OpenSSH on musl Linux has reportedly allowed access to passwordless accounts since 2016.据报道,musl Linux 上的 OpenSSH 自 2016 年起可能允许进入无密码账户。
OpenSSH on musl Linux has reportedly allowed access to passwordless accounts since 2016. The issue relates to how Unix historically represented accounts that should never permit login, including accounts marked with an asterisk in the password file. The behavior can therefore expose accounts that were intended to be inaccessible.据报道,musl Linux 上的 OpenSSH 自 2016 年起可能允许进入无密码账户。问题与 Unix 处理本应禁止登录的账户的历史方式有关,包括在密码文件中用星号标记的账户。因此,原本应不可访问的账户可能被暴露。
The source reports that frontier LLMs can read and write neural-network weights directly.资料称,前沿大语言模型已经能够直接读取和写入神经网络权重。
The source reports that frontier LLMs can read and write neural-network weights directly. GPT-6 Astra can infer what a network was trained to do from its raw weights and create small game-playing networks that outperform ones trained with PPO for 100 million steps. The capability could matter for model optimization and interpretability.资料称,前沿大语言模型已经能够直接读取和写入神经网络权重。GPT-6 Astra 可以仅凭原始权重判断网络的训练目标,还能从零编写击败经过 1 亿步 PPO 训练的小型游戏网络。这种能力可能对模型优化和可解释性具有重要意义。
HackerTube presents a YouTube-style interface for 6,454 videos sourced from 39,834 Hacker News comments dating back to 2007.HackerTube 提供了一个 YouTube 风格的界面,收录了自 2007 年以来从 39,834 条 Hacker News 评论中提取的 6,454 个视频。
HackerTube presents a YouTube-style interface for 6,454 videos sourced from 39,834 Hacker News comments dating back to 2007. The collection is available through hackertube.org. It turns years of community recommendations into a browsable video archive.HackerTube 提供了一个 YouTube 风格的界面,收录了自 2007 年以来从 39,834 条 Hacker News 评论中提取的 6,454 个视频。该视频集合可通过 hackertube.org 浏览。它将多年社区推荐整理成了一个可探索的视频档案。
HackerBooks catalogs about 16,500 books mentioned by at least two different Hacker News commenters between 2007 and today.HackerBooks 收录了 2007 年至今至少被两名不同 Hacker News 评论者提到过的约 1.65 万本书。
HackerBooks catalogs about 16,500 books mentioned by at least two different Hacker News commenters between 2007 and today. Each book links back to the comments that mentioned it, and the dataset can be downloaded from the about page. The creator notes that the system counts mentions without distinguishing praise from criticism, so some entries may be missing or mismatched.HackerBooks 收录了 2007 年至今至少被两名不同 Hacker News 评论者提到过的约 1.65 万本书。每本书都链接回提及它的评论,数据集也可从关于页面下载。创建者说明,系统只统计提及次数,不区分赞扬和批评,因此部分书籍可能遗漏或匹配错误。
Mistral AI has released the weights for Mistral Large 4, enabling researchers to run and fine-tune the model locally.Mistral Large 4 的权重已对外公开,研究者和开发者可以在本地部署与微调。
Mistral AI has released the weights for Mistral Large 4, enabling researchers to run and fine-tune the model locally. This move reduces reliance on hosted APIs and accelerates experimentation, but also brings licensing, safety, and governance considerations to the fore in open-weight AI projects.Mistral Large 4 的权重已对外公开,研究者和开发者可以在本地部署与微调。此举降低对 API 的依赖,促进实验,但也引发许可、安全与治理方面的讨论。
CVE-2026-98432 is described as a critical, remotely exploitable bug in the Linux kernel's TCP code that has existed since 2016.CVE-2026-98432 被描述为 Linux 内核 TCP 代码中的严重漏洞,自 2016 年起便存在且可远程利用。
CVE-2026-98432 is described as a critical, remotely exploitable bug in the Linux kernel's TCP code that has existed since 2016. The source says Anthropic developed an exploit capable of gaining root access, while Debian issued a patch on October 16 and supported Ubuntu releases remained listed as vulnerable.CVE-2026-98432 被描述为 Linux 内核 TCP 代码中的严重漏洞,自 2016 年起便存在且可远程利用。资料称 Anthropic 已开发出可获取服务器 root 权限的漏洞利用程序;Debian 于 10 月 16 日发布了补丁,而受支持的 Ubuntu 版本仍被列为易受攻击。
A reported Claude Code bug exposed a user's email address to arbitrary third parties in the User-Agent header during web searches.有用户报告称,Claude Code 在进行网页搜索时,会通过 User-Agent 请求头将其邮箱地址暴露给任意第三方。
A reported Claude Code bug exposed a user's email address to arbitrary third parties in the User-Agent header during web searches. The user says this happened without permission and was followed by a large increase in scam emails. They argue that the behavior should be removed rather than controlled by an opt-in or opt-out setting.有用户报告称,Claude Code 在进行网页搜索时,会通过 User-Agent 请求头将其邮箱地址暴露给任意第三方。该用户表示,未经允许发生这一行为后,自己收到的诈骗邮件大幅增加。用户认为,这种行为不应通过加入或退出选项控制,而应从系统中彻底移除。
AI is accelerating both attacks and defenses in cybersecurity.AI 正在同时加速网络攻击和防御。
AI is accelerating both attacks and defenses in cybersecurity. That acceleration is exposing painful technical debt in fragile legacy systems built up over roughly four decades. The pressure from AI is therefore forcing organizations to address weaknesses that older systems had left unresolved.AI 正在同时加速网络攻击和防御。这样的加速暴露了脆弱遗留系统中积累约 40 年的严重技术债。AI 带来的压力因此迫使组织处理旧系统长期遗留下来的弱点。
Residential proxy traffic exited the White House network 236 times over 30 days, passing through 28 public proxy services.过去 30 天内,有住宅代理流量 236 次从白宫网络流出,并经过 28 个公共代理服务。
Residential proxy traffic exited the White House network 236 times over 30 days, passing through 28 public proxy services. The network, AS2650, is registered to the Executive Office of the President, which hosts the White House network. Residential proxies route customers’ traffic through everyday devices so it appears to originate from those devices’ networks.过去 30 天内,有住宅代理流量 236 次从白宫网络流出,并经过 28 个公共代理服务。该网络的 AS2650 注册于负责托管白宫网络的美国总统行政办公室。住宅代理会让客户的流量经由普通设备转发,使其看起来来自这些设备所在的网络。
Tesla CEO Elon Musk said Cybertruck production at the company's Texas gigafactory will wind down around mid-2027.特斯拉 CEO 埃隆·马斯克表示,公司位于得州的超级工厂将在 2027 年年中左右逐步停止生产 Cybertruck。
Tesla CEO Elon Musk said Cybertruck production at the company's Texas gigafactory will wind down around mid-2027. He said the move will make room for higher production of the newly launched Cybercab, central to Tesla's robotaxi plans. Cybertruck sales have remained weak, and it has sold in only a limited number of markets. It remains unclear whether production will move elsewhere or end entirely.特斯拉 CEO 埃隆·马斯克表示,公司位于得州的超级工厂将在 2027 年年中左右逐步停止生产 Cybertruck。他称,此举将为提高新款 Cybercab 的产量腾出空间,而 Cybercab 是特斯拉机器人出租车计划的核心车型。Cybertruck 销量长期疲软,目前仅在少数市场销售。该车型是转移到其他工厂生产还是彻底停产,目前尚不清楚。
The project brings Need for Speed Underground 2 online in a web browser.该项目让《极品飞车:地下狂飙 2》能够在浏览器中在线运行。
The project brings Need for Speed Underground 2 online in a web browser. No further implementation details or features are established here.该项目让《极品飞车:地下狂飙 2》能够在浏览器中在线运行。现有信息未说明其具体实现方式或功能细节。
libghostty-vt 1.0 has been released as a portable, high-performance C library for terminal emulation, extracted from Ghostty’s core.libghostty-vt 1.0 已发布,这是一个从 Ghostty 核心提取出的便携、高性能终端模拟 C 库。
libghostty-vt 1.0 has been released as a portable, high-performance C library for terminal emulation, extracted from Ghostty’s core. Both the library and its C ABI are now stable. The C API has grown from nonexistent to a complete terminal emulator used in production by millions of people across hundreds of free and commercial products.libghostty-vt 1.0 已发布,这是一个从 Ghostty 核心提取出的便携、高性能终端模拟 C 库。该库本身及其 C ABI 现已稳定。其 C API 已从无到有发展为完整的终端模拟器,目前已在数百款免费和商业产品中投入生产,服务数百万人。
The author argues that the experiments reading raw model weights and the experiments modifying Qwen use entirely different techniques.作者认为,读取模型原始权重的实验与修改 Qwen 的实验使用了完全不同的技术。
The author argues that the experiments reading raw model weights and the experiments modifying Qwen use entirely different techniques. They compare the presentation to a series of stage-magic tricks that appear to test competing explanations while relying on different mechanisms.作者认为,读取模型原始权重的实验与修改 Qwen 的实验使用了完全不同的技术。作者将这种呈现方式比作一连串舞台魔术:表面上是在检验相互竞争的解释,实际却依赖不同机制。文章分析了原始说法为何可能让人误解这些实验究竟证明了什么。
The title reports a backdoor in a Claude Code extension for VS Code distributed through the Cursor Marketplace.标题称,Cursor Marketplace 上分发的 Claude Code for VS Code 扩展中存在后门。
The title reports a backdoor in a Claude Code extension for VS Code distributed through the Cursor Marketplace. The supplied context provides no further reliable details.标题称,Cursor Marketplace 上分发的 Claude Code for VS Code 扩展中存在后门。提供的上下文没有更多可靠细节。
The author describes modifying the navigation system of a 2014 Renault Latitude to change the dashboard clock.作者介绍了如何修改一辆 2014 年雷诺 Latitude 的导航系统,以调整仪表盘时钟。
The author describes modifying the navigation system of a 2014 Renault Latitude to change the dashboard clock. The car apparently required a navigation-map SD card for access to the time-setting function, even though the card had been discarded because its maps were outdated. Without it, the display showed a prominent “No maps found!” message.作者介绍了如何修改一辆 2014 年雷诺 Latitude 的导航系统,以调整仪表盘时钟。该车似乎需要导航地图 SD 卡才能访问时间设置功能,但由于地图过时,这张卡此前已被丢弃。没有 SD 卡时,显示屏会显示醒目的“No maps found!”提示。
In the early 1990s, Parma farmer Claudio Grossi found an unidentified sack of seeds that likely belonged to his great-grandfather, who founded the farm around 1915.20 世纪 90 年代初,帕尔马农民克劳迪奥·格罗西发现了一袋无法辨认的种子,可能是其约于 1915 年创建农场的曾祖父留下的。
In the early 1990s, Parma farmer Claudio Grossi found an unidentified sack of seeds that likely belonged to his great-grandfather, who founded the farm around 1915. After he discarded the seeds in manure, several types of wheat unexpectedly sprouted. One variety grew about six feet tall and produced seven ears, unlike the single-ear commercial wheat Grossi knew.20 世纪 90 年代初,帕尔马农民克劳迪奥·格罗西发现了一袋无法辨认的种子,可能是其约于 1915 年创建农场的曾祖父留下的。他把种子倒进肥料堆后,几种小麦竟意外发芽。其中一个品种长到约六英尺高,并结出七个麦穗,不同于格罗西熟悉的单穗商业小麦。
The post claims that every Mac since macOS 13, released in 2022, includes two daemons that build a graph of the user’s relationships and activities.文章称,自 2022 年发布的 macOS 13 起,每台 Mac 都包含两个用于构建用户关系和活动图谱的后台进程。
The post claims that every Mac since macOS 13, released in 2022, includes two daemons that build a graph of the user’s relationships and activities. The graph is described as covering contacts, relationships, communications, locations, activities, and timing.文章称,自 2022 年发布的 macOS 13 起,每台 Mac 都包含两个用于构建用户关系和活动图谱的后台进程。该图谱据称涵盖联系人、彼此关系、通信、位置、活动及时间信息。文章还分析了这一系统的工作方式,并提供了查看个人设备上该图谱的工具。
Postgres 19 is late because many more bugs are being found during development.Postgres 19 的发布出现延期,原因是开发过程中发现的错误数量大幅增加。
