Bluesky 已正式为 AT Protocol 注册商标，涵盖 "AT Protocol""atproto" 等相关词汇。此举主要是出于防御性考虑：因第三方曾威胁就该名称的使用提出法律挑战，Bluesky 通过取得商标来保护现有开发者社区免受潜在的法律干扰，并确保该协议继续作为生态系统内各方开发工作的稳固基础。

此决定的核心并非限制使用，而是确保协议保持开放与可访问。 Bluesky 强调，他们不打算对典型的、出于善意的使用收取许可费。所建立的法律框架旨在防止不良行为者滥用该名称，同时允许开发者继续开发、编写文档和制作兼容工具，对于常规的描述性引用无需获得正式许可。

对绝大多数贡献者而言，并不会有变化。鼓励以描述性方式为开源包或社区项目命名，但不得虚假暗示其实现是官方的或经 Bluesky 特别认可。只有当商标使用从描述性转为品牌化——例如用于公司名称、商业产品或官方周边时，才需要申请许可。

Bluesky 承认去中心化和开源精神与商标法的官僚现实之间存在张力，但指出美国法律要求对商标进行积极维护才能保有所有权。采取这一措施是遵循 Linux 、 Mozilla 和 Apache Foundation 等主要开源项目的做法，目的是防止有人试图将该名称据为己有。

展望未来，Bluesky 计划最终将商标所有权移交给一个独立的、专门负责协议治理的组织。目前为实现及时保护，商标暂由 Bluesky PBC 持有，但这是临时安排。在此期间，Bluesky 承诺继续向社区开放该商标，营造一个让开发者无须担心外部法律干扰即可放心开发的安全环境。 Bluesky has officially acquired the trademark for the AT Protocol, covering terms like AT Protocol, atproto, and their variants. This move was primarily a defensive action prompted by a third party that threatened legal challenges regarding the use of the name. By securing the trademark, Bluesky aims to protect the existing developer community from potential legal interference and ensure that the protocol remains a stable foundation for everyone building within its ecosystem.

The core intent behind this decision is not to restrict usage, but rather to ensure the protocol remains open and accessible. Bluesky emphasizes that they do not intend to charge for licenses for typical, good-faith usage. The legal framework they have established is designed to prevent bad actors from co-opting the name while allowing developers to continue building, writing documentation, and creating compatible tools without needing formal permission for standard descriptive references.

For the vast majority of contributors, nothing changes. Developers are encouraged to continue naming their open-source packages or community projects in descriptive ways, provided they do not falsely imply that their implementations are official or specifically endorsed by Bluesky. The requirement for a license only arises when the use of the trademark shifts from descriptive to brand-based, such as using it in a company name, a commercial product, or official merchandise.

Bluesky acknowledges the tension between the ethos of a decentralized, open-source protocol and the bureaucratic reality of trademark law. However, they note that the U.S. legal system necessitates the active defense of a mark to maintain ownership. By taking this step, they are following the established practices of other major open-source projects like Linux, Mozilla, and the Apache Foundation, ensuring the protocol is shielded from those who might otherwise try to claim it as their own.

Looking toward the future, Bluesky intends to eventually transfer ownership of the trademark to an independent, dedicated protocol governance organization. While the trademark is currently held by Bluesky PBC for practical and timely protection, they maintain that this is a temporary arrangement. In the meantime, the company is committed to keeping the mark freely available to the community, fostering a secure environment where developers can continue to build without fear of external legal disruption.

SQLite 被广泛认为是本地数据存储的行业标准：作为一个以库形式提供的关系型数据库引擎，它免去了自定义序列化器的需求。尽管它在从嵌入式项目到服务器软件中广泛使用，但引擎存在一个明显的问题——其默认配置对现代开发往往既不直观也不理想。这些源于历史的默认设置经常导致数据不一致、性能瓶颈和意外的运行时错误。

一个主要问题是外键约束默认被禁用。不强制执行外键会带来数据损坏的风险，特别是考虑到 SQLite 的 row-id 重用策略，悬空引用可能会意外指向无关数据。同样，SQLite 通过灵活的类型亲和性允许列存储与声明的模式不符的类型，这虽然增加了灵活性，但也常导致静默失败和难以发现的 bug 。虽然有 strict tables 可以缓解这一问题，但必须逐表启用，无法全局设置，给开发者增加了额外负担。

在并发和性能方面也存在不足：默认情况下并发写操作会立即返回 SQLITE_BUSY 错误而不是等待锁释放，且 Write-Ahead Log (WAL) 模式默认关闭，限制了性能。为达到现代数据库应有的功能和稳定性，开发者不得不反复手动配置各种 pragma（如 busy_timeout 和 journal_mode）。

