头条报道聚焦一起空中相撞事故：2026 年 5 月 17 日周日在爱达荷州 Mountain Home 空军基地举行的 Gunfighter Skies 航空展上，两架美国海军 EA-18G Growler 在空中相撞。两机共四名机组人员全部安全弹射并打开降落伞，飞机在基地附近撞击地面后起火爆炸。基地位于 Boise 以南约 50 英里，军事人员和现场观众均未受伤。

事故涉及来自 Whidbey Island 的 Electronic Attack Squadron 129 的两架飞机，事发时正在进行空中表演。观众 Shane Odgen 拍到的视频显示，两机发生接触后同步旋转，机组人员弹射并打开降落伞，随后飞机一同坠落并在撞击时爆炸。事故发生后基地被封锁，Idaho Transportation 关闭了从 Simco Road 到基地附近 SH-67 的 SH-167 公路，预计在调查期间将关闭数日。

366th Fighter Wing 联队指挥官 Col. David Gunter 对所有人员安然无恙表示感谢，并赞扬包括市县部门在内的应急队伍迅速响应。协助筹办航空展的志愿组织 Silver Wings of Idaho 发表声明，称当天情绪沉重但对四名飞行员生还感到宽慰，并对来自多机构的执法、消防、急救及其他应急人员展现出的专业与协作表示肯定。

Gunfighter Skies 是 Mountain Home 自 2018 年以来首次举办的航空展；2018 年一名滑翔伞运动员在表演中遇难。航空展行业一直在努力提升约 200 场年均活动的安全标准。 International Council of Air Shows 总裁兼首席执行官 John Cudahy 指出，过去美国航空展平均每年约有两人遇难，但过去十年这一数字已降至接近每年一人，2024 年和 2025 年均无航展致死记录。最近一次致命事故发生于 2022 年，两架老式军机在达拉斯一场活动中相撞，造成六人死亡。

National Weather Service 表示，坠机时能见度良好，阵风最高约 29 英里 / 小时。 U.S. Air Force Thunderbirds 在为期两天的活动中担任压轴表演，组织者将此次活动定位为对航空历史的庆祝与现代军事能力的展示。 The main story covers a mid-air collision between two U.S. Navy EA-18G Growler fighter jets during the Gunfighter Skies Airshow at Mountain Home Air Force Base in Idaho on Sunday, May 17, 2026. All four crew members from both aircraft safely ejected and deployed parachutes before the planes crashed and exploded into a fireball upon impact near the base, located about 50 miles south of Boise. No injuries were reported among personnel at the military base or spectators at the event.

The collision involved two jets from Electronic Attack Squadron 129 based in Whidbey Island, Washington, which were performing an aerial demonstration at the time of the incident. Spectator Shane Odgen captured video showing the two aircraft making contact, spinning in tandem as the crew members ejected, and their parachutes opening before the planes fell together and exploded on impact. The base was locked down following the crash, and Idaho Transportation closed SH-167 from Simco Road to SH-67 near the base, with the closure anticipated to last multiple days during the investigation.

Col. David Gunter, Wing Commander of the 366th Fighter Wing, expressed gratitude that everyone involved was safe and praised the quick response from emergency teams including city and county agencies. Silver Wings of Idaho, the volunteer organization that helped plan the airshow alongside the base, released a statement acknowledging the emotional difficulty of the day while emphasizing relief that all four pilots survived. The organization recognized the outstanding professionalism and teamwork shown by law enforcement, fire, EMS, and other emergency personnel from multiple agencies who responded to the situation.

The Gunfighter Skies event was the first airshow at Mountain Home Air Force Base since 2018, when a hang glider died during a performance. The air show industry has been working to improve safety at the roughly 200 events held annually across the United States. John Cudahy, president and CEO of the International Council of Air Shows, noted that while there used to be an average of about two deaths per year at U.S. air shows, the average has dropped to closer to one death per year over the past decade, with no air show deaths recorded in 2024 or 2025. The last fatal crash occurred in 2022 when two vintage military planes collided at an event in Dallas, killing six people.

The National Weather Service reported good visibility and winds gusting up to 29 mph around the time of the crash. The U.S. Air Force Thunderbirds demonstration squadron had headlined the show both days of the event, which organizers described as a celebration of aviation history and a showcase of modern military capabilities.

GenCAD 是 MIT 研究人员开发的一种新型 AI 系统，能够从二维图像生成完整的参数化 CAD 模型。与以往只输出网格或点云等简化表示的方法不同，GenCAD 直接产出可在工程软件中执行的 CAD 命令序列，不仅重建三维形体，还保留了完整的设计命令历史，便于工程师修改与二次编辑。

该系统解决了 AI 驱动设计中的一大难题：传统 CAD 的数据结构（边界表示，B-rep）复杂难以直接被 AI 处理，现有方法常以牺牲精度和可编辑性为代价。 GenCAD 通过以参数化 CAD 命令序列为工作对象——也就是在 CAD 软件中构建模型的逐步指令——规避了这一瓶颈。

其架构由四个关键模块协同工作：首先，自回归 Transformer 编码器将 CAD 命令序列压缩表示到潜在空间；其次，对比学习模型学习图像与 CAD 命令的联合表示，搭建两种模态之间的桥梁；第三，潜在扩散模型根据输入图像生成新的 CAD 潜在表示；最后，解码器将这些潜在表示还原为可被几何内核执行的 CAD 命令，从而生成三维实体模型。

除从图像生成 CAD 外，GenCAD 还可用于 CAD 检索：在数千个模型的数据库中找到与目标图像匹配的现有 CAD 程序。系统还能生成多样化样本——对同一输入图像提供多种不同的 CAD 解释，为设计师提供多个备选方案，从而显著加速设计流程，使工程师能从概念图快速得到可编辑、可用于制造的 CAD 模型。

研究人员认为 GenCAD 是向从图像实现更精确、可修改的三维建模迈出的重要一步。通过保留 CAD 模型的完整参数化历史，该方法保全了工程应用中关键的精度与可编辑性，具有在自动化设计流程、快速原型制作和设计空间探索等方面的潜在应用价值，可简化从概念图像到功能性、可编辑 CAD 模型的转化过程。 GenCAD is a new AI system developed by researchers at MIT that generates complete, parameterized CAD models from 2D images. Unlike previous approaches that produce simplified representations like meshes or point clouds, GenCAD outputs actual CAD command sequences that can be directly used in engineering software. This means it doesn't just create a 3D shape, but generates the entire history of design commands that engineers can modify and work with.