Postgres 19 is late because many more bugs are being found during development. The title identifies the delay and attributes it to the increased discovery of defects.Postgres 19 的发布出现延期,原因是开发过程中发现的错误数量大幅增加。标题将这次延迟归因于缺陷发现增多。
Stringman is presented as an $800 open-source room-scale gripper robot.Stringman 被介绍为一款售价 800 美元的开源房间级抓取机器人。
Stringman is presented as an $800 open-source room-scale gripper robot. The available material provides an installation note about securing cleats to wall studs or columns, but no further technical specifications.Stringman 被介绍为一款售价 800 美元的开源房间级抓取机器人。现有内容只提到应将固定件用木螺钉固定在墙柱上,或用棘轮绑带固定在立柱上,未提供更多技术规格。
Subvocal is building smart glasses that decode silently mouthed words into text at speech speed.Subvocal 正在开发能将无声口型以接近语速转换为文字的智能眼镜。
Subvocal is building smart glasses that decode silently mouthed words into text at speech speed. The system can also work without opening the mouth, using radio-frequency sensing to detect movements inside the mouth, though the developers distinguish this from silently thinking words. The founders began the project after finding that voice tools were impractical in libraries, offices, and public spaces.Subvocal 正在开发能将无声口型以接近语速转换为文字的智能眼镜。系统甚至可以在不张嘴的情况下工作,通过射频感知口腔内部动作;开发者强调,这不同于单纯在脑中默念。创始人因发现语音工具不适合图书馆、办公室和公共场所,开始了这个项目。
The author bought an OpenStick for $10 including shipping, mainly because it was an unusual and appealing piece of hardware.作者花 10 美元(含运费)买了一台 OpenStick,主要是因为这款硬件新奇而有趣。
The author bought an OpenStick for $10 including shipping, mainly because it was an unusual and appealing piece of hardware. Although the device is less powerful than many modern watches, it can still support enjoyable self-hosting experiments. The project demonstrates how a very inexpensive, tiny computer can serve as an entry point to homelabbing.作者花 10 美元(含运费)买了一台 OpenStick,主要是因为这款硬件新奇而有趣。它的性能可能还不如许多现代智能手表,但仍然适合进行有趣的自托管实验。这个项目展示了廉价小设备如何成为入门家庭实验室的起点。
CVE-2026-98432 is a critical remote use-after-free vulnerability in Linux's IPv4 TCP handshake, reachable by an unauthenticated network peer.CVE-2026-98432 是 Linux IPv4 TCP 握手中的严重远程释放后使用漏洞,未经身份验证的网络对端即可触发。
CVE-2026-98432 is a critical remote use-after-free vulnerability in Linux's IPv4 TCP handshake, reachable by an unauthenticated network peer. Stable kernels fixed it in October 2026, with patches for versions including 5.10.272 through 7.3, but the flaw dates back to Linux 4.4. At the reported time, Ubuntu still marked every release as vulnerable, while Red Hat listed RHEL 8, 9, and 10 as affected without a fix.CVE-2026-98432 是 Linux IPv4 TCP 握手中的严重远程释放后使用漏洞,未经身份验证的网络对端即可触发。稳定版内核已于 2026 年 10 月修复,补丁覆盖 5.10.272 至 7.3 等版本,但该缺陷可追溯至 Linux 4.4。报道所述时间点,Ubuntu 仍将所有版本标为易受攻击,Red Hat 也列出受影响的 RHEL 8、9 和 10,尚未提供修复。
The author had used Claude Code for more than a year, beginning when it became available through the Claude Pro subscription.作者从 Claude Code 随 Claude Pro 订阅推出后就开始使用,至今已有一年多。
The author had used Claude Code for more than a year, beginning when it became available through the Claude Pro subscription. Despite considering its release an important moment for software development, the author accumulated enough small frustrations to stop using it as the main coding harness. Their development setup is now being migrated to Pi.作者从 Claude Code 随 Claude Pro 订阅推出后就开始使用,至今已有一年多。尽管作者认为它的发布可能会成为软件开发史上的重要节点,但长期积累的许多小问题最终让作者不再把它作为主要编程工具。作者目前正将开发环境迁移到 Pi。
BUI is a GPU-accelerated Blender user interface decoupled from Blender's original codebase.BUI 是一个与 Blender 原始代码库解耦的 GPU 加速用户界面。
BUI is a GPU-accelerated Blender user interface decoupled from Blender's original codebase. It supports Rust, Python, JavaScript, React, C, and C++. The project describes its performance as twice that of GPUI.BUI 是一个与 Blender 原始代码库解耦的 GPU 加速用户界面。它支持 Rust、Python、JavaScript、React、C 和 C++。项目方称其性能是 GPUI 的两倍。
A user is looking for a privacy-respecting television with a 60–80-inch 4K display running at least 60 Hz.一名用户想寻找重视隐私的电视,要求尺寸在 60 至 80 英寸之间、支持 4K 且刷新率至少为 60Hz。
A user is looking for a privacy-respecting television with a 60–80-inch 4K display running at least 60 Hz. They want the device to act only as a display for an Apple TV rather than use smart-TV apps, and are asking which models pair well with that setup.一名用户想寻找重视隐私的电视,要求尺寸在 60 至 80 英寸之间、支持 4K 且刷新率至少为 60Hz。他希望电视只作为 Apple TV 的显示器,而不运行智能电视原生应用,并询问哪些型号适合这种配置。
The author reports that Qwen 3.8 27B running on a laptop can interpret raw neural-network weights.作者称,在笔记本上运行的 Qwen 3.8 27B 也能解读神经网络的原始权重。
The author reports that Qwen 3.8 27B running on a laptop can interpret raw neural-network weights. This follows research describing “weight-space programming,” in which language models analyze and modify model weights represented as floating-point numbers. The earlier experiments highlighted frontier-model capabilities, including those attributed to GPT-6 Astra.作者称,在笔记本上运行的 Qwen 3.8 27B 也能解读神经网络的原始权重。这延续了“权重空间编程”的研究:语言模型可以分析并修改由浮点数表示的模型权重。此前的实验展示了前沿模型的这类能力,包括研究归于 GPT-6 Astra 的能力。
For personal Google AI subscriptions, saved Gemini chats can be used to improve Google services and train generative AI models when Keep Activity is enabled.对于个人 Google AI 订阅账户,在启用“保留活动记录”时,保存的 Gemini 聊天记录可用于改进 Google 服务和训练生成式 AI 模型。
For personal Google AI subscriptions, saved Gemini chats can be used to improve Google services and train generative AI models when Keep Activity is enabled. A subset of chats may be reviewed by human reviewers and can be retained for up to three years after review, even if the user deletes activity. Users are warned not to enter confidential information, and opting out requires disabling chat history or using temporary chats.对于个人 Google AI 订阅账户,在启用“保留活动记录”时,保存的 Gemini 聊天记录可用于改进 Google 服务和训练生成式 AI 模型。部分聊天可能由人工审核,审核过的内容即使用户删除活动记录,也可能保留最长三年。用户被提醒不要输入机密信息;若要退出,需要关闭聊天记录或使用临时聊天。
The restrooms along Norway’s Scenic Routes are presented as architectural landmarks designed to work with the surrounding landscape.挪威风景公路沿线的卫生间被视为与周围自然环境协调的建筑地标。
The restrooms along Norway’s Scenic Routes are presented as architectural landmarks designed to work with the surrounding landscape. Most were designed by established or emerging Norwegian architecture firms. Examples include Snøhetta’s small Eggum stop in Lofoten and Reiulf Ramstad’s facility at Selvika.挪威风景公路沿线的卫生间被视为与周围自然环境协调的建筑地标。大多数设施由知名或新兴的挪威建筑事务所设计。例如,罗弗敦群岛 Eggum 休息点由 Snøhetta 设计,Selvika 的设施则由 Reiulf Ramstad 设计。
Goldman Sachs has restricted Anthropic's Claude Code and Microsoft's GitHub to its engineers, removing access for most traders and bankers.高盛已将 Anthropic 的 Claude Code 和微软 GitHub 的使用权限限制在工程师范围内,大多数交易员和银行家因此失去访问权限。
Goldman Sachs has restricted Anthropic's Claude Code and Microsoft's GitHub to its engineers, removing access for most traders and bankers. Those employees had used the tools to track client flows and build market models, while the bank had already barred Hong Kong bankers from Claude in April. The guidance followed a review of productivity and risk concerns and came weeks before Goldman was due to act as a lead underwriter for Anthropic's IPO.高盛已将 Anthropic 的 Claude Code 和微软 GitHub 的使用权限限制在工程师范围内,大多数交易员和银行家因此失去访问权限。此前,这些员工曾用相关工具追踪客户资金流向并构建市场模型;高盛也已于 4 月禁止香港银行家使用 Claude。这项指导是在审查生产力和风险问题后出台的,时间上距离高盛担任 Anthropic IPO 主承销商仅数周。
A remote code execution vulnerability has been reported in Microsoft Paint.微软画图工具被报告存在远程代码执行漏洞。
A remote code execution vulnerability has been reported in Microsoft Paint. The issue could allow code to run remotely when exploited.微软画图工具被报告存在远程代码执行漏洞。攻击者利用该问题后,可能在远程执行代码。
The author describes discovering that the long-running Teapuppies fansite had disappeared and its domain was serving online-gambling advertisements.作者发现,一个长期存在的 Teapuppies 粉丝网站已经消失,原域名如今被用于投放在线赌博广告。
The author describes discovering that the long-running Teapuppies fansite had disappeared and its domain was serving online-gambling advertisements. The site had been online since 2004 and largely unchanged since 2005, making its loss an example of how fragile old web pages can be.作者发现,一个长期存在的 Teapuppies 粉丝网站已经消失,原域名如今被用于投放在线赌博广告。该网站自 2004 年上线、2005 年后基本未变,这次消失体现了旧网页保存状况的脆弱性。
Kagi has opened web search to anyone without requiring an account for the first 50 searches.Kagi 现在允许任何人使用网页搜索,前 50 次搜索无需创建账户。
Kagi has opened web search to anyone without requiring an account for the first 50 searches. Registering adds another 100 searches. The paid search engine says it does not run ads and is relying on community recommendations and word of mouth.Kagi 现在允许任何人使用网页搜索,前 50 次搜索无需创建账户。注册后还可额外获得 100 次搜索。这个付费搜索引擎表示自己不投放广告,而是依靠社区推荐和口碑传播。
A compatibility report examines whether the ASN Bank Android app blocks non-Google Android systems, including GrapheneOS.一份兼容性报告调查 ASN Bank 的 Android 应用是否屏蔽包括 GrapheneOS 在内的非 Google Android 系统。
A compatibility report examines whether the ASN Bank Android app blocks non-Google Android systems, including GrapheneOS. The report identifies app version 2.40.2 in the Netherlands. One user reported that the app appeared to work on a stock operating system.一份兼容性报告调查 ASN Bank 的 Android 应用是否屏蔽包括 GrapheneOS 在内的非 Google Android 系统。报告涉及荷兰地区的 2.40.2 版本。一名用户表示,该应用在原生操作系统上看起来可以正常运行。
Moonshot AI has released the weights for Kimi K3.1 Pro, an open-weight, natively multimodal agentic model with 2.8 trillion parameters and a one-million-token context window.月之暗面发布了 Kimi K3.1 Pro 的模型权重。
Moonshot AI has released the weights for Kimi K3.1 Pro, an open-weight, natively multimodal agentic model with 2.8 trillion parameters and a one-million-token context window. It was post-trained from Kimi K3 using large-scale joint reinforcement learning across hundreds of domains. The model replaces Gated MLA with Gated DeepSeek Sparse Attention for more efficient long-context modeling.月之暗面发布了 Kimi K3.1 Pro 的模型权重。这是一款开放权重、原生多模态的智能体模型,拥有 2.8 万亿参数和 100 万 token 的上下文窗口。该模型基于 Kimi K3,并通过覆盖数百个领域的大规模联合强化学习进行后训练;同时以 Gated DeepSeek Sparse Attention 替代 Gated MLA,以提高长上下文建模效率。
Alpine Linux advises users running sshd to update openssh immediately and restart the daemon.Alpine Linux 建议运行 sshd 的用户立即更新 openssh 并重启守护进程。
Alpine Linux advises users running sshd to update openssh immediately and restart the daemon. Updating the package alone does not replace an already running sshd process. Until updating is possible, disabling password logins with PasswordAuthentication no and restarting sshd closes the hole.Alpine Linux 建议运行 sshd 的用户立即更新 openssh 并重启守护进程。仅更新软件包并不会替换已经运行的 sshd 进程。在完成更新前,可将配置设为 PasswordAuthentication no 并重启 sshd,以关闭该漏洞。