拒绝改变这些默认值的主要原因是对向后兼容性的严格坚持——修改既有行为可能会破坏现有软件。但这种做法也把 SQLite 限制在过去，阻碍其随现代工程实践演进。为在不损害遗留系统完整性的前提下打破这一僵局，引入类似 Rust 的 edition 系统可作为一种折中方案。

设计一个"超级 pragma"，例如 edition 标志，可以让开发者选择一套现代化、强化的默认设置。通过设置 edition，用户能够全局启用外键强制、严格类型、合理的写入超时以及像 WAL 这样的性能优化模式。基于版本的方法不仅能理清当前混乱的配置，还为未来改进提供可扩展的路径，使数据库在不抛弃旧项目的前提下不断进步。 SQLite is widely considered the industry standard for local data storage, functioning as a robust library-based relational database engine that removes the need for custom serializers. Despite its utility and prevalence in everything from embedded projects to server software, the engine suffers from a significant drawback: its default configurations are often counter-intuitive or suboptimal for modern development. These defaults, while historically rooted, frequently lead to data inconsistency, performance bottlenecks, and unexpected runtime errors.

A primary concern is the handling of foreign key constraints, which are disabled by default. Failing to enforce these constraints risks data corruption, especially since SQLite's row-id re-use policy can cause dangling references to inadvertently point to unrelated data. Similarly, SQLite allows columns to store types that do not match their defined schema through a flexible system of type affinity. While this permits a loose, dynamic approach, it often leads to silent failures and buggy code. Although strict tables exist to mitigate this, they must be enabled on a per-table basis rather than globally, creating an extra burden for developers.

The engine also presents issues regarding concurrency and performance. By default, concurrent write operations trigger immediate SQLITE_BUSY errors rather than waiting for a lock to clear, and performance is hindered by the fact that the Write-Ahead Log (WAL) mode is disabled out of the box. These hurdles force developers to rely on manual, repetitive configuration of various pragmas like busy_timeout and journal_mode to reach a level of functionality and stability expected from a modern database.

The persistent refusal to change these defaults is largely driven by a commitment to absolute backwards compatibility, as modifying established behaviors risks breaking existing software. However, this approach traps SQLite in its past, preventing it from evolving alongside current engineering standards. To break this cycle without compromising the integrity of legacy systems, the introduction of a Rust-style edition system could serve as an effective middle ground.

Implementing a super pragma, such as an edition flag, would allow developers to opt into a modern, hardened set of defaults. By setting an edition, a user could globally enable foreign key enforcement, strict typing, reasonable write timeouts, and optimized performance modes like WAL simultaneously. This version-based approach would not only clean up the current configuration landscape but also provide a scalable framework for future improvements, allowing the database to progress without leaving older projects behind.

在 20 世纪 80 年代，作者曾与自由软件运动的创始人 Richard Stallman 就开放软件的利弊争论多年。起初作者为专有控制辩护，认为公司需要所有权才能推动技术进步，但他最终认识到，软件是一种重要的知识体系，只有通过透明才能不断壮大。这样的认识转变反映了更广泛的历史潮流：像 GNU/Linux 这样的开源项目证明，公开的集体协作能够胜过私营机构的封闭系统，并最终成为现代数字世界的承重基石。

如今，先进人工智能的兴起呈现出类似但规模更大的冲突。随着前沿 AI 模型越来越被企业壁垒所封锁，这一领域面临停滞的风险。 AI 正日益成为科学、专业和个人推理的基础设施，缺乏透明性令人担忧。当模型被少数公司当作黑箱掌控时，用户就无法审查其底层机制，只能依赖那些输出难以完全验证或理解的"神谕"。

反对开放 AI 的人常以安全和潜在滥用为由，担心公布模型方法等于散布危险工具。然而，作者认为真正的安全并非依靠模糊不清来实现。正如科学界在公开研究的前提下管理风险一样，AI 行业也应以透明为先。封闭系统仍然易受泄露和越狱的威胁，而权力集中于少数公司之手，会让这些企业单方面塑造未来的数字资源库，并为他人的知识获取制定规则。

当前所谓促进开放 AI 的努力常常流于表面——提供可运行模型的代码，却隐瞒构建模型所用的数据与方法。结果是用户只能得到未经解释就能运行的"魔数"，而非一个开放、可复现的基础设施。这种开放只是企业暂时施予的恩惠，而非全球社区可以独立研究或改进的可靠共享资源。