The system addresses a major challenge in AI-driven design. Traditional CAD data structures, known as boundary representation or B-rep, are complex and difficult for AI models to work with directly. Most existing methods sacrifice the precision and editability that make CAD models valuable for manufacturing and design exploration. GenCAD overcomes this by working with parametric CAD command sequences, which are essentially the step-by-step instructions used to build a 3D model in CAD software.

The architecture combines four key components working together. First, an autoregressive transformer encoder learns to represent CAD command sequences in a compressed latent space. Second, a contrastive learning model bridges the gap between CAD commands and images by learning joint representations of both modalities. Third, a latent diffusion model generates new CAD representations based on input images. Finally, a decoder converts these latent representations back into actual CAD commands that can be executed by geometry kernels to produce 3D solid models.

Beyond just generating CAD from images, GenCAD can also perform CAD retrieval, finding existing CAD programs that match a given image from a database of thousands of models. The system demonstrates sample diversity, meaning it can produce multiple different CAD interpretations of the same input image, giving designers various options to choose from. This capability could significantly speed up the design process by allowing engineers to start with an image concept and quickly get editable, manufacturing-ready CAD models.

The researchers see GenCAD as an important step toward more precise and modifiable 3D modeling from images. By preserving the full parametric history of CAD models, their approach maintains the accuracy and editability that are critical for real-world engineering tasks. The work has potential applications in automated design processes, rapid prototyping, and design space exploration, making it easier to go from conceptual images to functional, editable CAD models.

文章 "Prolog Coding Horror" 向 Prolog 程序员发出警示，指出那些会导致代码有缺陷的常见反模式。作者借用了 Joseph Conrad 的 Heart of Darkness 作比，把这些编程陷阱描述为"恐怖"故事——它们破坏了逻辑编程的优雅与表达力。文章的核心观点是：在 Prolog 中对既有规范保持质疑是有价值的，但执着于过时或不纯的构造只是一种代价高昂而无益的叛逆。

作者列举了几类主要的"恐怖"来源。第一类是使用不纯且非单调的构造，比如 cut (!/0) 、 if-then-else (->)/2 和 var/1，这些会导致程序遗漏本应返回的解——作者认为这比返回错误结果更糟。第二类是通过 assertz/1 、 retract/1 等谓词修改全局数据库，这会引入隐式依赖，使程序脆弱并依赖执行顺序。第三类是不纯的输出：程序员直接把结果打印到终端，而不是让 Prolog 的顶层（toplevel）来报告结果，这既妨碍对输出的推理，也破坏了代码的关系性。

一个说明性例子是所谓的 "horror factorial" 实现。初始版本用了 cut 和低级算术，导致在通用查询下会丢失解。即便去掉 cut，仍因使用 is/2（要求变量已实例化）而存在问题。文章展示了如何用声明性手段改写：用 CLP(FD) 约束（#>/2 、 #=/2）替代低级算术，并去掉 cut，从而得到在最一般查询下也能正确工作的更通用程序。

作者强烈主张尽量停留在 Prolog 的纯、单调子集内。具体做法包括：用约束替代底层算术；通过谓词参数或半上下文记法来传递状态，而不是依赖全局变量；以声明式方式描述解，而不是直接打印。文章指出，低级构造还会增加教学难度，因为学生不得不同时面对声明式语义和操作式语义。

结论是：在 Prolog 中的"叛逆"应当朝向采用现代、声明式的特性，而不是固守过时和不纯的做法。通过拥抱纯性与约束，程序员既能写出更通用的逻辑程序，又能保持可接受的性能，从而充分发挥语言的关系性与声明式优势。 The article "Prolog Coding Horror" serves as a cautionary guide for Prolog programmers, warning against common anti-patterns that lead to defective code. It draws a parallel to Joseph Conrad's Heart of Darkness, framing these programming pitfalls as "horror" stories that undermine the elegance and power of logic programming. The core message is that while rebellion against industry norms can be valuable in Prolog, clinging to outdated or impure constructs is a misguided form of rebellion with high costs and no benefits.

The author identifies several major sources of horror in Prolog programming. The first is losing solutions by using impure and non-monotonic constructs like the cut operator (!/0), if-then-else (->)/2, and var/1. These can cause a program to fail to report intended solutions, which the author argues is worse than reporting wrong answers. The second horror is modifying the global database with predicates like assertz/1 and retract/1, which introduces implicit dependencies that make programs fragile and order-dependent. The third is impure output, where programmers print results directly to the terminal instead of letting the Prolog toplevel report them, which prevents reasoning about the output and breaks the relational nature of the code.

A particularly illustrative example is the "horror factorial" implementation. The initial version uses a cut and low-level arithmetic, causing it to lose solutions when queried generally. Even removing the cut leaves problems due to the use of (is)/2, which requires instantiated variables. The article demonstrates how moving to declarative constructs like CLP(FD) constraints (#>/2, #=/2) and removing the cut creates a more general program that works correctly for the most general query.