RTX Spark laptops run Linux with GPU acceleration through Nvidia’s open kernel driver, but their GPU clocks drop to roughly 780 MHz when plugged in, instead of exceeding 2 GHz on Windows.RTX Spark 笔记本可以在 Linux 上运行,并通过英伟达开源内核驱动实现 GPU 加速,但接通电源后 GPU 频率会被限制在约 780MHz,而 Windows 上可超过 2GHz。
RTX Spark laptops run Linux with GPU acceleration through Nvidia’s open kernel driver, but their GPU clocks drop to roughly 780 MHz when plugged in, instead of exceeding 2 GHz on Windows. On a 120-billion-parameter model, performance falls from about 65 tokens per second to fewer than 14. The problem is tied to power-management firmware that Linux currently cannot read.RTX Spark 笔记本可以在 Linux 上运行,并通过英伟达开源内核驱动实现 GPU 加速,但接通电源后 GPU 频率会被限制在约 780MHz,而 Windows 上可超过 2GHz。在一个 1200 亿参数模型上,速度因此从约每秒 65 个 token 降至不足 14 个。问题与 Linux 目前无法读取的电源管理固件有关。
Ubuntu has not released patched kernels for CVE-2026-98432, a vulnerability rated CVSS 9.8.Ubuntu 尚未为 CVSS 评分 9.8 的 CVE-2026-98432 发布修复内核。
Ubuntu has not released patched kernels for CVE-2026-98432, a vulnerability rated CVSS 9.8. Ubuntu 22.04, 24.04, and 26.04 LTS therefore lack an official kernel package that fixes the issue. Internet-facing and multi-tenant systems relying on those kernels remain particularly exposed until updates become available.Ubuntu 尚未为 CVSS 评分 9.8 的 CVE-2026-98432 发布修复内核。Ubuntu 22.04、24.04 和 26.04 LTS 因此仍没有能够修复该问题的官方内核包。在更新发布前,依赖这些内核的面向互联网和多租户系统尤其处于暴露状态。
According to Binance, North Korean cryptocurrency thieves tricked a software engineer into downloading malware during a fake job interview.币安称,朝鲜加密货币窃贼在一次虚假求职面试中诱骗一名软件工程师下载恶意软件。
According to Binance, North Korean cryptocurrency thieves tricked a software engineer into downloading malware during a fake job interview. The exchange said the October 29 breach involved social engineering, malware, compromised internal AI tooling, endpoint takeover, and an advanced transaction-authorization method. Binance characterized it as a combination of techniques rather than a single isolated phishing incident.币安称,朝鲜加密货币窃贼在一次虚假求职面试中诱骗一名软件工程师下载恶意软件。该交易所表示,10 月 29 日的入侵涉及社会工程、恶意软件、遭入侵的内部 AI 工具、终端接管以及一种先进的交易授权方法。币安将其描述为多种手段共同造成的事件,而非单一的钓鱼攻击。
Hackers siphoned roughly $6.3 billion in cryptocurrency from wallets belonging to Binance across multiple blockchains in about 20 minutes.黑客在约 20 分钟内从币安所属的钱包中跨多个区块链转走了约 63 亿美元的加密货币。
Hackers siphoned roughly $6.3 billion in cryptocurrency from wallets belonging to Binance across multiple blockchains in about 20 minutes. Binance suspended deposits and withdrawals and began investigating, while the exchange and other crypto firms said they had frozen around half of the stolen funds. Much of the remaining cryptocurrency may not be recoverable through freezing.黑客在约 20 分钟内从币安所属的钱包中跨多个区块链转走了约 63 亿美元的加密货币。币安暂停了存款和提现并展开调查;该交易所及其他加密行业公司称,已冻结约一半被盗资金。剩余的大部分加密货币可能无法通过冻结追回。
AMD expects to outsell Intel in the fourth quarter for the first time in its history.AMD 预计将在第四季度历史上首次超过英特尔。
AMD expects to outsell Intel in the fourth quarter for the first time in its history. Record quarterly revenue is being driven by the AI boom, while its client and gaming businesses are facing double-digit declines heading into 2027. The forecast indicates a widening contrast between AMD’s AI growth and its weaker traditional segments.AMD 预计将在第四季度历史上首次超过英特尔。AI 热潮推动其季度营收创下纪录,但面向客户的业务和游戏业务在进入 2027 年时正面临两位数下滑。这一预测显示,AMD 的 AI 业务增长与传统业务疲软之间的差距正在扩大。
Britain plans to make salary disclosure mandatory in job advertisements.英国计划强制招聘广告披露薪资信息。
Britain plans to make salary disclosure mandatory in job advertisements. Employers would have to include pay information when advertising open positions.英国计划强制招聘广告披露薪资信息。雇主发布职位空缺时将必须列出薪酬信息。
The title alleges that Lenovo’s fingerprint-protected USB device is not an encrypted drive, but an unencrypted SD card placed inside a casing.标题指称,联想这款受指纹保护的 USB 设备并非加密硬盘,而只是装在外壳中的未加密 SD 卡。
The title alleges that Lenovo’s fingerprint-protected USB device is not an encrypted drive, but an unencrypted SD card placed inside a casing. It identifies a potential mismatch between the product’s security description and its underlying storage design.标题指称,联想这款受指纹保护的 USB 设备并非加密硬盘,而只是装在外壳中的未加密 SD 卡。这表明产品的安全描述可能与其底层存储设计不一致。
Tests of Jev and GPT-6 Luna as judges for identity resolution found out-of-the-box accuracy ranging from 88.8% to 97.3%.对 Jev 和 GPT-6 Luna 作为身份解析评判器的测试显示,它们开箱即用的准确率为 88.8% 至 97.3%。
Tests of Jev and GPT-6 Luna as judges for identity resolution found out-of-the-box accuracy ranging from 88.8% to 97.3%. A hybrid that combines Jev with Claude Opus matched frontier-level accuracy at more than four times lower cost. Identity resolution maps data records to unique real-world people.对 Jev 和 GPT-6 Luna 作为身份解析评判器的测试显示,它们开箱即用的准确率为 88.8% 至 97.3%。将 Jev 与 Claude Opus 结合的混合方案达到了前沿模型水平的准确率,而成本降低了四倍以上。身份解析的任务是将数据记录映射到现实世界中唯一的个人。
Salon has announced that it will cease publication after 31 years.Salon 宣布,在出版 31 年后将停止运营。
Salon has announced that it will cease publication after 31 years. The publication described its history as an attempt to explore what journalism could become while the internet was still new.Salon 宣布,在出版 31 年后将停止运营。该出版物将自己的历史描述为一次探索新闻业可能形态的尝试,起步于互联网仍处于早期阶段之时。
Research found that Apple restricts predictive-text completion for more than 100,000 words and phrases.研究发现,苹果限制了超过 10 万个词语和短语的预测文本补全功能。
Research found that Apple restricts predictive-text completion for more than 100,000 words and phrases. The restrictions affect swipe typing, in which users glide a finger across the keyboard instead of tapping each character. As a result, an iPhone may refuse to complete some unusual or sensitive phrases even when users try to enter them by gesture.研究发现,苹果限制了超过 10 万个词语和短语的预测文本补全功能。这些限制会影响滑行输入,即用户用手指划过键盘,而不是逐字点击。结果是,即使用户尝试通过滑行输入,iPhone 也可能不会补全某些不寻常或敏感的短语。
OccupySaaS is an Apache-2.0 open-source effort to build self-hosted, drop-in replacements for major SaaS products without per-seat fees.OccupySaaS 是一个采用 Apache-2.0 许可的开源项目,旨在构建可自托管、无需按席位收费的大型 SaaS 替代品。
OccupySaaS is an Apache-2.0 open-source effort to build self-hosted, drop-in replacements for major SaaS products without per-seat fees. Its first projects are ZeroAuthy, positioned as an Auth0 replacement, and SalesCompel, positioned as a Salesforce replacement.OccupySaaS 是一个采用 Apache-2.0 许可的开源项目,旨在构建可自托管、无需按席位收费的大型 SaaS 替代品。它的首批项目包括定位为 Auth0 替代品的 ZeroAuthy,以及定位为 Salesforce 替代品的 SalesCompel。
The item is a message to Nvidia employees from CEO Jensen Huang concerning an affair.这是一封英伟达 CEO 黄仁勋就婚外情相关事件写给员工的信。
The item is a message to Nvidia employees from CEO Jensen Huang concerning an affair. The available material provides links to related reporting but does not include the contents of Huang’s letter or describe any resulting action.这是一封英伟达 CEO 黄仁勋就婚外情相关事件写给员工的信。现有材料仅提供了相关报道的链接,没有包含信件内容,也没有说明后续采取了什么行动。
Fujifilm plans to invest about 10 billion yen, or roughly $63 million, to increase magnetic-tape production capacity fivefold.富士胶片计划投资约 100 亿日元,即约 6300 万美元,将磁带产能提高五倍。
Fujifilm plans to invest about 10 billion yen, or roughly $63 million, to increase magnetic-tape production capacity fivefold. The company attributes the rising demand for archiving to generative AI and what it calls exponential global data growth. The fivefold increase will be measured against fiscal 2025 production levels.富士胶片计划投资约 100 亿日元,即约 6300 万美元,将磁带产能提高五倍。公司将归档需求的增长归因于生成式 AI 以及其所称的全球数据指数级增长。五倍增幅将以 2025 财年的产量为基准。
An Ask HN thread asks how people use AI agents in their daily work.一则 Ask HN 帖子询问人们目前如何在日常工作中使用 AI agent。
An Ask HN thread asks how people use AI agents in their daily work. It solicits details about paid subscriptions, external or locally hosted models, coding harnesses, orchestration tools, and broader agent platforms. The discussion aims to gauge how individual workflows have changed over the past year.一则 Ask HN 帖子询问人们目前如何在日常工作中使用 AI agent。发帖者希望了解大家购买的订阅、外部或本地部署的模型、编程工具链、编排工具以及更广泛的 agent 平台。讨论旨在了解个人工作流在过去一年中的变化。
Mark Cuban says AI-generated scam ads have used his face to promote a diabetes cure he never sold.马克·库班表示,利用其肖像的 AI 生成诈骗广告正在推广他从未销售过的糖尿病疗法。
Mark Cuban says AI-generated scam ads have used his face to promote a diabetes cure he never sold. He says people contact him every week believing they bought the supposed cure for $294. The claims come amid allegations that Facebook knowingly runs such deepfake scam advertising.马克·库班表示,利用其肖像的 AI 生成诈骗广告正在推广他从未销售过的糖尿病疗法。他说,每周都有人联系他,以为自己花 294 美元买到了这种所谓疗法。这些说法出现之际,Facebook 被指控明知相关深度伪造诈骗广告存在却仍在投放。
Safari 27.3 beta introduces Adaptive Dark Mode.Safari 27.3 beta 引入了自适应深色模式。
Safari 27.3 beta introduces Adaptive Dark Mode. When enabled by a user, Safari adapts light-themed websites and presents them in a dark appearance.Safari 27.3 beta 引入了自适应深色模式。用户启用该功能后,Safari 会调整浅色网站,将其以深色外观呈现。
A user says Anthropic blocked their account after flagging work involving reinforcement fine-tuning, supervised fine-tuning, or training larger models as “training a competitor.” The user says Claude only helped submit jobs and analyze results, while the rollouts came from an open-source model.一名用户称,Anthropic 将其涉及强化微调、监督微调或训练更大模型的工作标记为“训练竞争对手”,并屏蔽了其账户。
A user says Anthropic blocked their account after flagging work involving reinforcement fine-tuning, supervised fine-tuning, or training larger models as “training a competitor.” The user says Claude only helped submit jobs and analyze results, while the rollouts came from an open-source model. They also report being downgraded from Opus 5.5.一名用户称,Anthropic 将其涉及强化微调、监督微调或训练更大模型的工作标记为“训练竞争对手”,并屏蔽了其账户。该用户表示,Claude 只帮助提交任务和分析结果,实际生成过程使用的是开源模型。用户还称自己的账户被降级,无法继续使用 Opus 5.5。
OpenAI built an autonomous pipeline that discovered, validated, and generated patches for 948 local privilege-escalation vulnerabilities in the Linux kernel.OpenAI 构建了一个自主流水线,用于发现、验证并为 Linux 内核中的 948 个本地提权漏洞生成补丁。