要走出困境，就必须重新承诺将开源 AI 作为公共产品来资助和支持。尽管私营部门的创新依然重要，但不应成为唯一选择。通过设立公共计算资源拨款、促进大学与非营利组织间的协作，并要求用公共资金开发的 AI 必须保持开放，社会可以阻止开放知识的流失。最初的开源运动已证明，共享的基础能够带来更快、更稳健的创新，我们必须在 AI 领域被永久关闭于公众之外之前，将这些经验教训付诸实施。 In the 1980s, the author spent years debating the merits of open software with Richard Stallman, the founder of the free software movement. While the author initially defended proprietary control, arguing that companies required ownership to advance technology, he eventually recognized that software represents a vital body of knowledge that grows stronger through transparency. This shift in perspective mirrored a broader historical trend, where open source initiatives like GNU/Linux proved that collective, public collaboration could outpace the closed systems of private entities, ultimately becoming the load-bearing foundation for the modern digital world.

Today, the emergence of advanced artificial intelligence presents a parallel conflict, though on a much larger scale. As frontier AI models are increasingly locked behind corporate walls, the field faces the risk of stagnation. Because AI is effectively becoming the primary infrastructure for scientific, professional, and personal reasoning, the lack of transparency is concerning. When models are treated as black boxes controlled by a few firms, users lose the ability to audit the underlying processes, forcing them to rely on "oracles" whose outputs cannot be fully verified or understood.

Critics of open AI often cite safety and the potential for misuse, fearing that publishing the methods behind these models is equivalent to distributing dangerous tools. However, the author argues that true security is not achieved through obscurity. Just as the scientific community manages risks while keeping research public, the AI industry should prioritize transparency. Closed systems remain vulnerable to leaks and jailbreaks, while the concentration of power in a few companies allows those entities to unilaterally shape the digital library of the future, dictating the terms of knowledge access for everyone else.

Current efforts to promote open AI often fall short, providing the code to run a model while concealing the data and methods used to build it. This creates a state where users are left with "magic numbers" that function without explanation, rather than an open, replicable foundation. This form of openness is merely a temporary favor granted by corporations, rather than a reliable, shared resource that can be independently studied or improved by the global community.

The way forward requires a renewed commitment to funding and supporting open source AI as a public good. While private sector innovation remains important, it should not be the sole option available. By establishing public compute grants, fostering collaboration between universities and nonprofits, and mandating that AI developed with public funds remains open, society can prevent the erosion of open knowledge. The success of the initial open source movement proves that shared foundations lead to faster, more robust innovation, and it is imperative that we apply these lessons before the field of AI becomes permanently closed to the public.

Puter Labs 推出了一项可以在单个浏览器标签页内完整运行 Firefox 的新项目。它把驱动 Firefox 的 Gecko 引擎编译成 WebAssembly，从而在另一个浏览器的环境中重现传统桌面浏览器的界面和渲染能力，作为一个完整的应用运行，而不是简单的网页。

要让这个"浏览器内的浏览器"正常工作，用户需要进行一些环境配置。由于项目依赖 WebAssembly JSPI，Firefox 用户需在 about:config 中手动启用 javascript.options.wasm_js_promise_integration 标志。完成配置后，引擎会在标签页中初始化并启动 Firefox 界面，为网络探索提供一个沙箱环境。

项目还为希望试验性能和架构的用户提供了高级选项，包括通过 WebGL 实现的 GPU 加速，以及一个将 JavaScript 转译为 WebAssembly 的实验性即时编译器（JIT）。尽管该 JIT 功能仍属试验阶段，但可作为在此虚拟化场景中提升引擎速度和效率的测试平台。

嵌套浏览器的网络由 Wisp 服务器管理，充当 WebSocket 代理。当用户通过该 Firefox 实例浏览网页时，流量会通过由 Puter 托管的该代理转发。整个方案由 Puter.js 支持——一个用于在浏览器中处理复杂网络任务的库，使内部的 Firefox 实例能够在多层虚拟化下无缝加载外部网页内容。 Puter Labs has introduced a project that runs a full version of the Firefox web browser entirely within a single browser tab. This is achieved by compiling the Gecko engine, which powers Firefox, into WebAssembly. By doing so, the interface and rendering capabilities of a traditional desktop browser are recreated inside the environment of another web browser, functioning as a complete application rather than a simple web page.

To ensure this browser-in-a-browser functions correctly, users need specific environment configurations. Because the project relies on WebAssembly JSPI, those using Firefox must manually adjust their settings by accessing the about:config menu and enabling the javascript.options.wasm_js_promise_integration flag. Once this is configured, the engine initializes and launches the Firefox interface within the tab, providing a sandbox for web exploration.

The project also includes advanced features for users who want to experiment with performance and architecture. There is an option for GPU acceleration, which utilizes WebGL to handle the rendering of web content, alongside an experimental Just-In-Time compiler that translates JavaScript into WebAssembly. While this JIT feature is marked as experimental, it serves as a testing ground for improving the speed and efficiency of the browser engine within this virtualized context.