The author strongly advocates for staying within the pure, monotonic subset of Prolog. This means using constraints instead of low-level arithmetic, threading state through predicate arguments or semicontext notation instead of global variables, and describing solutions declaratively rather than printing them directly. The article emphasizes that low-level constructs make Prolog harder to teach and learn because they force students to grapple with both declarative and operational semantics simultaneously.

The conclusion reinforces that rebellion in Prolog should be directed toward using modern, declarative features rather than clinging to outdated ones. By embracing purity and constraints, programmers can create more general logic programs while maintaining acceptable performance. The article positions this approach as the path to writing great Prolog code that fully leverages the language's relational nature and declarative power.

在佛蒙特州的乡村，手机信号常常完全覆盖不到，一位名叫 Patrick Schlott 的电气工程师正把老式公用电话带回现代。 Schlott 在电动垂直起降飞行器制造商 Beta Technologies 工作，受 Futel 和 PhilTel 等免费电话项目的启发，他修复并在全州各地安装了可免费使用的公用电话。通过采用互联网语音协议（VoIP）和模拟电话适配器，他将这些复古设备接入互联网，让用户可以在美国和加拿大境内拨打免投币电话。

Schlott 的行动始于 2023 年初，那时他在父母的车库里发现了一部旧电话听筒。出于对复古技术的热爱以及对信号覆盖薄弱地区通信需求的实际考虑，他决定搭建一个公共电话网络。首个装置于 2024 年 3 月安装在 North Tunbridge General Store，在说服店主相信公用电话对旅客和当地居民尤其在紧急情况时有用后投入使用。

技术上，他要把原本为地线启动或环路启动电话线设计的旧式公用电话改造为兼容现代 VoIP 系统。 Schlott 用模拟电话适配器把模拟电话和基于互联网的数字通话桥接起来。他之所以选用公用电话而不是普通内部电话，是因为公用电话一眼就能被认作公共设施，结构坚固适合户外使用，同时带有独特的历史风味。

截至 2025 年 11 月，Schlott 已在图书馆、学校和市政厅等处安装了七部电话，且还有更多计划。需求主要由两方面推动：乡村地区缺乏手机信号，以及佛蒙特州将于 2026 年 9 月起在学校中禁止使用智能手机的新法规。家长和校方把这些公用电话视为学生打电话的必要替代方案。

该项目通过捐款和他自掏腰包资助，Schlott 负担二手公用电话的购买费用，价格依状态与来源从 50 美元到 700 美元不等，另有少量每月 VoIP 服务费。虽然电话供公众免费使用，但他保留了投币机构，既为保留历史感，也迎合黑客社区的兴趣，未来还计划纯粹出于趣味性恢复投币功能。

这些电话还设有特别的拨号代码：拨 0 可通过软电话应用直接联系到 Schlott 本人，其他代码可接通 Vermont United Way（211）、查号服务（411）和自杀热线（988）。 E911 支持紧急呼叫，每个地点的地址都已向服务提供商注册。 Schlott 的工作不仅提供了一项重要服务，也在庆祝老式电信技术的坚固耐用与丰厚历史。 In rural Vermont, where cell service is often nonexistent, an electrical engineer named Patrick Schlott is reviving the old-fashioned pay phone for the modern era. Working at Beta Technologies, an eVTOL manufacturer, Schlott was inspired by similar free-phone projects like Futel and PhilTel to restore and install free-to-use pay phones across the state. Using Voice over Internet Protocol (VoIP) technology, he connects these vintage devices to the internet via analog telephone adapters, allowing users to make coinless calls anywhere in the United States or Canada.

Schlott's journey began in early 2023 when he found an old phone handset in his parents' garage. Combining his passion for vintage technology with a practical need for communication in areas with poor cell coverage, he decided to create a public phone system. His first installation was at the North Tunbridge General Store in March 2024, after convincing the owners that a pay phone would be useful for travelers and locals alike, especially in emergencies.

The technical process involves converting old pay phones, which were originally designed for ground-start or loop-start telephone lines, to work with modern VoIP systems. Schlott uses analog telephone adapters to bridge the gap between the analog phones and digital internet-based calling. He chose pay phones over regular courtesy phones because they are instantly recognizable as public infrastructure, rugged enough to withstand outdoor conditions, and carry a sense of historical charm.

As of November 2025, Schlott has installed seven phones in locations such as libraries, schools, and a town hall, with more planned. The demand is driven by two main factors: the lack of cell service in rural areas and a new Vermont law banning smartphone use in schools starting in September 2026. Parents and school administrators see these pay phones as a necessary alternative for students to make calls.

Funding the project through donations and his own money, Schlott covers the costs of acquiring used pay phones, which can range from $50 to $700 depending on condition and source, and the minimal monthly VoIP service fees. While the phones are free to use, he has kept the coin mechanisms intact for their historical and hacker-friendly appeal, with future plans to incorporate coin functionality purely for fun.

The phones also feature special dialing codes: dialing 0 connects to Schlott himself via a softphone app, while other codes link to services like Vermont United Way (211), directory assistance (411), and the suicide hotline (988). Emergency calls are supported through E911, with each location's address registered with the provider. Schlott's work not only provides a vital service but also celebrates the rugged durability and rich history of vintage telecommunications technology.