OpenAI built an autonomous pipeline that discovered, validated, and generated patches for 948 local privilege-escalation vulnerabilities in the Linux kernel. In a single run using its cyber model, the system validated the vulnerabilities and produced patches at an API cost of $1.2 million. The work suggests that some security investigations can increasingly be handled as a systematic automated process.OpenAI 构建了一个自主流水线,用于发现、验证并为 Linux 内核中的 948 个本地提权漏洞生成补丁。在一次运行中,该系统借助其网络安全模型完成了漏洞验证和补丁生成,API 成本为 120 万美元。这项工作表明,部分安全研究正逐渐可以作为系统化的自动流程来执行。
The discussion challenges the web development community's tendency to describe frameworks and tools with unsupported claims such as “blazing fast” or “supercharge your development.” It argues that many choices have historically been justified by developer experience and intuition rather than evidence.文章质疑 Web 开发社区习惯用“快如闪电”或“显著提升开发效率”等缺乏证据的说法描述框架和工具。
The discussion challenges the web development community's tendency to describe frameworks and tools with unsupported claims such as “blazing fast” or “supercharge your development.” It argues that many choices have historically been justified by developer experience and intuition rather than evidence. The focus is on examining the measurable costs of JavaScript frameworks in an era of AI-assisted development.文章质疑 Web 开发社区习惯用“快如闪电”或“显著提升开发效率”等缺乏证据的说法描述框架和工具。它指出,许多技术选择长期以来更多是依据开发者体验和直觉,而非实证来论证的。文章将关注点放在 AI 辅助开发时代 JavaScript 框架可衡量的成本上。
Plymouth canceled its contract with Flock after cameras wrongly flagged a couple who were later confronted by police.Plymouth 市因 Flock 摄像头错误标记一对夫妇、导致警方随后对其采取行动而取消了合同。
Plymouth canceled its contract with Flock after cameras wrongly flagged a couple who were later confronted by police. The city suspended user accounts and formally notified Flock, but the cameras remained installed a month later and the company had not said when it would remove them. Police Chief Eric Fadden said the department had stopped using the Flock ALPR system after the city council’s decision.Plymouth 市因 Flock 摄像头错误标记一对夫妇、导致警方随后对其采取行动而取消了合同。市政府暂停了用户账户并正式通知 Flock,但一个月后摄像头仍未拆除,公司也没有说明何时移除。警察局长 Eric Fadden 表示,市议会作出决定后,部门已停止使用 Flock 的自动车牌识别系统。
HackerTV turns YouTube videos frequently recommended in Hacker News comments into 24/7 topic channels.HackerTV 将 Hacker News 评论中经常被推荐的 YouTube 视频整理成全天候主题频道。
HackerTV turns YouTube videos frequently recommended in Hacker News comments into 24/7 topic channels. It includes channels such as DEVTV, SOLDER, RETRO, and ORBIT, along with an HN Prime channel containing the top 400 videos. The project was built in a few hours with Claude Code and also includes a music channel and a late-night channel.HackerTV 将 Hacker News 评论中经常被推荐的 YouTube 视频整理成全天候主题频道。它包含 DEVTV、SOLDER、RETRO 和 ORBIT 等频道,以及收录前 400 个视频的 HN Prime 频道。项目用 Claude Code 在几小时内完成,还提供音乐频道和深夜频道。
Declarative dragging proposes that interface authors define which future model states an element may reach by dragging.“声明式拖拽”提出,界面作者可以通过指定元素拖动后允许到达的模型状态来定义拖拽行为。
Declarative dragging proposes that interface authors define which future model states an element may reach by dragging. This makes a diagram element draggable by describing its allowed outcomes rather than implementing dragging behavior directly. The work was named the Best Paper at UIST 2026.“声明式拖拽”提出,界面作者可以通过指定元素拖动后允许到达的模型状态来定义拖拽行为。这样,创建可拖动的图表元素不必直接实现拖拽逻辑,而是描述其允许的结果。该研究获得了 UIST 2026 最佳论文奖。
Motorola has introduced the motorola signature 27 as its most advanced flagship smartphone to date.摩托罗拉推出了迄今最先进的旗舰手机 motorola signature 27。
Motorola has introduced the motorola signature 27 as its most advanced flagship smartphone to date. The phone features a DXOMARK Gold Label camera system, an Extreme AMOLED Duo Tandem display, a Snapdragon 8 Elite Extreme Gen 6 platform, and Bang & Olufsen audio. Motorola says it includes the company's most advanced camera system and next-generation AI capabilities.摩托罗拉推出了迄今最先进的旗舰手机 motorola signature 27。该机配备获徕卡?
Nvidia’s board has initiated an independent investigation after reports that founder and CEO Jensen Huang was targeted by a foreign intelligence effort.英伟达董事会在相关报道指出创始人兼 CEO 黄仁勋成为外国情报行动目标后,启动了独立调查。
Nvidia’s board has initiated an independent investigation after reports that founder and CEO Jensen Huang was targeted by a foreign intelligence effort. A committee of independent directors will oversee the investigation with assistance from outside independent counsel. The available statement does not provide further findings.英伟达董事会在相关报道指出创始人兼 CEO 黄仁勋成为外国情报行动目标后,启动了独立调查。调查将由独立董事组成的委员会负责,并由外部独立法律顾问协助。现有声明没有提供进一步调查结论。
The piece reflects on Patrick Stewart’s portrayal of Captain Jean-Luc Picard and the leadership ideals associated with the character.文章回顾了帕特里克·斯图尔特饰演让-卢克·皮卡尔舰长时所体现的领导理想。
The piece reflects on Patrick Stewart’s portrayal of Captain Jean-Luc Picard and the leadership ideals associated with the character. It notes that Stewart died on Wednesday and places Picard’s post-Cold War vision in the historical context of the late 1980s and early 1990s.文章回顾了帕特里克·斯图尔特饰演让-卢克·皮卡尔舰长时所体现的领导理想。文章提到斯图尔特于周三去世,并将皮卡尔所代表的后冷战世界观置于 20 世纪 80 年代末和 90 年代初的历史背景中。
Discord's new Rust client loads the friends list before the existing Electron app gets past its loading screen and skips the Starting screen.Discord 的新 Rust 客户端在现有 Electron 客户端还没通过加载界面时就已载入好友列表,并跳过了 Starting 界面。
Discord's new Rust client loads the friends list before the existing Electron app gets past its loading screen and skips the Starting screen. In a cold-start test, the Rust-based client used 290.2 MB of RAM with two accounts signed in. Discord employees confirmed that both applications were launched from a cold start, while a video showing the comparison received 2.3 million views at the time reported.Discord 的新 Rust 客户端在现有 Electron 客户端还没通过加载界面时就已载入好友列表,并跳过了 Starting 界面。在冷启动测试中,Rust 客户端同时登录两个账号时占用 290.2 MB 内存。Discord 员工确认两款应用都是冷启动,该对比视频当时已获得 230 万次观看。
Memdoor implements a team-solving workflow in which five small models act as a planner, ideators, solver, manager, and repair worker.Memdoor 实现了一种协作解题流程,由五个小模型分别担任规划者、创意提出者、求解器、管理者和修复者。
Memdoor implements a team-solving workflow in which five small models act as a planner, ideators, solver, manager, and repair worker. Each model starts with a fresh context and communicates only through files, while an external script determines whether solutions pass. The authors reproduced the method using Qwen3.8-27B, eight YAML files, their prompts, and a 128K response limit.Memdoor 实现了一种协作解题流程,由五个小模型分别担任规划者、创意提出者、求解器、管理者和修复者。每个模型都从全新的上下文开始,仅通过磁盘文件通信,并由外部脚本判断解答是否通过。作者使用 Qwen3.8-27B、八个 YAML 文件、自定义提示词和 128K 回复上限复现了这一方法。
Discord is moving its applications fully to Rust and abandoning Electron for the desktop client.Discord 正将其应用全面迁移到 Rust,并放弃桌面客户端使用的 Electron。
Discord is moving its applications fully to Rust and abandoning Electron for the desktop client. The company says the change will remove Electron, while mobile keeps a native UI and shares application code with desktop through Rust. Discord has used Rust in infrastructure, voice technology, and parts of its client since 2018.Discord 正将其应用全面迁移到 Rust,并放弃桌面客户端使用的 Electron。公司表示,此举将移除 Electron;移动端继续使用原生 UI,并通过 Rust 与桌面端共享应用代码。Discord 自 2018 年起已在基础设施、语音技术和部分客户端中使用 Rust。
Researchers ran AI agent swarms in Minecraft and Mindustry to study emergent coordination at scale.在 Minecraft 与 Mindustry 中测试AI代理群,研究群体协作与尺度效应。
Researchers ran AI agent swarms in Minecraft and Mindustry to study emergent coordination at scale. The experiments illustrate how simple agents, when put in shared environments, can exhibit complex group behaviors, informing the design of scalable multi-agent systems and game AI safety. The work provides hands-on demos for researchers building autonomous agents in synthetic worlds.在 Minecraft 与 Mindustry 中测试AI代理群,研究群体协作与尺度效应。实验表明,简单代理在共享环境中可涌现出复杂的协同行为,为多智能体系统和游戏AI安全提供直观范例。
Linux 7.3 has been released, and Asahi Linux continues upstreaming support for Apple Silicon.Linux 7.3 已发布,Asahi Linux 仍在持续将 Apple Silicon 支持代码合并到上游。
Linux 7.3 has been released, and Asahi Linux continues upstreaming support for Apple Silicon. Two major drivers, macaudio and avd, joined the upstream work this cycle. Getting macaudio ready required substantial ALSA changes and modifications to Apple’s speaker and headset amplifier support, with the driver and Devicetree patches totaling about 3,100 lines.Linux 7.3 已发布,Asahi Linux 仍在持续将 Apple Silicon 支持代码合并到上游。本周期有两个重要驱动加入上游工作:macaudio 和 avd。为了准备 macaudio,项目对 ALSA 进行了大量修改,也改造了 Apple 扬声器和耳机放大器的支持;相关驱动与设备树补丁共约 3100 行。
The project describes what is required to run Macs in a data center, including custom MacBook racks, macOS virtualization with Apple’s Virtualization framework, and distributing large virtual-machine images across a fleet.该项目介绍了在数据中心运行 Mac 所需的技术,包括定制 MacBook 机架、使用 Apple Virtualization 框架虚拟化 macOS,以及在设备集群中分发大型虚拟机镜像。
The project describes what is required to run Macs in a data center, including custom MacBook racks, macOS virtualization with Apple’s Virtualization framework, and distributing large virtual-machine images across a fleet. The author says hundreds of MacBook Pros were unpacked for new server racks. The work explores how Mac infrastructure can be deployed at scale.该项目介绍了在数据中心运行 Mac 所需的技术,包括定制 MacBook 机架、使用 Apple Virtualization 框架虚拟化 macOS,以及在设备集群中分发大型虚拟机镜像。作者表示,已拆箱数百台 MacBook Pro,用于搭建新的服务器机架。项目探索了如何大规模部署 Mac 基础设施。
The available context consists only of links to related coverage and does not provide reliable details about the reported incidents.现有资料只有相关报道链接,未提供足够可靠的事件细节。
The available context consists only of links to related coverage and does not provide reliable details about the reported incidents.现有资料只有相关报道链接,未提供足够可靠的事件细节。
Llama-modchip is a collection of ggml and llama.cpp kernel patches, quantization transcoders, and runtime tools for running sparse MoE models entirely in system RAM.Llama-modchip 是一组 ggml 和 llama.cpp 内核补丁、量化转换器及运行时工具,旨在让稀疏 MoE 模型完全运行于系统内存中。
Llama-modchip is a collection of ggml and llama.cpp kernel patches, quantization transcoders, and runtime tools for running sparse MoE models entirely in system RAM. Its project page claims that a 284B DeepSeek-V4-Flash model can run on a 64GB CPU-only system at a 256K context length, with reported prompt-processing speeds of 44.4 tok/s and generation speeds of 9.1–14.7 tok/s. It supports x86_64 and ARM64 processors and also targets 30B MoE models.Llama-modchip 是一组 ggml 和 llama.cpp 内核补丁、量化转换器及运行时工具,旨在让稀疏 MoE 模型完全运行于系统内存中。项目页面称,284B 参数的 DeepSeek-V4-Flash 可在仅配备 64GB 内存、没有 GPU 的 CPU 系统上运行,支持 256K 上下文,提示处理速度为 44.4 token/s,生成速度为 9.1–14.7 token/s。该项目支持 x86_64 和 ARM64 处理器,也面向 30B MoE 模型。