Networking within this nested browser is managed through a Wisp server, which acts as a WebSocket proxy. When users browse the web through this instance of Firefox, their traffic is routed through this proxy, which is hosted by Puter. This setup is supported by Puter.js, a library designed to handle complex networking tasks in the browser, allowing the internal instance of Firefox to load external web content seamlessly despite the layers of virtualization involved.

Grok Build 是 xAI 开发的基于终端的 AI 代码代理，旨在作为与代码库交互的一体化工具。它以全屏终端用户界面运行，能够理解项目文件、执行 shell 命令、进行网络搜索并管理长时间运行的任务。设计上保持灵活，既支持交互式使用，也支持用于脚本或持续集成流水线的无头模式，并可通过 Agent Client Protocol 集成到现有编辑器中。

该工具用 Rust 编写，仓库作为 xAI 主 monorepo 的定期同步公共镜像。用户可以通过简单的安装脚本在 macOS 、 Linux 和 Windows 上安装预构建二进制包；也提供了从源码构建的明确说明，但需要 Rust toolchain 和 protoc 。启动后，主二进制 xai-grok-pager 会通过浏览器完成用户认证流程。

项目按职责划分为多个 crates：核心包括 TUI 、代理运行时本身、诸如文件编辑和终端访问的工具实现，以及用于版本控制和命令执行的工作区管理。这样的模块化设计便于关注点分离，通用逻辑与共享代码则集中存放在指定目录中。

仓库内的开发流程优先针对单个 crate 进行测试和验证，因为完整工作区的构建可能非常耗时。尽管为开发者提供了丰富资源，项目明确指出当前不接受外部贡献。代码主要以 Apache License, Version 2.0 授权，并对第三方及 vendored 组件（例如 Mermaid diagram stack 和其他移植的工具实现）给出了相应声明。

项目提供了详尽的文档，内容涵盖键盘快捷键、斜杠命令，以及 MCP 服务器、插件和自定义钩子的配置等各方面。结合可扩展和可定制的特性，Grok Build 被定位为可以深度融入开发者工作流的实用助手。 Grok Build is the terminal-based AI coding agent developed by xAI, designed to serve as a comprehensive tool for interacting with your codebase. Operating as a full-screen terminal user interface, the agent is capable of understanding project files, executing shell commands, performing web searches, and managing long-running tasks. Its design is intended to be flexible, supporting interactive use, headless operation for scripting or continuous integration pipelines, and integration into existing editors through the Agent Client Protocol.

The tool is written in Rust and the repository acts as a periodically synced public mirror of the main xAI monorepo. Users can install prebuilt binaries for macOS, Linux, and Windows via simple installation scripts. For those who prefer to build from source, the project provides clear instructions, though it requires the Rust toolchain and protoc. Once launched, the primary binary, known as xai-grok-pager, facilitates user authentication through a browser-based flow.

The project structure is organized into various crates that handle specific responsibilities within the agent runtime. The core includes packages for the TUI, the agent runtime itself, tool implementations like file editing and terminal access, and workspace management for version control and execution. This modularity allows for a cleaner separation of concerns, with common and shared logic kept in designated directories.

Development within the repository is structured to prioritize per-crate testing and validation, as full-workspace builds can be time-consuming. While the repository provides resources for developers, it specifies that external contributions are not currently accepted. The code is primarily licensed under the Apache License, Version 2.0, with specific notices provided for third-party and vendored components, such as the Mermaid diagram stack and other ported tool implementations.

Comprehensive documentation is available for those looking to understand the intricacies of the tool, covering everything from keyboard shortcuts and slash commands to the setup of MCP servers, plugins, and custom hooks. This documentation, along with the project's ability to be extended and customized, underscores its role as a functional assistant designed to integrate deeply into a developer's workflow.

Alex Turner 解释了自己从 Google DeepMind 辞职的原因，源于 2026 年初发生的一系列事件。 Department of Homeland Security 的特工在致人死亡事件后，他发现 Google 向该机构提供云服务，从而对科技公司参与政府行动感到越来越不安。特别是在 Pentagon 向 Anthropic 等公司施压，要求签署在致命性自主武器和大规模监控方面几乎不设限制的军事 AI 合同时，他认为这是 AI 业界划定伦理底线的关键时刻，但许多知名人士和组织并未明确表态，这让他深感失望。