在美国各地，人们开始自行对 Flock Safety 的监控摄像头采取行动。自 2025 年 4 月以来，至少有 25 台摄像头在五个州被毁，从 California 到 Virginia 不等。这些破坏并非统一组织发起，而是愤怒的居民对他们认为与联邦移民执法有关联的侵入性监控网络的即时反应。 Virginia 的一名男子 Jeffrey S. Sovern 在六个月内系统性地拆毁了 13 台摄像头，面临 25 项刑事指控；他公开了自己的动机，援引第四修正案，并希望此案能成为遏制侵入性监控的催化剂。

破坏方式各不相同。在 La Mesa，市议会在面对压倒性反对声浪仍投票保留摄像头后，Fletcher Parkway 上的两台摄像头数周后被发现损坏。在 Eugene 和 Springfield，六台摄像头被从杆上剪下，至少一台被喷漆，现场还留有一张写着"Hahaha get wrecked ya surveilling fucks."的纸条。在 Suffolk，Sovern 拆除了支杆并拔走了线路、电池和太阳能板。 Greenview 有两台摄像头被剪落，Lisbon, Connecticut 有一台被砸毁。这一做法出现在蓝州、红州、城市和郊区，呈跨地域蔓延之势。

Flock Safety 在大约 6000 个美国社区部署了其以 AI 为基础的摄像头，能够扫描每一辆经过车辆的车牌，公司估值约为 75 亿美元。表面上宣称是为了社区安全，但实际上这一监控网络会向联邦移民执法提供数据。数据显示，地方与州警察对联邦移民目的进行了超过 4000 次查询，其中包括明确标注为"ICE"、"ICE+ERO"和"ICE WARRANT"的搜索。仅在 Virginia，警方在 12 个月内就对 Flock 网络进行了近 3000 次与移民相关的搜索。在 Washington state，至少有八个执法机构开启了与 U.S. Border Patrol 直接共享其 Flock 网络的功能。甚至有一个 Texas 学区的摄像头被来自多个州的 30 个执法机构为了移民目的进行过查询。

公众的愤怒还因他们觉得反对无效而加剧。在 La Mesa，市议会就续签 Flock 合约举行了场面爆满的会议，绝大多数出席者反对，但议会仍投票通过。 San Diego 也出现类似情形：大量居民出席并反对 Flock，官员却批准继续合作。当选民感到他们的担忧被忽视时，一些人就不再通过官方渠道表达意见。城市开始隐瞒摄像头位置：Louisville 为保密 Flock 摄像头位置而提起诉讼，称公开坐标可能危及公共安全，但更明显的担忧似乎是防止被破坏。 Norfolk 在 2025 年 12 月败诉，被迫公开 Hampton Roads 地区 600 个摄像头的位置。

Flock 的回应并未平息争议。 CEO Garrett Langley（38 岁）曾宣称大规模监控可消除美国所有犯罪。公司对破坏事件的官方声明称"我们尊重并重视对我们技术的担忧与反馈，建立信任对我们很重要"。但在城市隐瞒摄像头位置、警方在没有搜查令的情况下发起 ICE 搜索、学校摄像头被远在数州之外的机构查询、以及议会在压倒性反对声中仍续约的现实下，建立信任显得困难。 Sovern 的案件在 Reddit 上几乎获得普遍支持；当陪审团的舆论将你视为民间英雄时，对检控方来说是个难题。

这类破坏可能会继续，因为引发冲突的因素并未消失。尽管 Flock 否认，ICE 相关的查询仍在进行；城市仍然忽视公众反对；摄像头在 6000 多个社区持续扩展；逮捕行为往往带来所谓的"殉道者"。与此同时，Amazon 已终止与 Flock 的 Ring 合作，已有 46 个城市正式拒绝 Flock 摄像头，包括 Austin 、 Eugene 、 Mountain View 、 Santa Cruz 和 Alameda County 。政治风向正在变化，但对一些人来说，官方渠道推进得太慢了。争论的核心并非仅仅是 Flock 这个品牌，而是它所代表的监控基础设施：不顾公众反对仍不断扩展、声称本地化却向联邦数据库输送数据、挂着社区安全的幌子却助长驱逐机制。眼下，有人正用钳子将这一观点表达出来。 Across the United States, people are taking matters into their own hands when it comes to Flock Safety surveillance cameras. Since April 2025, at least 25 cameras have been destroyed in five states, from California to Virginia. The destruction isn't coordinated, it's just angry residents responding to what they see as an intrusive surveillance network with ties to federal immigration enforcement. One Virginia man, Jeffrey S. Sovern, faces 25 criminal charges for systematically destroying 13 cameras over six months, and he's been open about his motivations, citing the Fourth Amendment and hoping his case becomes a catalyst for rolling back intrusive surveillance.

The destruction has taken various forms. In La Mesa, California, two cameras were found destroyed on Fletcher Parkway weeks after the city council voted to keep them despite overwhelming public opposition. In Eugene and Springfield, Oregon, six cameras were cut down from poles, at least one spray-painted, with a note left behind reading "Hahaha get wrecked ya surveilling fucks." In Suffolk, Virginia, Sovern dismantled mounting poles and removed wiring, batteries, and solar panels. Two cameras were cut down in Greenview, Illinois, and one was smashed in Lisbon, Connecticut. The pattern spans blue states, red states, cities, and suburbs.

Flock Safety operates in approximately 6,000 U.S. communities with AI-powered cameras scanning every license plate that passes, and the company is valued at $7.5 billion. While the pitch is neighborhood safety, the reality is a surveillance network that feeds into federal immigration enforcement. Data shows more than 4,000 lookups by local and state police were conducted for federal immigration purposes, including searches explicitly tagged "ICE," "ICE+ERO," and "ICE WARRANT." In Virginia alone, police performed nearly 3,000 immigration-related searches on the Flock network over 12 months. In Washington state, at least eight law enforcement agencies enabled direct sharing of their Flock networks with U.S. Border Patrol. One Texas school district had cameras searched by 30 law enforcement agencies from multiple states for immigration purposes.