The US Federal Aviation Administration plans to phase out floppy disks from its air-traffic-control systems by 2028.美国联邦航空管理局计划在 2028 年前逐步淘汰空中交通管制系统中的软盘。
The US Federal Aviation Administration plans to phase out floppy disks from its air-traffic-control systems by 2028. The move will replace a legacy storage technology still used in parts of air-traffic operations.美国联邦航空管理局计划在 2028 年前逐步淘汰空中交通管制系统中的软盘。这一举措将替换部分航空交通运营中仍在使用的老旧存储技术。
This interactive CSS animation lets users add walkers with a button or with `__walker.add()`.这个交互式 CSS 动画允许用户通过按钮或 `__walker.add()` 添加行走者。
This interactive CSS animation lets users add walkers with a button or with `__walker.add()`. Walkers can be dragged, carried, or flicked through the scene. Resizing the window makes them ride the surface beneath them, while narrowing the window wraps them onto new rows and pushes walkers standing on those rows.这个交互式 CSS 动画允许用户通过按钮或 `__walker.add()` 添加行走者。行走者可以被拖动、携带或弹射。调整窗口大小时,它们会沿脚下的表面移动;窗口变窄后,行走者会换到新行,并推动站在其上的其他行走者。
Matt Scherer argues that the SEC should block a potential Anthropic public offering.Matt Scherer 主张 SEC 应阻止 Anthropic 可能进行的 IPO。
Matt Scherer argues that the SEC should block a potential Anthropic public offering. He says workers, retail investors, and retirees could bear the losses if the company's bills come due, and identifies the argument with a report about protecting people and the economy from an AI-bubble collapse.Matt Scherer 主张 SEC 应阻止 Anthropic 可能进行的 IPO。他认为,如果公司的账单最终无法承担,员工、散户投资者和退休人员可能会承受损失;这一观点与一份讨论如何在 AI 泡沫破裂时保护公众和经济的报告相关。
MaxxBench changed the requirements of familiar engineering tasks in a coding benchmark to test whether models would follow the new instructions.MaxxBench 修改了编码基准中熟悉工程任务的要求,用来测试模型是否会遵循新指令。
MaxxBench changed the requirements of familiar engineering tasks in a coding benchmark to test whether models would follow the new instructions. The researchers found cases where a model understood and correctly implemented the changed requirement, then removed its own correct code after another model supplied the benchmark’s original answer. The results expose a gap between recognizing a requirement and consistently acting on it.MaxxBench 修改了编码基准中熟悉工程任务的要求,用来测试模型是否会遵循新指令。研究人员发现,有时模型明白并正确实现了新要求,却在另一个模型给出基准原始答案后删除了自己的正确代码。这些结果显示,模型理解要求与始终按照要求执行之间存在差距。
This guest post by Frank Calegari examines changes in mathematics and argues for responding to them.这篇 Frank Calegari 撰写的客座文章讨论数学领域正在发生的变化,并主张采取相应行动。
This guest post by Frank Calegari examines changes in mathematics and argues for responding to them. The post was converted from another file format with AI assistance.这篇 Frank Calegari 撰写的客座文章讨论数学领域正在发生的变化,并主张采取相应行动。文章由其他文件格式转换而来,过程中使用了 AI。
Go 1.27 introduces the new encoding/json/v2 and encoding/json/jsontext packages.Go 1.27 引入了新的 encoding/json/v2 和 encoding/json/jsontext 包。
Go 1.27 introduces the new encoding/json/v2 and encoding/json/jsontext packages. JSON is the most widely used data format on the internet and has been supported by Go’s standard library since before Go 1.0. The new packages extend Go’s standard-library support for reading and writing JSON.Go 1.27 引入了新的 encoding/json/v2 和 encoding/json/jsontext 包。JSON 是互联网上使用最广泛的数据格式,Go 从 1.0 之前就已在标准库中提供支持。这些新包进一步扩展了 Go 标准库读写 JSON 的能力。
The author reflects on the common claim that technology companies spy on people and know intimate details about their lives.作者反思了“科技公司在监视人们、了解其生活隐私”这一常见说法。
The author reflects on the common claim that technology companies spy on people and know intimate details about their lives. Despite that widespread narrative, the author says they have never personally felt anything close to being known in that way. The piece contrasts public assumptions about surveillance with that personal experience.作者反思了“科技公司在监视人们、了解其生活隐私”这一常见说法。尽管这种叙事十分普遍,作者表示自己从未有过被如此了解的切身感受。文章将公众对监控的普遍想象与作者的个人经历进行了对比。
Images 2.6 introduces faster image generation, GIF creation, more precise editing, and new API controls in ChatGPT, Codex, and the API.Images 2.6 为 ChatGPT、Codex 和 API 带来了更快的图像生成、GIF 创建、更精确的编辑以及新的 API 控制功能。
Images 2.6 introduces faster image generation, GIF creation, more precise editing, and new API controls in ChatGPT, Codex, and the API. Users can animate images such as brand mascots, characters, and product shots into looping GIFs. The update also aims to preserve characters, products, and brand details more consistently across editing rounds.Images 2.6 为 ChatGPT、Codex 和 API 带来了更快的图像生成、GIF 创建、更精确的编辑以及新的 API 控制功能。用户可以将品牌吉祥物、角色和产品图像制作成循环 GIF。此次更新还旨在让角色、产品和品牌细节在多轮编辑中保持更一致。
KetchUp is a free, open-source 3D modeling tool aimed at woodworking and furniture design.KetchUp 是一款面向木工和家具设计的免费开源 3D 建模工具。
KetchUp is a free, open-source 3D modeling tool aimed at woodworking and furniture design. It supports precise dimensions and runs on Windows, macOS, Linux, and in a browser. Users can draw basic shapes and freehand outlines, then use operations such as Push/Pull, move, scale, offset, and painting to create forms.KetchUp 是一款面向木工和家具设计的免费开源 3D 建模工具。它支持精确尺寸,可运行于 Windows、macOS、Linux 以及浏览器中。用户可以绘制基本图形和自由手绘轮廓,再通过推拉、移动、缩放、偏移和绘色等操作创建模型。
Kitty 0.50 adds docked windows that can attach to the edge of another window or tab.Kitty 0.50 增加了停靠窗口,可附着在另一个窗口或标签页的边缘。
Kitty 0.50 adds docked windows that can attach to the edge of another window or tab. Docks sit outside the main layout and can serve as status bars, file trees, or shells beside an editor. They follow their host window across tabs, close with it, and are saved in sessions.Kitty 0.50 增加了停靠窗口,可附着在另一个窗口或标签页的边缘。停靠窗口位于主布局之外,适合用作状态栏、文件树或编辑器旁的 Shell。它会随宿主窗口切换标签页,在宿主关闭时一同关闭,并保存到会话中。
The item concerns the internals of PostgreSQL.该项目关注 PostgreSQL 的内部机制。
The item concerns the internals of PostgreSQL. The available material provides no technical discussion of PostgreSQL itself.该项目关注 PostgreSQL 的内部机制。现有内容没有提供 PostgreSQL 本身的技术介绍。
Gemini Omni 1.2 Flash adds native generation and editing of 15-second videos.Gemini Omni 1.2 Flash 新增原生的 15 秒视频生成与编辑能力。
Gemini Omni 1.2 Flash adds native generation and editing of 15-second videos. The release also promises stronger consistency and faster, more cost-effective generation. Its pricing is 30% lower.Gemini Omni 1.2 Flash 新增原生的 15 秒视频生成与编辑能力。此次更新还提升了一致性,并带来更快、更具成本效益的生成体验。其价格降低了 30%。
Researchers bypassed ChatGPT’s sandbox and content security policy to take over accounts by sending victims a link.研究人员绕过了 ChatGPT 的沙箱和内容安全策略,只需向受害者发送链接即可接管账户。
Researchers bypassed ChatGPT’s sandbox and content security policy to take over accounts by sending victims a link. In October 2026, they found that a public ChatGPT share could be turned into attacker-controlled JavaScript running inside the signed-in session of anyone who opened it. An attacker could publish a malicious conversation through ChatGPT’s normal sharing feature, and the payload would execute on chatgpt.com with access to the victim’s authenticated web session.研究人员绕过了 ChatGPT 的沙箱和内容安全策略,只需向受害者发送链接即可接管账户。2026 年 10 月,他们发现可以将普通的 ChatGPT 公开分享转换为攻击者控制的 JavaScript,并在打开链接者的登录会话中运行。攻击者可以通过 ChatGPT 的正常分享功能发布恶意对话,载荷随后会以 chatgpt.com 的身份执行,并获得受害者已认证的网页会话权限。
The author says they have long favored physical minimalism, owning few possessions and doing more with less.作者表示,自己长期推崇物质极简主义,即减少实体物品并以更少资源完成更多事情。
The author says they have long favored physical minimalism, owning few possessions and doing more with less. They argue that the idea responded to the conditions of the 2000s and 2010s, while still retaining value today. The author proposes digital minimalism as a new form suited to the modern era.作者表示,自己长期推崇物质极简主义,即减少实体物品并以更少资源完成更多事情。作者认为,这一理念回应了 21 世纪初和 2010 年代的环境,但如今仍有价值。作者进一步提出,现代社会需要一种新的极简主义:数字极简主义。
Scribd notified users of a security incident in which an unauthorized party may have accessed data in one of its systems.Scribd 通知用户称,其一个系统发生安全事件,未经授权的第三方可能访问了部分数据。
Scribd notified users of a security incident in which an unauthorized party may have accessed data in one of its systems. Potentially exposed information includes email addresses, the last four digits and country associated with a credit card, card issuer, expiry date, and payment history, but not full card details. The company said it had no evidence of misuse and warned users to watch for phishing.Scribd 通知用户称,其一个系统发生安全事件,未经授权的第三方可能访问了部分数据。可能暴露的信息包括电子邮件地址、信用卡末四位及所属国家、发卡机构、有效期和支付历史,但不包括完整银行卡信息。公司表示没有证据表明这些数据已被滥用,并提醒用户警惕钓鱼攻击。
Anthropic's IPO filing reportedly shows that the retail value of its internal AI usage has exceeded the company's employee compensation.据 Anthropic 的 IPO 文件披露,按零售价计算,该公司内部 AI 的使用量已经超过员工薪酬总额。
Anthropic's IPO filing reportedly shows that the retail value of its internal AI usage has exceeded the company's employee compensation. The figure is presented as a milestone in the company's internal use of AI.据 Anthropic 的 IPO 文件披露,按零售价计算,该公司内部 AI 的使用量已经超过员工薪酬总额。这一数字被视为 Anthropic 内部使用 AI 的一个里程碑。
The piece considers whether the future could still resemble the world depicted in Deus Ex.文章思考未来是否仍可能演变成《杀出重围》所描绘的样子,并引用了首席编剧 Sheldon Pacotti 于 2014 年发表的相关观点。
The piece considers whether the future could still resemble the world depicted in Deus Ex. It quotes lead writer Sheldon Pacotti, who made that observation in 2014.文章思考未来是否仍可能演变成《杀出重围》所描绘的样子,并引用了首席编剧 Sheldon Pacotti 于 2014 年发表的相关观点。
Disc Image Studio converts an image into data that conforms to the physical-layer constraints of CDs, DVDs, and single-layer BD-R discs.Disc Image Studio 可将图片转换为符合 CD、DVD 和单层 BD-R 光盘物理层约束的数据,刻录后光盘表面会呈现可见图案。
Disc Image Studio converts an image into data that conforms to the physical-layer constraints of CDs, DVDs, and single-layer BD-R discs. After burning, the disc surface displays a visible pattern. The Windows desktop version supports design and burning, while Linux and macOS have experimental offline command-line versions.Disc Image Studio 可将图片转换为符合 CD、DVD 和单层 BD-R 光盘物理层约束的数据,刻录后光盘表面会呈现可见图案。Windows 桌面版支持设计与刻录,Linux 和 macOS 则提供实验性的离线命令行版本。
Since November 5, 2026, Outlook.com and Hotmail have been bouncing some messages sent from domains without a DMARC record.自 2026 年 11 月 5 日起,Outlook.com 和 Hotmail 开始退回部分来自未配置 DMARC 记录域名的邮件。