Turner 一开始试图在公司内部推动改变，并动员有影响力的 AI 研究人员。他曾尝试与 Stuart Russell（长期致力于反对自主武器）和 Yoshua Bengio 等人接触。尽管这些人在公众面前以重视 AI 安全著称，Turner 认为在面对来自 U.S. government 的具体且高风险的政治压力时，他们未能采取实质性行动。 Turner 对 International Association for Safe and Ethical AI (iaseai) 尤为失望，认为该组织未能兑现支持 Anthropic 拒绝不道德合同条款的承诺。

在公司内部，Turner 向 Google 领导层游说，主张对政府合同建立一套有原则、约束力强的框架。他获得了同事广泛支持，在致 Jeff Dean（Google Chief Scientist）的一封请愿书上收集了超过 250 个签名。 Turner 希望通过增加内部成本并提出清晰、合乎伦理的替代方案，迫使公司在签署无限制军事协议时优先考虑其公开宣示的 AI 原则。他甚至与 Jeff Dean 进行了直接会面，希望借此争取其在反对 Pentagon 最后通牒中的领导作用。

尽管有这些努力，Google 最终仍与 Pentagon 签署了一份机密协议，允许将相关技术用于"任何合法的政府目的"，但在防止自主武器或大规模监控方面几乎没有具有约束力的保护条款。 Turner 认为，Google 的领导层，包括此前曾签署反对致命性自主武器承诺的高层，最终将政治权宜和公司利益置于所宣称的道德承诺之上。他注意到 Demis Hassabis 和其他高管在协议签署后仍留任，这让他判断公司管理层奉行的"参与决策（seat at the table）"策略并未带来实质性的伦理护栏。

在耗尽了所有内部影响途径后，Turner 认为自己已无法凭良心继续在该公司工作。他认为 Google DeepMind 的内部治理实践以失败告终，原因在于追逐利润和对政治报复的恐惧最终压倒了组织的道德承诺。协议签署后他选择辞职，认为要实现真正改变需要有打破既有界限的勇气，并且仅依赖有权势领导人的私人道德承诺而缺乏制度化的执行，无法有效应对存在性的 AI 风险。 Alex Turner explains his reasons for resigning from Google DeepMind, stemming from a series of events in early 2026. Following the killing of citizens by Department of Homeland Security agents, Turner discovered that Google provided cloud services to the agency. He grew increasingly concerned about the tech industry's involvement in government operations, particularly when the Pentagon pressured companies like Anthropic to sign contracts for military AI without restrictions on lethal autonomous weapons or mass surveillance. Turner viewed this as a critical moment for the AI industry to establish ethical boundaries, yet he found that prominent figures and organizations often failed to take a stand.

Turner initially sought to influence change by working within Google and by mobilizing influential AI researchers. He attempted to engage leaders such as Stuart Russell, who had built a long career campaigning against autonomous weapons, and Yoshua Bengio. Despite their public reputations for prioritizing AI safety, Turner felt these luminaries avoided taking meaningful action when faced with specific, high-stakes political pressure from the U.S. government. Turner was particularly disappointed by the International Association for Safe and Ethical AI (iaseai), which he believes failed to fulfill its promise to support Anthropic's right to refuse unethical contract terms.

Internally, Turner lobbied Google leadership to adopt a robust, principled framework for government contracts. He garnered significant support from colleagues, collecting over 250 signatures on a petition addressed to Jeff Dean, Google's Chief Scientist. Turner hoped that by creating internal costs and presenting a clear, ethically sound counteroffer, he could force the company to prioritize its stated AI principles over the pressure to sign unrestricted military deals. He even held a direct meeting with Jeff Dean, hoping to secure his leadership in opposing the Pentagon's ultimatum.

Despite these efforts, Google ultimately signed a classified agreement with the Pentagon, allowing for "any lawful government purpose" with few binding protections against autonomous weaponry or mass surveillance. Turner felt that Google's leadership, including those who had previously signed pledges against lethal autonomous weapons, prioritized political expediency and corporate interests over their stated ethical commitments. He observed that Demis Hassabis and other senior figures remained at the company even after the deal was signed, leading Turner to conclude that the "seat at the table" strategy favored by management had failed to produce meaningful ethical guardrails.

Having exhausted his internal avenues for influence, Turner decided he could no longer work for the company in good conscience. He argues that the internal governance experiment at Google DeepMind failed because the pressure for profit and the fear of political retribution ultimately outweighed the organization's moral commitments. Following the signing of the deal, Turner chose to resign, believing that real change requires a willingness to break established boundaries and that relying on the private ethical pledges of powerful leaders without institutional enforcement is an insufficient strategy for mitigating existential AI risk.