The anger is compounded by the feeling that public opposition doesn't matter. In La Mesa, the city council held a packed meeting about continuing its Flock contract where the overwhelming majority opposed the cameras, but the council voted to keep them anyway. San Diego had the same dynamic, huge turnout against Flock at council meetings, yet officials approved continuation. When elected officials tell constituents their concerns don't matter, some people stop using official channels. Cities are now hiding camera locations, with Louisville suing to keep Flock camera locations secret, claiming releasing them could compromise public safety, though the real concern appears to be vandalism. Norfolk lost a lawsuit in December 2025 that forced disclosure of 600 camera locations in Hampton Roads.

Flock's response hasn't helped. CEO Garrett Langley, 38, has claimed that mass surveillance could eliminate all crime in America. The company's official statement about the destruction said they "respect and value concerns and feedback raised about our technology, and building trust is important to us." But building trust is difficult when cities hide camera locations, police run ICE searches without warrants, school cameras get searched by agencies three states away, and council meetings with overwhelming opposition end in contract renewals. Sovern's case has generated near-universal support on Reddit, and when your jury pool thinks you're a folk hero, that's a problem for prosecutors.

The destruction is likely to continue because the triggers remain in place. ICE searches continue despite Flock's denials, cities keep ignoring opposition, cameras keep expanding across 6,000 communities and growing, and arrests create martyrs. Meanwhile, Amazon already killed its Ring-Flock partnership, and 46 cities have formally rejected Flock cameras, including Austin, Eugene, Mountain View, Santa Cruz, and Alameda County. The political winds are shifting, but for some people, the official channels are too slow. This isn't really about Flock specifically, it's about what Flock represents: surveillance infrastructure that expands regardless of public opposition, claims to be local while feeding federal databases, and promises community safety while enabling deportation machinery. Right now, that argument is being made with vice grips.

Semble 是一款为 AI 代理量身打造的代码搜索库，能快速且准确地从代码仓库中检索相关代码片段。它解决了 grep 等传统方法和全文读取带来的低效问题——这些方法会消耗大量 tokens 并增加延迟。通过只返回必要的精确代码片段，Semble 比基于 grep 的方案节省约 98% 的 tokens，同时维持高质量的检索效果。

该工具完全在 CPU 上运行，无需 API 密钥、 GPU 或外部服务，轻量且易于部署。性能表现优异：平均对一个仓库进行索引约需 250 毫秒，查询响应约 1.5 毫秒。在覆盖 19 种语言、 63 个仓库的 1,250 次查询对比测试中，Semble 的 NDCG@10 达到 0.854，检索质量与专用 transformer 模型相当，但在索引和查询速度上明显更快。

Semble 使用 Chonkie 将文件按代码语义拆分为片段；在评分时采用两种互补方法：通过 Model2Vec embeddings（potion-code-16M 模型）进行语义相似度匹配，和通过 BM25 对标识符与 API 名称做词汇匹配。两者得分以 Reciprocal Rank Fusion 合并，随后依据多种代码感知信号重排序，包括基于查询类型的自适应权重、针对符号查询的定义提升、标识符词干匹配、文件连贯性加分，以及对测试文件和遗留代码的噪声惩罚。

该库提供多种集成方式：可作为 MCP 服务器运行，兼容 Claude Code 、 Cursor 、 Codex 、 OpenCode 等 MCP 兼容代理，便于在开发环境内直接搜索代码库；也可通过 AGENTS.md 或 CLAUDE.md 的 bash 集成，让无法访问 MCP 的子代理仍能调用其能力；此外还有独立 CLI，便于脚本化和命令行使用。

主要特性包括支持本地路径与远程 git 仓库、本地目录的自动文件监控与重索引，以及用于估算搜索节省量的 token 节省追踪器。提供 Python API 以便程序化集成自定义工具。 Semble 在 MIT 许可证下开源，自发布以来在 GitHub 上已获得超过 1,100 个 star，受到广泛关注。 Semble is a code search library designed specifically for AI agents, providing fast and accurate retrieval of relevant code snippets from repositories. It addresses the inefficiency of traditional methods like grep and full-file reading, which consume excessive tokens and latency. By returning only the precise chunks of code needed, Semble uses approximately 98% fewer tokens than grep-based approaches while maintaining high retrieval quality.

The tool operates entirely on CPU without requiring API keys, GPUs, or external services, making it lightweight and easy to deploy. It achieves impressive performance benchmarks, indexing an average repository in about 250 milliseconds and answering queries in roughly 1.5 milliseconds. In comparative testing across 1,250 queries over 63 repositories in 19 languages, Semble achieved an NDCG@10 score of 0.854, matching the retrieval quality of specialized transformer models while being significantly faster at both indexing and query time.

Semble works by splitting files into code-aware chunks using Chonkie, then scoring queries against these chunks through two complementary methods: semantic similarity via Model2Vec embeddings using the potion-code-16M model, and lexical matching via BM25 for identifiers and API names. These scores are combined using Reciprocal Rank Fusion, followed by reranking with code-aware signals such as adaptive weighting based on query type, definition boosts for symbol queries, identifier stem matching, file coherence bonuses, and noise penalties for test files and legacy code.

The library offers multiple integration paths for different use cases. It can run as an MCP server compatible with Claude Code, Cursor, Codex, OpenCode, and other MCP-compatible agents, allowing direct codebase searching from within development environments. Alternatively, it provides bash integration through AGENTS.md or CLAUDE.md configuration files, enabling sub-agents that cannot access MCP tools to still leverage its capabilities. A standalone CLI is also available for scripting and direct command-line usage.

Key features include support for both local paths and remote git repositories, automatic file watching and re-indexing for local directories, and a savings tracker that estimates token savings across searches. The Python API allows programmatic access for custom tooling integration. Semble is open source under the MIT license and has gained significant traction with over 1,100 GitHub stars since its release.

