰政府的数字主权。该栈将包括 AI、基于 NixOS 的操作系统、云服务和协作工具。
- Cohere now offers a confidential computing platform for artificial intelligence. The company claims that customer data is never visible to Cohere itself or any cloud providers that are in use; data is processed on GPUs whose memory is encrypted and isolated.Cohere 现在提供用于人工智能的机密计算平台。该公司声称，客户数据对于 Cohere 本身或所使用的任何云提供商都是不可见的；数据在内存经过加密和隔离的 GPU 上进行处理。
- Perplexity has announced Hybrid Compute, a feature that allows it to run models and use files and tools directly on a user’s Mac. The company claims that sensitive data will never leave the user’s computer.Perplexity 宣布了 Hybrid Compute 功能，允许它直接在用户的 Mac 上运行模型并使用文件和工具。该公司声称敏感数据永远不会离开用户的计算机。
Hardware硬件
It’s too easy to view consumer devices as innocuous things that sit around and do their job silently. Recent devices include cameras, microphones, and even EEG sensors that are constantly collecting data. Where is that data sent, how is it used, and who might have access to it? These questions need to be asked more often.人们太容易将消费设备视为无害的、静静地完成工作的物件。最近的设备包括摄像头、麦克风，甚至是不断收集数据的脑电图（EEG）传感器。这些数据被发送到哪里，如何使用，谁可能接触到它们？这些问题需要更频繁地提出。
- LG Smart Televisions have been found to record conversations and other audio, even while turned off. The conversations are sent back to LG. If the set is disconnected from the network, it will attempt to find open WiFi access points to deliver its data.LG 智能电视被发现即使在关机状态下也会录制对话和其他音频。对话会被传回 LG。如果电视与网络断开连接，它会尝试寻找开放的 WiFi 接入点来传输数据。
- In part because of backlash against Meta’s camera-enabled glasses and their abuse, its AI glasses now come with or without a camera, and can be used as hearing aids. Well-documented abuse aside, virtual reality will only succeed if there are fashionable, easily wearable products.部分由于对 Meta 摄像头眼镜及其滥用的抵制，其 AI 眼镜现在有带摄像头和不带摄像头两种版本，并可用作助听器。撇开记录在案的滥用行为不谈，虚拟现实只有在出现时尚、易于佩戴的产品时才能成功。
- Headphones, earbuds, and other devices equipped with EEG sensors are appearing on the market. They’re advertised for monitoring fatigue, monitoring sleep, and similar applications. It’s time to ask what happens at the interface between neurology and AI.配备 EEG 传感器的耳机、耳塞和其他设备正在市场上出现。它们被宣传用于监测疲劳、睡眠监测和类似应用。现在是时候问问神经学与 AI 的接口处会发生什么了。
- Microduck is a small bipedal AI-driven robot. It’s trained in simulation with open source software, and the model that results can be shared on Hugging Face. It’s affordable and is available for preorder now, shipping by Christmas.Microduck 是一款小型双足 AI 驱动机器人。它使用开源软件在模拟中进行训练，由此产生的模型可以在 Hugging Face 上共享。它价格实惠，现已开放预订，将于圣诞节前发货。
WebWeb
- Cloudflare now supports HTTP Vary, which allows servers to serve different kinds of files at the same URL. This is the “ugliest part” of the HTTP standard. It makes caching very difficult, and it probably should be avoided.Cloudflare 现在支持 HTTP Vary，允许服务器在同一 URL 下提供不同类型的文件。这是 HTTP 标准中“最丑陋的部分”。它使缓存变得非常困难，可能应该避免使用。
- WebMCP is a proposed standard that gives websites a small API to register tools that agents can discover and call. It was developed by Google and Microsoft.WebMCP 是一个提议的标准，它为网站提供了一个小型 API，用于注册智能体可以发现并调用的工具。它由 Google 和 Microsoft 开发。
- A new Twitter? Operation Bluebird is relaunching Twitter, the service bought by Elon Musk and renamed X.一个新的 Twitter？Operation Bluebird 正在重新启动 Twitter，即埃隆·马斯克购买并更名为 X 的服务。
Biology生物学
- Anthropic has built a biology lab for experimenting with AI-enabled drug development. Claude assisted in the discovery of an enzyme that might be able to perform CRISPR-like gene editing.Anthropic 建立了一个生物实验室，用于实验 AI 辅助的药物开发。Claude 协助发现了一种可能能够执行类似 CRISPR 基因编辑的酶。
- To improve its training data for biological applications, the OpenAI Foundation (OpenAI’s nonprofit parent organization) is buying data from failed biotech companies.为了改善其用于生物应用的训练数据，OpenAI 基金会（OpenAI 的非营利性母组织）正在从倒闭的生物技术公司购买数据。
- Google has released AlphaGenome Atlas, a database of every possible single letter change to human DNA, and what that change will do.Google 发布了 AlphaGenome Atlas，这是一个包含人类 DNA 每个可能的单字母变化及其影响的数据库。