Thinking Machines 发布了 Inkling，一个拥有 9750 亿参数的 Mixture-of-Experts Transformer 模型。作为通用基础模型，Inkling 设计为便于定制，支持最高 100 万个 token 的上下文窗口，并可原生在文本、图像和音频上进行推理。此外，公司还提供了 Inkling-Small 的预览版，这是一个为降低延迟和成本而优化的 2760 亿参数版本。

Inkling 的一大特色是可控的推理开销，允许用户根据具体需求在性能与 token 效率之间进行权衡。其架构对多模态输入采用无编码器方法并使用相对位置嵌入，开发者发现这比常规方案在长序列外推方面更为有效。模型在 45 万亿个 token 上进行了预训练，并通过大规模强化学习进一步精炼；值得注意的是，训练过程中出现了链式思考推理随着时间自发压缩的现象。

该模型在设计上强调认识论，注重置信度校准和对指令的遵循。为提高可靠性，研究团队使用自动化评分系统训练 Inkling，其中包括通过网页搜索验证事实声明的 claims grader 。安全评估显示，Inkling 具备强大的内置防护能力，能够有效拒绝有害请求，并在良性任务上保持高性能。

为便于实际应用，Thinking Machines 通过其 Tinker 平台开放了 Inkling 的微调能力。该生态系统使用户能够定制模型、评估表现并将其部署到不同的基础设施提供商上。团队还演示了一个案例，Inkling 成功自我微调以采用特定风格约束，展示了其作为开发工作流中具代理能力伙伴的潜力。

Inkling 的完整权重已在 Hugging Face 上发布，其中包含为 NVIDIA Blackwell 系统优化的版本。通过提供基础模型与深度定制工具，Thinking Machines 致力于帮助组织构建反映专业知识和特定运营需求的 AI 。 Thinking Machines has released Inkling, a Mixture-of-Experts transformer model with 975 billion total parameters. Designed as a broad foundation model, Inkling is built to serve as a base for customization, supporting a context window of up to 1 million tokens and native reasoning across text, images, and audio. Alongside the primary model, the company has provided a preview of Inkling-Small, a 276-billion-parameter version optimized for lower latency and cost.

Inkling distinguishes itself through controllable thinking effort, allowing users to balance performance against token efficiency based on their specific needs. Its architecture utilizes an encoder-free approach for multimodal inputs and relative positional embeddings, which the developers found to be more effective for long-sequence extrapolation than standard alternatives. The model was pretrained on 45 trillion tokens and further refined through large-scale reinforcement learning, which, notably, resulted in an emergent compression of its chain-of-thought reasoning over time.

The model is designed with a focus on epistemics, emphasizing calibrated confidence and instruction following. To ensure reliability, researchers trained Inkling using automated grading systems, including a claims grader that verifies factual assertions via web search. Safety evaluations indicate that Inkling demonstrates strong built-in safeguards, effectively refusing harmful requests while maintaining high performance on benign tasks.

To facilitate practical application, Thinking Machines has made Inkling available for fine-tuning through their Tinker platform. This ecosystem enables users to customize the model, evaluate its performance, and deploy it across various infrastructure providers. Demonstrating the model's self-improvement capabilities, the team showcased an instance where Inkling successfully fine-tuned itself to adopt a specific stylistic constraint, illustrating its potential as an agentic partner in development workflows.

Inkling's full weights are available on Hugging Face, including versions optimized for NVIDIA Blackwell systems. By providing both the foundational model and the tools for deep customization, the company aims to support organizations in developing AI that reflects specialized knowledge and unique operational requirements.

Clocks.dev 是一个由 Lev Miseri 创建的数字时钟设计创意库，展示了用实验性方式可视化时间流逝的各种可能性。平台汇集了许多独特且常常打破常规的计时界面，远离传统指针或标准数字显示，每一款时钟都是开源的，鼓励对美学与计时交叉感兴趣的设计师与开发者共同创作与交流。

这个集合囊括了从高度技术化、数学化的诠释到更俏皮或极简的设计等广泛的理念。例如，有的时钟依靠几何推导与平面时间研究，使用复杂公式计算小时与分钟之间的空间关系；也有把时间简化到最基本元素的作品，比如基于"困倦"情绪的时钟，或通过密集文字域并高亮显示来指示时间的布局，模仿了 word clock 的风格但带有鲜明的数字化气质。

除了功能性展示，网站还收录了对数据呈现与排版进行创新的设计。用户可以看到将数字组织成广阔数字田、使用二进制信号，或以重复数字创造视觉节奏的界面。这些设计把看时间这件平凡的行为变成了动态的互动体验，挑战了我们对数字与时间单位在屏幕上应如何排列的固有认知。

通过为这些实验性项目提供展示与分享的空间，Clocks.dev 展示了在数字媒体语境下对时间解读的多样性。无论是受巴洛克风格启发的复杂布局，还是干净利落的极简数字序列，平台都在形式与功能之间寻找平衡。它仍在不断演进，开发者可在此贡献自己的设计，确保这个集合持续以新颖且富有创造力的方式记录和计算每一秒。 Clocks.dev serves as a creative repository for digital clock designs, acting as a showcase for experimental ways to visualize the passage of time. Created by Lev Miseri, the platform invites users to explore a collection of unique, often unconventional timekeeping interfaces that move far beyond the traditional ticking hands or standard numerical displays. Each clock featured on the site is open-source, encouraging a collaborative spirit among designers and developers interested in the intersection of aesthetics and chronometry.