一位名为 Night­mare‑Eclipse 的安全研究人员发布了名为 YellowKey 的工具，声称可以完全绕过 BitLocker 的整盘加密。该利用方法是把特定的 "FsTx" 文件夹复制到以 NTFS 、 FAT32 或 exFAT 格式化的 USB 驱动器上，将受 BitLocker 保护的机器重启到 Windows Recovery Environment，然后按一套特定的输入序列操作。若步骤无误，会弹出一个命令行 shell，从而在不需要任何密码的情况下无限制访问加密卷。

Nightmare‑Eclipse 认为，这个问题看起来不像普通的程序缺陷，更像是 Microsoft 在 BitLocker 中刻意留下的后门。其理由是触发问题的组件仅存在于官方的 Windows Recovery Environment 镜像中，并且在运行系统上与标准 Windows 安装镜像的表现不同。此外，该问题只影响 Windows 11 和 Server 2022/2025，而 Windows 10 不受影响，研究者认为这点很难用偶然解释，更倾向于故意设计。

据报道，第三方研究人员已验证 YellowKey 与公开的 GitHub 资料描述一致。 Nightmare‑Eclipse 还发布了第二个名为 GreenPlasma 的提权工具，声称能实现权限提升，但并未公开可直接拿来获得 SYSTEM 级访问的完整概念验证代码。研究者表示可能会在下个月的 Patch Tuesday 前披露更多细节。

Nightmare‑Eclipse 有过针对 Microsoft 的公开对抗历史，曾以 Chaotic Eclipse 名义活动并发布过 Red Sun 等漏洞，且声称 Microsoft 损害了其职业和声誉。研究者一贯的对抗性做法为其关于 YellowKey 是刻意后门的大胆结论提供了背景说明。

对于担心 BitLocker 安全性的人，缓解建议是不要仅依赖单一加密方案。安全专业人士建议评估经过良好审查的整盘加密替代方案，如 VeraCrypt 。这一事件也凸显了独立安全研究者与大型科技公司在漏洞披露和主流安全工具信任度方面持续存在的紧张关系。 A security researcher known as "Nightmare-Eclipse" has released a tool called YellowKey that allegedly bypasses BitLocker's full-volume encryption entirely. The exploit works by copying a specific "FsTx" folder to a USB drive formatted with NTFS, FAT32, or exFAT, rebooting a BitLocker-protected machine into the Windows Recovery Environment, and following a particular sequence of inputs. If done correctly, a command shell appears granting unrestricted access to encrypted volumes without requiring any passwords.

Nightmare-Eclipse argues this vulnerability looks less like a typical bug and more like an intentional backdoor built into BitLocker by Microsoft. Their reasoning centers on the fact that the component triggering the issue only exists in the official Windows Recovery Environment image and behaves differently on live systems compared to standard Windows installation images. The researcher also notes that only Windows 11 and Server 2022/2025 are affected, while Windows 10 is not, which they find suspicious and difficult to explain as anything other than deliberate design.

Third-party researchers have reportedly confirmed that YellowKey works as described in the public GitHub materials. Nightmare-Eclipse also released a second exploit called GreenPlasma, said to enable privilege escalation, though they did not publish full proof-of-concept code for achieving SYSTEM-level access. The researcher suggested they may disclose further details ahead of next month's Patch Tuesday.

Nightmare-Eclipse has a history of targeting Microsoft and accusing the company of hostility toward external security researchers. Previously operating as "Chaotic Eclipse," they released other vulnerabilities like Red Sun while claiming Microsoft damaged their career and reputation. The researcher's confrontational approach adds context to the bold claims about YellowKey being an intentional backdoor.

For those concerned about BitLocker's security, mitigation advice centers on not relying solely on any single encryption system. Security professionals recommend evaluating well-reviewed full-disk encryption alternatives such as VeraCrypt. The situation highlights ongoing tensions between independent security researchers and large technology companies over vulnerability disclosure and the integrity of widely-used security tools.

世界卫生组织周六宣布，蔓延刚果民主共和国和乌干达的埃博拉疫情构成"国际关注的突发公共卫生事件"。就在前一天，非洲主要公共卫生机构报告了刚果东北部与数十例疑似死亡相关的疫情。截至周六，两国首都金沙萨和坎帕拉均已出现确诊病例。伊图里省是最先发现疫情的地区，报告有 246 例疑似病例和 80 例死亡，但仅有 8 例经实验室确证。此次疫情涉及埃博拉病毒邦迪布约型，目前尚无获批疫苗或特效治疗。世卫组织指出，疫情真实规模存在很大不确定性，实际病例数可能远高于已检测和上报的数字。

将疫情列为"国际关注的突发公共卫生事件"意味着需要协调的国际应对，旨在促使各国为可能的传播作准备并共享疫苗、治疗手段与资源。然而，应对工作面临严峻挑战，尤其是近期美国若干关键卫生机构经费或功能被削减。报道指出，美国国际开发署在遏制以往疫情中发挥过重要作用，但去年被特朗普政府关闭；政府还削减了对疾病控制与预防中心的经费，并于一月宣布退出世卫组织。全球卫生专家担忧，此次疫情的初次报告可能比实际爆发时间滞后，通常世卫组织或其他机构会更早发现病例激增。