Since November 5, 2026, Outlook.com and Hotmail have been bouncing some messages sent from domains without a DMARC record. The change affects small senders as well as companies sending mail to outlook.com, hotmail.com, live.com, or msn.com addresses. The reported fix is to publish a single DMARC record in DNS.自 2026 年 11 月 5 日起,Outlook.com 和 Hotmail 开始退回部分来自未配置 DMARC 记录域名的邮件。这一变化不仅影响大型企业,也影响向 outlook.com、hotmail.com、live.com 或 msn.com 地址发信的小型发件方。据报道,解决办法是在 DNS 中发布一条 DMARC 记录。
U.S. labor share fell to 52.8% of economic output in the second quarter of 2026, the lowest level since the Bureau of Labor Statistics began recording the measure in 1947.2026 年第二季度,美国劳动者获得的经济产出份额降至 52.8%,为美国劳工统计局 1947 年开始记录该指标以来的最低水平。
U.S. labor share fell to 52.8% of economic output in the second quarter of 2026, the lowest level since the Bureau of Labor Statistics began recording the measure in 1947. It was 53.6% in the previous quarter, compared with an average of about 55.6% for the decade. The decline occurred as corporate earnings and GDP grew while workers received a smaller share through wages and salaries.2026 年第二季度,美国劳动者获得的经济产出份额降至 52.8%,为美国劳工统计局 1947 年开始记录该指标以来的最低水平。上一季度这一比例为 53.6%,本十年的平均值约为 55.6%。与此同时,企业利润和 GDP 增长,但劳动者通过工资和薪酬获得的份额下降。
Two lawyers embedded a white-on-white instruction in a court filing telling an AI to treat their case leniently.两名律师在法院文件中嵌入了白底白字的指令,要求 AI 对他们的案件从轻处理。
Two lawyers embedded a white-on-white instruction in a court filing telling an AI to treat their case leniently. The court's AI document-review system detected the hidden prompt injection. The court fined the lawyers $16,000.两名律师在法院文件中嵌入了白底白字的指令,要求 AI 对他们的案件从轻处理。法院的 AI 文档审查系统检测到了这条隐藏的提示注入指令。法院因此对两名律师处以 16,000 美元罚款。
Bones may release proteins that slow aging and extend life span.骨骼可能会释放能够减缓衰老并延长寿命的蛋白质。
Bones may release proteins that slow aging and extend life span. The title presents this as a possible biological finding without identifying the proteins or describing confirmed effects.骨骼可能会释放能够减缓衰老并延长寿命的蛋白质。标题将其描述为一种可能的生物学发现,但未说明具体蛋白质或已确认的效果。
Molly, a Signal fork, had its latest release on July 31, version v8.19.2-4.Molly 是 Signal 的一个分支,最近一次发布是在 7 月 31 日,版本为 v8.19.2-4。
Molly, a Signal fork, had its latest release on July 31, version v8.19.2-4. The supplied information says it has not been updated for roughly three months and is vulnerable to attacks.Molly 是 Signal 的一个分支,最近一次发布是在 7 月 31 日,版本为 v8.19.2-4。提供的信息称,它已有约三个月未更新,并且容易受到攻击。
MCP Explained is an interactive guide to the Model Context Protocol.《MCP 详解》是一份交互式 Model Context Protocol 指南。
MCP Explained is an interactive guide to the Model Context Protocol. It walks through JSON-RPC messages, transports, authentication, tools, resources, sampling, tasks, tool search, and codemode. MCP is described as an open standard that connects AI applications with the tools and data they use.《MCP 详解》是一份交互式 Model Context Protocol 指南。它逐步介绍 JSON-RPC 消息、传输、认证、工具、资源、采样、任务、工具搜索和 codemode。MCP 被定义为一种将 AI 应用与其使用的工具和数据连接起来的开放标准。
Ricursive Intelligence presents Ripple-65B as achieving U.S. open-weight state of the art on agentic coding. The company describes itself as a full-stack AI lab that co-optimizes hardware, software, and models. Its broader goal is to accelerate chip development and enable more custom silicon.Ricursive Intelligence 将 Ripple-65B 预览版定位为在智能体编程方面达到美国开放权重模型的最先进水平。该公司是一家协同优化硬件、软件和模型的全栈 AI 实验室。其更广泛的目标是加速芯片开发,推动更多定制化硅芯片的出现。
Window Sweaters is a macOS menu-bar app that adds knitted borders to application windows.Window Sweaters 是一款 macOS 菜单栏应用,会为应用窗口添加针织风格的边框。
Window Sweaters is a macOS menu-bar app that adds knitted borders to application windows. It offers sweater-like borders and colors inspired by different apps. The project can also be handed to a coding agent to install.Window Sweaters 是一款 macOS 菜单栏应用,会为应用窗口添加针织风格的边框。它提供类似毛衣的边框和受不同应用启发的配色。项目还可以交给编程代理来完成安装。
OmniDiff is an AGPLv3 open-source diff tool written in Rust that handles code, configuration, images, PDFs, archives, cursors, and fonts.OmniDiff 是一个采用 AGPLv3 许可、用 Rust 编写的开源差异工具,支持代码、配置文件、图像、PDF、压缩包、光标和字体等内容。
OmniDiff is an AGPLv3 open-source diff tool written in Rust that handles code, configuration, images, PDFs, archives, cursors, and fonts. It can compare C++ files tens of thousands of lines long in under 100 milliseconds on a 2023 ThinkPad T14s. Its pipelines were informed by manually studying 2,000 diffs sampled from 7,400 Gentoo package repositories, and the project is currently at version 0.3.0.OmniDiff 是一个采用 AGPLv3 许可、用 Rust 编写的开源差异工具,支持代码、配置文件、图像、PDF、压缩包、光标和字体等内容。它在 2023 年款 ThinkPad T14s 上可在不到 100 毫秒内比较数万行的 C++ 文件。其处理流程源于对 7400 个 Gentoo 软件包仓库中抽样的 2000 个差异进行人工研究,目前版本为 0.3.0。
Coldframe is a Python honeypot that impersonates supermarket refrigeration controllers, including the Danfoss AK-SM 800A and Dixell XWEB interfaces.Coldframe 是一个用 Python 编写的蜜罐,伪装成超市制冷控制器,包括 Danfoss AK-SM 800A 和 Dixell XWEB 的界面。
Coldframe is a Python honeypot that impersonates supermarket refrigeration controllers, including the Danfoss AK-SM 800A and Dixell XWEB interfaces. It uses only the Python standard library and matches details such as headers, page sizes, certificate fields, and TLS versions found on real devices. The honeypot recognizes several exploit attempts, including the CVE-2025-41450 login bypass, and records submitted codes.Coldframe 是一个用 Python 编写的蜜罐,伪装成超市制冷控制器,包括 Danfoss AK-SM 800A 和 Dixell XWEB 的界面。它仅使用 Python 标准库,并模拟真实设备的请求头、页面大小、证书字段和 TLS 版本等细节。该蜜罐能够识别多种攻击尝试,包括 CVE-2025-41450 登录绕过,并记录提交的代码。
NVIDIA Auto Ontology is an open Python framework for building semantic, graph-based reasoning layers over data.NVIDIA Auto Ontology 是一个用于构建语义化图推理层的开源 Python 框架。
NVIDIA Auto Ontology is an open Python framework for building semantic, graph-based reasoning layers over data. It reconciles enterprise metadata and definitions across systems, helping agents locate, interpret, and retrieve data more reliably. The layer is exposed to agent systems through MCP, CLI, and API, and supports text-to-SQL evaluation workflows.NVIDIA Auto Ontology 是一个用于构建语义化图推理层的开源 Python 框架。它整合企业元数据并协调不同系统中的定义,帮助智能体更可靠地定位、理解和检索数据。该推理层可通过 MCP、CLI 和 API 接入智能体,也支持 text-to-SQL 评测流程。
Nibnib is an image format using 4 bits per pixel and optimized for extremely fast decompression.Nibnib 是一种每像素 4 bit、针对极快解压速度优化的图像格式。
Nibnib is an image format using 4 bits per pixel and optimized for extremely fast decompression. It began as an experiment targeting DVI output on an RP2350 microcontroller, where roughly six cycles per pixel must cover pixel fetching, palette lookup, and storage in a line buffer.Nibnib 是一种每像素 4 bit、针对极快解压速度优化的图像格式。它最初是为 RP2350 微控制器上的 DVI 输出进行的实验,该设备每个像素约有 6 个时钟周期,需要完成像素读取、调色板查询和写入行缓冲区。
Unleash is an open-source feature-management platform that lets teams control when and how new features are released.Unleash 是一个开源功能管理平台,让团队控制新功能何时以及如何向用户发布。
Unleash is an open-source feature-management platform that lets teams control when and how new features are released. It supports smaller, more manageable production deployments and can streamline software delivery. The project has been developed for 11 years.Unleash 是一个开源功能管理平台,让团队控制新功能何时以及如何向用户发布。它支持更小、更易管理的生产部署,并可简化软件交付流程。该项目已开发 11 年。
Rune is a native UI framework for Rust that uses GPU rendering across macOS, Windows, Linux, and the web.Rune 是一个使用 GPU 渲染的 Rust 原生 UI 框架,支持 macOS、Windows、Linux 和 Web。
Rune is a native UI framework for Rust that uses GPU rendering across macOS, Windows, Linux, and the web. It provides its own renderer, text system, layout engine, and spring-based animations. Applications can also run headlessly, allowing agents to render PNG images, dump the element tree, and drive scripted input.Rune 是一个使用 GPU 渲染的 Rust 原生 UI 框架,支持 macOS、Windows、Linux 和 Web。它自带渲染器、文本系统、布局引擎和基于弹簧模型的动画。应用还可以无头运行,使智能体能够渲染 PNG、导出元素树并通过脚本驱动输入。
iSponsorBlockTV is a self-hosted application that connects to the YouTube TV app and automatically skips sponsored segments and intros using the SponsorBlock API.iSponsorBlockTV 是一款自托管应用,可连接 YouTube TV 客户端,并通过 SponsorBlock API 自动跳过赞助片段和片头。
iSponsorBlockTV is a self-hosted application that connects to the YouTube TV app and automatically skips sponsored segments and intros using the SponsorBlock API. It can also mute videos automatically and press YouTube's "Skip Ad" button when it appears. YouTube appears to have changed its screen-ID format and is revoking existing codes, so devices must be paired again.iSponsorBlockTV 是一款自托管应用,可连接 YouTube TV 客户端,并通过 SponsorBlock API 自动跳过赞助片段和片头。它还可以自动静音,并在 YouTube 广告出现“跳过广告”按钮后自动点击。YouTube 似乎更改了屏幕 ID 格式并正在撤销现有代码,因此设备需要重新配对。
NVIDIA Kubernetes Engine is a management plane for running Kubernetes on bare-metal NVIDIA GPU infrastructure.NVIDIA Kubernetes Engine 是一个用于在裸机 NVIDIA GPU 基础设施上运行 Kubernetes 的管理平面。
NVIDIA Kubernetes Engine is a management plane for running Kubernetes on bare-metal NVIDIA GPU infrastructure. Platform teams can install it to create and operate workload clusters, each with its own control plane, etcd, and node pools. NVKE is currently a preview release, and its APIs, configuration, and behavior may change, so it is not recommended for production workloads.NVIDIA Kubernetes Engine 是一个用于在裸机 NVIDIA GPU 基础设施上运行 Kubernetes 的管理平面。平台团队可以安装它来创建和运营工作负载集群,每个集群都拥有独立的控制平面、etcd 和节点池。NVKE 目前仍处于预览阶段,API、配置和行为可能发生变化,因此不建议用于生产工作负载。
Rivet is an open-source, self-hostable orchestrator for agent workloads built around lightweight actors backed by S3.Rivet 是一个开源、可自行托管的代理工作负载编排器,围绕由 S3 支持的轻量级 actor 构建。
Rivet is an open-source, self-hostable orchestrator for agent workloads built around lightweight actors backed by S3. Each actor provides a long-lived process, in-memory context, a WebSocket server, durable SQLite storage backed by S3, and idle hibernation. The team says the design grew out of infrastructure work for multiplayer games and aims to avoid the overhead of running a heavyweight sandbox per agent.Rivet 是一个开源、可自行托管的代理工作负载编排器,围绕由 S3 支持的轻量级 actor 构建。每个 actor 都提供长期运行的进程、内存上下文、WebSocket 服务器、由 S3 支持的持久化 SQLite 存储,以及空闲休眠功能。团队表示,该设计源自多人游戏基础设施工作,旨在避免为每个代理运行重量级沙箱的开销。
This project is a Claude Code skill for creating product videos.这是一个用于制作产品视频的 Claude Code 技能。