The collection highlights an impressive range of conceptual approaches, from highly technical and mathematical interpretations to more whimsical and minimalist designs. For instance, some clocks rely on geometric proofs and planar time studies, utilizing complex formulas to calculate the spatial relationship between hours and minutes. Others strip time down to its simplest components, such as a "sleepy" mood-based clock or a dense word-field layout that highlights text to indicate the current time, mimicking the familiar style of word clocks but with a distinct digital flair.

Beyond the purely functional, the site hosts designs that play with data representation and typography. Users can find clocks that organize digits in expansive number fields, utilize binary signals, or repeat digits in creative patterns to form a visual rhythm. These designs transform the mundane act of checking the time into a dynamic interaction, as the various interfaces challenge our conventional understanding of how digits and units of time should be ordered on a screen.

By providing a space for these experimental projects, Clocks.dev illustrates how diverse the interpretation of time can be when framed through digital media. Whether through the use of baroque-inspired layouts or clean, minimalist numerical sequences, the platform demonstrates a commitment to both form and function. It remains an evolving resource where developers can contribute their own designs, ensuring the collection continues to grow with new and inventive ways to count the seconds.

Codex Micro 是由 OpenAI 与 Work Louder 合作开发的一款专业硬件，专为作为 agent 驱动工作（agentic work）的指挥中心而设计。该设备定价为 230 美元，旨在将关键的代理操作置于触手可及之处，弥合数字化 AI 工作流与物理交互之间的裂缝，成为用户工作空间的延伸，从而更直观地管理 AI 辅助任务。

设备的核心在于其交互性能，能为用户提供实时反馈。每个按键都配备可自定义的 RGB 灯光，对应某个代理的状态。通过这种视觉信号，用户可以一目了然地判断代理是在思考、运行、等待还是已完成任务，无需在浏览器标签或窗口间来回切换以跟踪进度。

硬件包含多种为编程与开发流程量身定制的物理控制装置。专用操纵杆（joystick）可让用户即时触发常用的 Codex 功能，例如调试错误、重构代码或审查 pull requests 。设备还设有功能键（command keys），作为常用操作的专用快捷键，比如新建聊天、按键通话（push-to-talk）或接受与拒绝建议。

对于需要更细致调整 AI 表现的用户，设备配备了旋转编码器拨盘（rotary encoder dial），可随时调节 AI 代理的推理强度。转动拨盘可在简单任务中选择更快、更轻量的模式，或在复杂项目中提升推理能力。机身采用 CNC 加工的聚碳酸酯与铝材制造，支持 USB-C 与 Bluetooth 连接，兼容 Mac 和 Windows 环境。

每台设备还附带一套完整的 Codex Icon Keyset，包含 32 个自定义键帽，方便用户标注各自的工作流；包装内同样配有低矮型机械轴（low-profile mechanical switches），以保证耐用且回馈灵敏的敲击手感。通过将这些物理输入直接整合进软件界面，Codex Micro 旨在为使用 agent 驱动 AI 系统的开发者与高级用户简化交付与发布流程。 The Codex Micro is a specialized hardware tool developed through a collaboration between OpenAI and Work Louder, designed specifically to serve as a command center for agentic work. Priced at $230, this device aims to bridge the gap between digital AI workflows and physical interaction by keeping essential agent operations within immediate reach. It functions as an extension of the user's workspace, allowing for more intuitive navigation of AI-assisted tasks.

At the heart of the device are its interactive capabilities, which provide users with real-time feedback. Each key is equipped with customizable RGB lighting that corresponds to the status of a specific agent. This visual signaling allows users to see at a glance whether an agent is currently thinking, running, waiting, or has completed a task, eliminating the need to constantly toggle between browser tabs or windows to track progress.

The hardware includes several tactile controls tailored for coding and development workflows. A dedicated joystick allows users to trigger common Codex skills instantly, such as debugging errors, refactoring code, or reviewing pull requests. Additionally, the device features command keys that act as dedicated shortcuts for frequent actions like starting new chats, using push-to-talk, or accepting and rejecting suggestions.