伊图里省的政治局势使防控更加复杂。总统费利克斯·齐塞克迪在 2021 年因包括 Allied Democratic Forces 和 Cooperative for the Development of Congo militia 在内的反叛组织发动袭击而宣布该省进入戒严，实际上实施了军事管制，限制了基本自由，也削弱了民众对当局在卫生与基本服务方面的信任。平民遭受严重暴力侵害，再加上美国国际开发署经费削减导致部分诊所被迫关闭，局势进一步恶化。该省毗邻乌干达和南苏丹，频繁的跨境流动构成重大挑战；此外，乌干达定于 6 月 3 日的天主教节日预计将吸引来自刚果的信徒前往朝圣。

人道主义危机、高度的人口流动性和庞大的非正规医疗网络，都加大了病毒传播风险。遏制埃博拉需要迅速的公共卫生行动，包括接触者追踪、及时隔离治疗和安全的遗体处理。乌干达卫生部长表示，该国拥有一支通过电子系统开展挨家挨户监测的社区卫生工作者队伍。世卫组织于 5 月 5 日接到疑似病例报告并派出调查小组，但初步现场样本呈阴性，因为现场检测设备只能识别扎伊尔型病毒；随后，位于金沙萨的国家生物医学研究所于周四确认了埃博拉阳性病例。

历史上曾发生过多次埃博拉疫情，最严重的是 2014–2015 年西非暴发，造成超过 1.1 万名人员死亡。此后，刚果和乌干达多次通过迅速的公共卫生应对成功控制了疫情。无国界医生组织表示，正准备在伊图里省迅速扩大医疗行动，对短期内出现大量病例和死亡、以及疫情跨区跨境扩散表示严重关切。疫情蔓延到首都将带来额外挑战，因为在人口密集的城市环境中传染病传播速度更快。 The World Health Organization declared the Ebola outbreak spreading across the Democratic Republic of Congo and Uganda a global health emergency on Saturday. The announcement came just one day after Africa's leading public health authority reported an outbreak in northeast Congo linked to dozens of suspected deaths. By Saturday, cases had been confirmed in the capital cities of both countries. Kinshasa and Kampala. In Congo's Ituri province, where the outbreak was first identified, there were 246 suspected cases and 80 deaths, though only eight cases had been confirmed through laboratory testing. This outbreak involves the Bundibugyo species of Ebola, for which there is no approved vaccine or therapeutics. The W.H.O. noted significant uncertainties about the true scale of the outbreak, cautioning that the actual numbers could be far larger than what has been detected and reported.

The declaration of a "public health emergency of international concern" signals a public health risk requiring coordinated international response, intended to prompt countries to prepare for potential spread and share vaccines, treatments, and resources. However, the response faces significant challenges, particularly given recent cuts to key U.S. health agencies. The U.S. Agency for International Development, which played a major role in containing previous outbreaks, was shuttered by the Trump administration last year. The administration also cut funding for the Centers for Disease Control and Prevention and withdrew from the W.H.O. in January. Global health experts expressed concern that the first reports of this outbreak emerged late in its development, as a surge in cases is typically detected much earlier by the W.H.O. or other organizations.

The political situation in Ituri province complicates containment efforts. President Félix Tshisekedi declared a state of siege in the province in 2021 following attacks by insurgent groups, including the Allied Democratic Forces and the Cooperative for the Development of Congo militia. This effectively imposes martial law, limiting fundamental freedoms and eroding public trust in authorities regarding health and basic services. Civilians have been exposed to extreme violence, and the state's failure to protect them has been compounded by U.S.A.I.D. cuts that forced some clinics to close. The province borders Uganda and South Sudan, making frequent cross-border movement a major challenge, with Uganda's upcoming Catholic festival on June 3 expected to attract worshipers from Congo.

The risk of spreading is exacerbated by this humanitarian crisis, high population mobility, and a large network of informal healthcare facilities. Containing Ebola requires rapid public health response, including contact tracing, prompt isolation and treatment, and safe burial practices. Uganda's health minister stated the country has a strong system of community health workers who conduct door-to-door surveillance through an electronic health system. The W.H.O. was notified about suspected cases on May 5 and sent a team to investigate, but initial samples tested negative because field equipment could only detect the Zaire species. Later, the National Institute of Biomedical Research in Kinshasa confirmed positive cases for Ebola on Thursday.

This follows previous Ebola epidemics, such as the devastating 2014-2015 outbreak in West Africa that killed over 11,000 people. Since then, outbreaks in Congo and Uganda have been contained due to quick public health responses. Doctors Without Borders is preparing to rapidly scale up its medical response in Ituri, expressing extreme concern over the number of cases and deaths in such a short time, combined with the spread across multiple health zones and across borders. The spread to the capitals poses an additional challenge, as infectious diseases can spread more rapidly in dense urban settings.

rkdebian 是一个构建系统，用于为搭载 Rockchip RK3562 处理器的 Doogee U10 平板生成完整的 Debian 12 Bookworm 镜像。整个系统完全从零逆向实现，未依赖厂商文档或官方支持，借助 Claude 、 Codex 等 AI 工具并以开源仓库为起点。生成的镜像可从 SD 卡启动，不会改动平板内置的 Android 存储，也无需解锁引导加载程序。

Doogee U10 采用 Rockchip RK3562 SoC，配备四核 Cortex-A53 、 4 GB 内存和 10.1 英寸显示屏。项目支持广泛的硬件功能，包括显示与触摸屏、 Wi‑Fi 、蓝牙、音频与麦克风、加速度计、闪光灯和电池管理。还提供通过 Panfrost 的 3D 加速（实验性）和可用的摄像头管线，但色彩调教仍需优化。一个显著功能是启用了 RK3562 的 NPU 核心，借助 Rockchip 的 RKLLM 栈实现本地 LLM 推理；基准显示 Qwen3-0.6B 模型生成速度约为每秒 5 个 token 。