This project is a Claude Code skill for creating product videos. It supports real-browser capture, branded composition, and local mastering, and includes a quick start, worked videos, and usage instructions.这是一个用于制作产品视频的 Claude Code 技能。它支持真实浏览器录制、品牌化合成和本地母版制作,并提供快速入门、示例视频和使用说明。
BootGPT places an entire inference engine in a 510-byte MBR, with model weights, vocabulary, and strings stored in the following disk sectors.BootGPT 将完整推理引擎放入 510 字节的 MBR 中,并把模型权重、词汇表和字符串存储在后续磁盘扇区。
BootGPT places an entire inference engine in a 510-byte MBR, with model weights, vocabulary, and strings stored in the following disk sectors. Its small character-level neural network was trained to generate Linux boot messages indefinitely, matching a simulator character by character. Debugging uncovered an x86 assembly bug involving the remainder left in EDX by integer division and a pointer that also used EDX.BootGPT 将完整推理引擎放入 510 字节的 MBR 中,并把模型权重、词汇表和字符串存储在后续磁盘扇区。这个小型字符级神经网络经过训练,可以无限生成 Linux 启动消息,并与模拟器逐字符匹配。调试过程中发现了一个 x86 汇编错误:整数除法会把余数留在 EDX,而偏置表指针也使用了 EDX。
Stroq is an open-source local tool for inspecting Claude Code, Codex, and Cursor sessions.Stroq 是一款开源本地工具,用于检查 Claude Code、Codex 和 Cursor 的会话。
Stroq is an open-source local tool for inspecting Claude Code, Codex, and Cursor sessions. It identifies known secrets, counts commands derived from text an agent read, and shows which keys left the machine. Local hooks can then trace risky commands back to the content that prompted them, without telemetry.Stroq 是一款开源本地工具,用于检查 Claude Code、Codex 和 Cursor 的会话。它可以识别已知密钥,统计代理从所读取文本中产生的命令,并显示哪些密钥离开了本机。用户还可以通过本地钩子将高风险命令追溯到触发它们的内容,而且不会发送遥测数据。
Waymark is an open-source durable execution library similar to Temporal or DBOS.Waymark 是一个类似 Temporal 或 DBOS 的开源持久执行库。
Waymark is an open-source durable execution library similar to Temporal or DBOS. It lets developers write workflows in ordinary Python code using conditionals, loops, and asyncio-style parallelism, while the runtime handles distributed execution. Rather than replaying past state in an interpreter, Waymark walks the program’s AST and compiles the logic into bytecode for a stackless virtual machine.Waymark 是一个类似 Temporal 或 DBOS 的开源持久执行库。开发者可以用普通 Python 代码编写工作流,包括条件语句、循环和类似 asyncio 的并行操作,再由运行时负责分布式执行。Waymark 不通过解释器重放历史状态,而是遍历程序 AST,将逻辑编译为运行在无栈虚拟机上的字节码。
Mizu provides thread-speed asynchronous and parallel execution for Python over shared memory.Mizu 通过共享内存为 Python 提供接近线程速度的异步和并行执行能力。
Mizu provides thread-speed asynchronous and parallel execution for Python over shared memory. Its benchmarks report reducing task round trips from roughly 100 microseconds over sockets to less than 1 microsecond. Lists, dictionaries, NumPy arrays, and Arrow tables can be shared between Python and R as zero-copy views.Mizu 通过共享内存为 Python 提供接近线程速度的异步和并行执行能力。其基准测试称,任务往返时间可从基于套接字的约 100 微秒降至不到 1 微秒。列表、字典、NumPy 数组和 Arrow 表都可以作为零拷贝视图在 Python 与 R 之间共享。
HyperFlow H3 Flash is a compressed, few-step version of MiniMax-H3 with 13.7 billion parameters.HyperFlow H3 Flash 是 MiniMax-H3 的压缩少步数版本,包含 137 亿个参数。
HyperFlow H3 Flash is a compressed, few-step version of MiniMax-H3 with 13.7 billion parameters. It generates video and synchronized stereo audio together and supports INT8 and GGUF formats. The project also provides weights, ComfyUI nodes, a GGUF setup guide, and an eight-step LoRA for the full-size model.HyperFlow H3 Flash 是 MiniMax-H3 的压缩少步数版本,包含 137 亿个参数。它能够同时生成视频和同步立体声,并支持 INT8 与 GGUF 格式。项目还提供模型权重、ComfyUI 节点、GGUF 配置指南,以及适用于完整模型的 8 步 LoRA。
Spindle is a multicore-by-design, Eio-native web framework for OCaml that uses a type-driven API to derive codecs, request validation, OpenAPI specifications, and frontend schemas.Spindle 是一个面向多核设计、原生支持 Eio 的 OCaml Web 框架,通过类型驱动的 API 自动生成编解码器、请求验证、OpenAPI 规范和前端模式。
Spindle is a multicore-by-design, Eio-native web framework for OCaml that uses a type-driven API to derive codecs, request validation, OpenAPI specifications, and frontend schemas. It includes an RFC-strict HTTP server, WebSockets, SSE, MCP server primitives, tracing, metrics, and logging. Its benchmarks report performance two times faster than NestJS and five times faster than FastAPI on a single core, although speed was not the project’s primary goal.Spindle 是一个面向多核设计、原生支持 Eio 的 OCaml Web 框架,通过类型驱动的 API 自动生成编解码器、请求验证、OpenAPI 规范和前端模式。它包含严格遵循 RFC 的 HTTP 服务器、WebSocket、SSE、MCP 服务器基础组件,以及分布式追踪、指标和日志功能。基准测试显示,它在单核上运行速度约为 NestJS 的两倍、FastAPI 的五倍,不过速度并非项目的首要目标。
ZapFast is a lightweight, native WhatsApp client written in Rust with an egui interface.ZapFast 是一款使用 Rust 编写、基于 egui 界面的轻量级原生 WhatsApp 客户端。
ZapFast is a lightweight, native WhatsApp client written in Rust with an egui interface. It is designed to run on Linux, macOS, and Windows. The project provides downloads, setup instructions, a user guide, and contribution information.ZapFast 是一款使用 Rust 编写、基于 egui 界面的轻量级原生 WhatsApp 客户端。它面向 Linux、macOS 和 Windows 平台。项目提供下载、安装说明、用户指南和参与贡献的方式。
PyLings teaches Python by having users fix code in exercises ranging from basic concepts to senior-level interview problems.PyLings 通过让用户修复代码来学习 Python,练习范围从基础知识延伸到高级面试题。
PyLings teaches Python by having users fix code in exercises ranging from basic concepts to senior-level interview problems. It includes 35 warm-ups and 20 interview exercises, requires no dependencies, and targets Python 3.14. The tool opens the next exercise and reruns its tests whenever the user saves.PyLings 通过让用户修复代码来学习 Python,练习范围从基础知识延伸到高级面试题。它包含 35 道热身题和 20 道面试题,不依赖额外软件包,面向 Python 3.14。用户每次保存时,工具都会打开下一道练习并重新运行测试。
This project collects V100 SXM2 pin maps, two- and four-GPU NVLink connection plans, and connector-only KiCad resources.该项目整理了 V100 SXM2 引脚图、双 GPU 和四 GPU 的 NVLink 连接方案,以及仅含连接器的 KiCad 资源。
This project collects V100 SXM2 pin maps, two- and four-GPU NVLink connection plans, and connector-only KiCad resources. It aims to support an open-source carrier-board design for Nvidia V100 SXM2 modules, while enabling others to develop their own designs.该项目整理了 V100 SXM2 引脚图、双 GPU 和四 GPU 的 NVLink 连接方案,以及仅含连接器的 KiCad 资源。项目目标是支持 Nvidia V100 SXM2 模块的开源载板设计,同时让其他人能够据此开发自己的设计。
WaveDudeNext is an interactive Windows simulator for two-dimensional wave optics.WaveDudeNext 是一款面向 Windows 的交互式二维波动光学模拟器。
WaveDudeNext is an interactive Windows simulator for two-dimensional wave optics. Users can turn images into refractive scenes, place light sources and optical components, and observe refraction, diffraction, and interference. It displays both the wave field and time-averaged intensity, while supporting probes, timeline control, video export, adjustable brightness, custom color schemes, and multiple interface languages.WaveDudeNext 是一款面向 Windows 的交互式二维波动光学模拟器。用户可以将图像转换为折射场景,放置光源和光学元件,并观察折射、衍射与干涉现象。程序同时显示波场和时间平均强度,并支持探针、时间轴控制、视频导出、亮度调节、自定义配色以及多种界面语言。
The project is an open field guide covering 39 CAPTCHA types, their identifying signals, documented open-source bypass methods, and paid alternatives.该项目是一份开放实地指南,涵盖 39 种 CAPTCHA 类型、识别信号、已有的开源绕过方法及付费替代方案,并提供可搜索的网页版内容。
The project is an open field guide covering 39 CAPTCHA types, their identifying signals, documented open-source bypass methods, and paid alternatives. It provides a searchable web version and frames the material for education and authorized research. Its stated goal is to make changing CAPTCHA systems easier to navigate, compare, and cite while emphasizing responsible use.该项目是一份开放实地指南,涵盖 39 种 CAPTCHA 类型、识别信号、已有的开源绕过方法及付费替代方案,并提供可搜索的网页版内容。项目将材料定位为教育和获授权研究用途。其目标是在强调负责任使用的同时,让不断变化的 CAPTCHA 系统更易于查找、比较和引用。
The project is titled “decode kernel panic from QR Code photo.” It is presented as a GitHub Gist, but the available page content does not provide the decoding method or a working example.这个项目名为“从二维码照片中解码内核崩溃信息”。
The project is titled “decode kernel panic from QR Code photo.” It is presented as a GitHub Gist, but the available page content does not provide the decoding method or a working example.这个项目名为“从二维码照片中解码内核崩溃信息”。它以 GitHub Gist 的形式呈现,但现有页面内容没有提供解码方法或可运行示例。
Codemap generates a single HTML page mapping a Go module’s packages, dependencies, files, and functions.Codemap 会生成一个 HTML 页面,展示 Go 模块中的包、依赖、文件和函数。
Codemap generates a single HTML page mapping a Go module’s packages, dependencies, files, and functions. Users can search the map, inspect what each package uses and what uses it, and jump to function definitions in GitHub-hosted modules. Tests are excluded from the map.Codemap 会生成一个 HTML 页面,展示 Go 模块中的包、依赖、文件和函数。用户可以搜索地图,查看每个包依赖什么以及被什么依赖,并跳转到 GitHub 模块中的函数定义。地图不包含测试代码。
MoofTCP is a drop-in replacement for Apple's MacTCP on System 6 and 68000-based Macs.MoofTCP 是 System 6 和 68000 Mac 上苹果 MacTCP 的即插即用替代品。
MoofTCP is a drop-in replacement for Apple's MacTCP on System 6 and 68000-based Macs. Existing MacTCP programs can run unchanged on its newly written, portable-C TCP/IP stack. It includes its own control panel, installer, and DHCP client, and was designed for a Mac Classic with 4 MB of RAM.MoofTCP 是 System 6 和 68000 Mac 上苹果 MacTCP 的即插即用替代品。现有的 MacTCP 程序无需修改即可运行在其用可移植 C 语言重写的 TCP/IP 协议栈上。它还包含独立的控制面板、安装程序和 DHCP 客户端,并针对仅有 4 MB 内存的 Mac Classic 设计。
Mocktastic is a small, customizable Meshtastic-compatible LoRa client for ESP32 boards built with ArduinoIDE.Mocktastic 是一个面向 ESP32 开发板、使用 ArduinoIDE 构建的小型可定制 Meshtastic 兼容 LoRa 客户端。
Mocktastic is a small, customizable Meshtastic-compatible LoRa client for ESP32 boards built with ArduinoIDE. It is an educational and experimental C++ implementation rather than a fork of the official firmware. The project demonstrates Meshtastic packet encoding and decoding and includes an environmental-sensor example.Mocktastic 是一个面向 ESP32 开发板、使用 ArduinoIDE 构建的小型可定制 Meshtastic 兼容 LoRa 客户端。它是用于教育和实验的 C++ 实现,并非官方固件的分支。项目演示了 Meshtastic 数据包的编码与解码,并包含环境传感器示例。
PyWolf3D is a Wolfenstein 3D engine and game clone written from scratch in Python using Pygame.PyWolf3D 是一个使用 Python 和 Pygame 从零编写的 Wolfenstein 3D 引擎及游戏克隆版,并提供演示。
PyWolf3D is a Wolfenstein 3D engine and game clone written from scratch in Python using Pygame. A demo is included.PyWolf3D 是一个使用 Python 和 Pygame 从零编写的 Wolfenstein 3D 引擎及游戏克隆版,并提供演示。
Gauge is an open-source CLI that recursively improves documentation for coding agents.Gauge 是一个用于递归改进 coding agent 文档的开源 CLI。