For those requiring more nuanced control over AI performance, the device incorporates a rotary encoder dial. This feature enables users to adjust the reasoning level of their AI agents on the fly. By turning the dial, users can opt for faster, lighter performance during simple tasks or increase the reasoning power for more complex projects. The unit is constructed from CNC-machined polycarbonate and aluminum, supporting both USB-C and Bluetooth connectivity for compatibility with Mac and Windows environments.

Every purchase includes the device itself alongside a comprehensive Codex Icon Keyset, providing 32 custom keycaps that allow users to label their specific workflows. The package also includes low-profile mechanical switches, ensuring a durable and responsive typing experience. By integrating these physical inputs directly into the software interface, the Codex Micro seeks to streamline the shipping process for developers and power users working with agentic AI systems.

一台服役已十三年的 HP StoreVirtual 服务器，配备两颗 Ivy Bridge Xeon E5‑2690 v2 处理器且无 GPU，已成功运行 Google 的 Gemma 4 26B‑A4B 语言模型。尽管缺少 AVX2 、 FMA3 等常见现代指令集，它仍能以约每秒 5 个 token 的速度生成文本。该项目实际证明，现代 AI 能力可以在遗留企业硬件上得到利用，挑战了"高性能 AI 只能靠云订阅或昂贵现代基础设施"的观念。

最初团队试图复现另一篇关于在旧硬件上运行语言模型的文章中的流程，但首次尝试失败：用于 Mixture‑of‑Experts 优化的专用 ik_llama.cpp 分支依赖于 Intel 在 2014 年 Haswell 架构中才支持的 AVX2 指令。借助 AI 助手分析构建失败并排查代码库，团队找到了必要的修改，使软件在遇到缺失指令时能够回退到标量计算。

技术上的关键在于修复模型计算调度器中的一个"静默失败"。图构建器生成了 Ivy Bridge 上没有对应计算路径的操作，导致模型产出看似流畅但毫无意义的文本。通过用可移植的标量循环重写缺失操作，并确保图构建器在禁用特定优化标志时能正确处理，模型才得以正常运行。这个过程凸显了现代工程的一项关键能力：审计并改造性能关键代码以适配具体硬件约束。

除了具体的技术成就外，该实验还强调了实操经验在维护遗留系统时的重要性。成功并不在于从头重写整个引擎，而在于系统性地诊断故障并实施有针对性的补丁。关闭诸如为新指令集设计的运行时权重重新打包等选项后，这台服务器现在可作为 AI 任务的可行本地回退方案，表明爱好者和企业有可能在本应报废的旧硬件上维持高效的本地 AI 能力。 A repurposed, thirteen-year-old HP StoreVirtual server, equipped with two Ivy Bridge Xeon E5-2690 v2 processors and no GPU, has been successfully configured to run Google's Gemma 4 26B-A4B language model. While this hardware lacks the modern instruction sets, such as AVX2 and FMA3, typically required for such tasks, it now generates text at a respectable five tokens per second. This project serves as a practical demonstration that modern artificial intelligence capabilities can be harnessed on legacy enterprise hardware, challenging the assumption that high-performance AI is exclusively accessible through cloud subscriptions and expensive, modern infrastructure.

The project began as an attempt to replicate a workflow described in another article regarding running language models on older hardware. However, the initial attempt failed because the specialized ik_llama.cpp fork, which includes optimizations for Mixture-of-Experts models, relied on AVX2 instructions that were not introduced until Intel's Haswell generation in 2014. By utilizing an AI assistant to analyze the build failures and troubleshoot the codebase, the necessary modifications were identified to allow the software to fall back on scalar math instead of reaching for non-existent CPU instructions.

The technical core of the fix involved addressing a silent failure within the model's compute dispatcher. Because the graph builder was emitting operations for which no compute path existed on the Ivy Bridge architecture, the model produced fluent-looking but entirely nonsensical output. By re-implementing the missing operations using portable scalar loops and ensuring the graph builder correctly handles cases where specialized optimization flags are disabled, the model was finally able to function. This process highlights a critical skill for modern engineering, namely the ability to audit existing, performance-critical code and make it compatible with specific hardware constraints.

Beyond the specific technical achievement, this experiment underscores the value of hands-on expertise in navigating legacy systems. The success of the project did not require rewriting the entire engine from scratch, but rather systematically diagnosing why the system was failing and implementing targeted patches. By turning off flags like runtime weight repacking, which is designed for newer instruction sets, the server now operates as a viable local fallback for AI tasks. This highlights the potential for enthusiasts and businesses alike to maintain high-functioning, local AI capabilities on older hardware that might otherwise be discarded.