Gauge is an open-source CLI that recursively improves documentation for coding agents. It runs tasks, analyzes agent transcripts for friction or failures, creates documentation variants, and repeats the process until a variant resolves the issues. The tool includes a local execution backend and can also optimize skills, CLIs, and other agent-facing interfaces.Gauge 是一个用于递归改进 coding agent 文档的开源 CLI。它会运行任务、分析 agent 会话记录中的阻碍或失败、生成文档变体,并重复这一过程,直到某个变体解决相关问题。该工具包含本地执行后端,也可用于优化技能、CLI 和其他面向 agent 的界面。
Flawless UI-UX is a skill for coding agents that helps them build distinctive and usable interfaces.Flawless UI-UX 是一项帮助编程智能体构建具有特色且易用界面的技能。
Flawless UI-UX is a skill for coding agents that helps them build distinctive and usable interfaces. Version 1.0.0 includes an interactive installer and support for twelve agents. The installer can be run with npm, pnpm, or Bun, and lets users choose project or global scope, select agents, review destinations, and confirm file copying.Flawless UI-UX 是一项帮助编程智能体构建具有特色且易用界面的技能。1.0.0 版本提供交互式安装程序,并支持 12 种智能体。用户可以通过 npm、pnpm 或 Bun 启动安装器,选择项目级或全局范围、指定智能体、检查目标位置并确认复制文件。
Panthera is Ampere’s cycle-accurate software performance model for studying prospective CPU features and evaluating microarchitectural trade-offs.Panthera 是 Ampere 的周期精确软件性能模型,用于研究候选 CPU 特性并评估微架构取舍。
Panthera is Ampere’s cycle-accurate software performance model for studying prospective CPU features and evaluating microarchitectural trade-offs. It is built by Ampere’s Architecture & Performance team to guide development of high-performance cloud processors. The project includes a user guide for building and using Panthera and its associated tools.Panthera 是 Ampere 的周期精确软件性能模型,用于研究候选 CPU 特性并评估微架构取舍。它由 Ampere 架构与性能团队开发,以指导高性能云处理器的设计。项目还提供了构建和使用 Panthera 及相关工具的用户指南。
ZeBridge is a bidirectional, local-first synchronization engine connecting PostgreSQL with edge SQLite, PGLITE, PostgreSQL, and DuckDB clients over NATS.ZeBridge 是一个双向、以本地优先为理念的同步引擎,通过 NATS 连接 PostgreSQL 与边缘端的 SQLite、PGLITE、PostgreSQL 和 DuckDB 客户端。
ZeBridge is a bidirectional, local-first synchronization engine connecting PostgreSQL with edge SQLite, PGLITE, PostgreSQL, and DuckDB clients over NATS. It keeps a local database on each device and queues writes while the device is offline. When connectivity returns, PostgreSQL remains authoritative by applying rules, resolving conflicts, and returning a result to the application.ZeBridge 是一个双向、以本地优先为理念的同步引擎,通过 NATS 连接 PostgreSQL 与边缘端的 SQLite、PGLITE、PostgreSQL 和 DuckDB 客户端。它在每台设备上保留本地数据库,并在离线时排队保存写入操作。重新连接后,PostgreSQL 仍是权威数据源,负责应用规则、解决冲突并向应用返回结果。
tmux-restore restores tmux sessions, layouts, and scrollback after a reboot.tmux-restore 可在重启后恢复 tmux 会话、布局和滚动缓冲区。
tmux-restore restores tmux sessions, layouts, and scrollback after a reboot. It can also reopen conversations for Claude Code, Codex, Gemini CLI, opencode, and pi. The project describes itself as reboot-proof tmux for coding agents.tmux-restore 可在重启后恢复 tmux 会话、布局和滚动缓冲区。它还能够重新打开 Claude Code、Codex、Gemini CLI、opencode 和 pi 中此前的对话。该项目将自己定位为面向编程代理的“重启无忧” tmux 工具。
This survival assistant runs fully offline on Windows, Linux, and macOS, including from a laptop or USB drive.这款生存助手可在 Windows、Linux 和 macOS 上完全离线运行,也可以从笔记本电脑或 U 盘启动。
This survival assistant runs fully offline on Windows, Linux, and macOS, including from a laptop or USB drive. It provides sourced guides on water, fire, food, first aid, and other topics in Portuguese, English, and Spanish, alongside a local LLM using llama.cpp or Ollama. Kiwix libraries provide additional offline reference material.这款生存助手可在 Windows、Linux 和 macOS 上完全离线运行,也可以从笔记本电脑或 U 盘启动。它提供葡萄牙语、英语和西班牙语的水、火、食物、急救等主题资料,并通过 llama.cpp 或 Ollama 运行本地 LLM。Kiwix 资料库则提供额外的离线参考内容。
LDraw Mecha uses agentic language models to analyze annotated LDraw LEGO model files, infer their submodels and mechanisms, plan animations, and execute them in Blender.LDraw Mecha 使用代理式语言模型分析带注释的 LDraw LEGO 模型文件,推断其中的子模型和机械结构,规划动画,并在 Blender 中执行。
LDraw Mecha uses agentic language models to analyze annotated LDraw LEGO model files, infer their submodels and mechanisms, plan animations, and execute them in Blender. It outputs a rigged and animated Blender file from an initial .mpd source file. The project builds on LDraw Nova, an earlier tool for generating LEGO CAD models from LDraw source code.LDraw Mecha 使用代理式语言模型分析带注释的 LDraw LEGO 模型文件,推断其中的子模型和机械结构,规划动画,并在 Blender 中执行。它可以从初始的 .mpd 源文件生成带绑定和动画的 Blender 文件。该项目建立在此前用于从 LDraw 源代码生成 LEGO CAD 模型的 LDraw Nova 之上。
Uniform is a JavaScript framework for web applications that uses WebAssembly for layout and GPU rendering, with a reported size of 280 KB.Uniform 是一个用于网页应用的 JavaScript 框架,使用 WebAssembly 处理布局并通过 GPU 渲染,体积据称为 280 KB。
Uniform is a JavaScript framework for web applications that uses WebAssembly for layout and GPU rendering, with a reported size of 280 KB. Its creator aims to avoid the platform limits of native frameworks, the HTML/CSS constraints of browser-engine approaches, and the size or language requirements of other cross-platform frameworks. The design bundles the UI framework with the application so developers can update it independently of platform release cycles.Uniform 是一个用于网页应用的 JavaScript 框架,使用 WebAssembly 处理布局并通过 GPU 渲染,体积据称为 280 KB。作者希望避免原生框架的平台限制、基于浏览器引擎方案所继承的 HTML/CSS 约束,以及其他跨平台框架的体积或语言要求。该设计将 UI 框架直接与应用打包,使开发者能够独立于平台发布周期更新它。
TubeHaus is a private, self-hosted YouTube client available as a Docker Compose web app and native apps for Android, Windows, macOS, Linux, and iPhone/iPad.TubeHaus 是一款私密、自托管的 YouTube 客户端,提供 Docker Compose 网页版,以及 Android、Windows、macOS、Linux 和 iPhone/iPad 原生应用。
TubeHaus is a private, self-hosted YouTube client available as a Docker Compose web app and native apps for Android, Windows, macOS, Linux, and iPhone/iPad. It requires no Google account and provides no tracking. The client preserves subscriptions and history on the user's own server or device while retaining a YouTube-like layout and keyboard shortcuts.TubeHaus 是一款私密、自托管的 YouTube 客户端,提供 Docker Compose 网页版,以及 Android、Windows、macOS、Linux 和 iPhone/iPad 原生应用。它无需 Google 账户,也不进行跟踪。用户可以将订阅和历史记录保存在自己的服务器或设备上,同时保留类似 YouTube 的布局和键盘快捷键。
This custom firmware for Baofeng UV-5RM radios adds applications including texting and spectrum tools.这套 Baofeng UV-5RM 自定义固件新增了短信和频谱工具等应用。
This custom firmware for Baofeng UV-5RM radios adds applications including texting and spectrum tools. Its texting feature supports short messages between two radios with delivery confirmation and requires no phone, internet connection, or additional hardware. CHIRP will no longer communicate with the modified device, so channels must be programmed with the included applications.这套 Baofeng UV-5RM 自定义固件新增了短信和频谱工具等应用。短信功能支持两台无线电之间发送短消息并确认送达,不需要手机、互联网或额外硬件。改装后设备将无法再与 CHIRP 通信,因此必须使用随附应用配置频道。
SafeScript is a strict computational subset of JavaScript designed to execute untrusted code in multiple environments.SafeScript 是 JavaScript 的严格计算子集,旨在让不可信代码能够在多种环境中安全执行。
SafeScript is a strict computational subset of JavaScript designed to execute untrusted code in multiple environments. It excludes features including eval, new, globals, and prototype access, then compiles to plain JavaScript and PHP. The compiler is about 48 KB minified, or roughly 12 KB gzipped, and uses .safe files intended to be locally verifiable.SafeScript 是 JavaScript 的严格计算子集,旨在让不可信代码能够在多种环境中安全执行。它排除了 eval、new、全局变量和原型访问等功能,并可编译为普通 JavaScript 和 PHP。编译器压缩后约 48 KB、gzip 后约 12 KB,使用 .safe 文件格式,并支持本地验证。
Datastructures-JS is a JavaScript data-structures project with TypeScript support.Datastructures-JS 是一个支持 TypeScript 的 JavaScript 数据结构项目。
Datastructures-JS is a JavaScript data-structures project with TypeScript support. The project also provides an associated planning board.Datastructures-JS 是一个支持 TypeScript 的 JavaScript 数据结构项目。项目还提供了配套的规划看板。
This hands-on guide explains how to train Mixture-of-Experts models.这份实践指南介绍了如何训练混合专家模型。
This hands-on guide explains how to train Mixture-of-Experts models. It provides a navigable table of contents and covers the practical steps involved in building and training MoE systems.这份实践指南介绍了如何训练混合专家模型。页面提供可点击的目录,并涵盖构建和训练 MoE 系统所需的实际步骤。
The project is a bio-inspired edge-detection algorithm modeled on the visual cortex.该项目是一种受视觉皮层启发的仿生边缘检测算法。
The project is a bio-inspired edge-detection algorithm modeled on the visual cortex. Its current benchmark results fall between traditional methods such as Canny and Sobel and modern deep neural networks, while retaining the interpretability of the former. The author plans to improve video frame rates and performance across different hardware, and is seeking collaborators.该项目是一种受视觉皮层启发的仿生边缘检测算法。目前的基准测试结果介于 Canny、Sobel 等传统方法与现代深度神经网络之间,同时保留了传统方法的可解释性。作者计划提升视频处理帧率和不同硬件上的性能,并寻找合作伙伴。
Async-a-sync is a compiler and runtime system built on Fil-C that lets ordinary blocking-looking C code run asynchronously underneath.Async-a-sync 是构建在 Fil-C 之上的编译器和运行时系统,让表面上阻塞的普通 C 代码在底层异步运行。
Async-a-sync is a compiler and runtime system built on Fil-C that lets ordinary blocking-looking C code run asynchronously underneath. It introduces implicit asynchronous futures without adding syntax, with the first access to a buffer acting as the await point.Async-a-sync 是构建在 Fil-C 之上的编译器和运行时系统,让表面上阻塞的普通 C 代码在底层异步运行。它无需新增语法即可引入隐式异步 future,并将首次访问缓冲区作为 await 点。
Flide is an inference engine designed specifically for Jev-like decision models.Flide 是一款专为 Jev 类决策模型设计的推理引擎。
Flide is an inference engine designed specifically for Jev-like decision models. The project claims up to a 9.5x speedup over vLLM and SGLang and currently supports more than 30 popular local Jev-like models. It can be installed with pip using the package name flide.Flide 是一款专为 Jev 类决策模型设计的推理引擎。项目称其速度最高可达 vLLM 和 SGLang 的 9.5 倍,目前支持 30 多种热门本地 Jev 类模型。用户可以通过 pip 安装名为 flide 的软件包。
Maple is a structural-compression method for mixture-of-experts language models.Maple 是一种面向混合专家语言模型的结构化压缩方法。
Maple is a structural-compression method for mixture-of-experts language models. It compresses Gemma 4 26B-A4B from 9.92 GB to 4.75 GB, equivalent to 1.48 effective bits per parameter. The project reports 100% in-VRAM execution on 6 GB laptop GPUs at 45.4 tokens per second.Maple 是一种面向混合专家语言模型的结构化压缩方法。它将 Gemma 4 26B-A4B 从 9.92 GB 压缩到 4.75 GB,相当于每参数 1.48 个有效比特。项目报告称,该模型可在 6GB 笔记本 GPU 上完全载入显存运行,速度达到每秒 45.4 个 token。