构建系统通过环境变量和命令行参数高度可配置，允许选择不同的 GPU 栈（Mali 或 Panfrost）、显示服务（Wayland 或 X11）、 CPU 调速策略和 UI 会话（如 Phosh）。系统包含多种体积优化选项，并支持设备上的 OTA 更新，无需重刷 SD 卡。项目还实现了若干 Phosh 定制集成，如电源模式映射、后置 LED 的原生手电筒切换、正确的电源键行为以及锁屏方向记忆。

仓库中包含预编译的 Mali GPU 二进制、 Rockchip MPP 包和 Seekwave 的 Wi‑Fi/ 蓝牙驱动，并对各自来源做了适当标注。构建过程会生成采用 GPT 分区的镜像，分别为引导加载程序、 U‑Boot 、内核和根文件系统创建独立分区，首次启动时会自动扩展以填满 SD 卡。镜像提供标准用户和 root 的默认凭据，强烈建议首次启动后立即修改密码。 This project, called rkdebian, is a build system that creates a complete Debian 12 Bookworm image for the Doogee U10 tablet, which uses the Rockchip RK3562 processor. The entire system was reverse-engineered from scratch without any official vendor documentation or support, using AI tools like Claude and Codex alongside open-source repositories as a starting point. The resulting image boots from an SD card, leaving the tablet's internal Android storage completely untouched and requiring no bootloader unlock.

The Doogee U10 features a Rockchip RK3562 SoC with four Cortex-A53 cores, 4 GB of RAM, and a 10.1-inch display. The project supports a wide range of hardware features including the display, touchscreen, Wi-Fi, Bluetooth, audio, microphone, accelerometer, flashlight, and battery management. It also includes experimental support for 3D acceleration via Panfrost and functional camera pipelines, though color tuning still needs calibration. A notable feature is the activation of the RK3562's NPU core for local LLM inference using Rockchip's RKLLM stack, with benchmarks showing the Qwen3-0.6B model running at about 5 tokens per second for generation.

The build system is highly configurable through environment variables and CLI flags, allowing users to select different GPU stacks (Mali or Panfrost), display servers (Wayland or X11), CPU governors, and UI sessions (Phosh). It includes various size optimization options and supports OTA updates on the device without reflashing the SD card. The project also implements several Phosh-specific integrations like power mode mapping, native torch toggle for the rear LED, proper power button behavior, and lock screen orientation memory.

The repository includes pre-built Mali GPU binaries, Rockchip MPP packages, and Seekwave Wi-Fi/Bluetooth drivers, with proper attribution to their respective sources. The build process produces a GPT-partitioned image with separate partitions for the bootloader, U-Boot, kernel, and root filesystem, which automatically expands to fill the SD card on first boot. Default credentials are provided for both a standard user and root account, with strong recommendations to change passwords on first boot.

John Gruber 对 Steven Levy 的观点提出了异议。 Levy 认为 Apple 的下一任 CEO 需要推出一款"杀手级 AI 产品"，像 iPhone 那样定义 AI 时代。 Gruber 认为 Levy 被 AI 炒作带偏了；Apple 一向不是单纯发布底层技术，而是打造出色的产品和体验。 iPod 不是关于 MP3 或微型硬盘，而是关于听音乐。 iPhone 定义了移动时代，但 Apple 不必在那个时代衍生出的每个市场（比如社交媒体）都称霸，才能保持成功。

Gruber 对 Levy 预测 AI 代理会很快取代 iPhone 生态、主动处理叫车等事务持怀疑态度。他把这种预测称为"被炒作冲昏头脑的幻想"，质疑这样的系统如何运行、依赖什么硬件，以及人们是否会觉得它是便利而非令人不适。他怀疑到 2030 年人们会放弃用手机打车，无论是语音还是传统点按。

他认为用手表、耳机或眼镜等更小的设备来替代手机进行 AI 交互并不合理，因为这些设备很可能仍需与随身携带的手机配对。 Apple 在制造小型个人计算设备方面已是世界一流，但手机在相机、屏幕和日常使用中仍居核心地位。放弃手机转而使用独立小型设备的想法难以成立。

Gruber 的核心观点是：AI 是一项技术，而不是独立的产品类别。他将其比作无线网络——Wi‑Fi 、蜂窝和蓝牙已渗透到 Apple 的所有产品中，但它们本身并不是某个单一的"杀手级产品"。 AI 同样会融入所有产品，而不会靠某一台设备来定义时代。 Apple 无法忽视 AI，因为它无处不在，但这并不意味着必须做出一款革命性的 AI 产品才能保持竞争力。 John Gruber pushes back against Steven Levy's argument that Apple's next CEO needs to launch a "killer AI product" to define the AI age the way the iPhone defined the mobile era. Gruber finds Levy's reasoning to be caught up in AI hype, arguing that Apple's philosophy has never been about shipping raw technology but about shipping great products and experiences. The iPod wasn't about MP3 files or tiny hard drives, it was about music. The iPhone defined mobile, but Apple doesn't need to dominate every market that era opened up, like social media, to remain successful.

Gruber is particularly skeptical of Levy's prediction that AI agents will soon replace the iPhone ecosystem by proactively handling tasks like summoning ride-shares without any user friction. He calls this a "fever dream high-on-the-hype fantasy," questioning the practicalities of how such a system would work, what hardware would support it, and whether people would even find it desirable rather than creepy. He doubts that by 2030, people will be using anything other than their phones to hail rides, whether through voice commands or traditional tapping.

Gruber argues that the idea of smaller devices like watches, earbuds, or glasses replacing phones for AI interactions doesn't make sense, since those devices would likely still pair with the phone you're carrying anyway. He points out that Apple is already the best in the world at making smaller personal computing devices, yet the phone remains central for cameras, screens, and general use. The notion that we'd abandon the phone for independent smaller devices seems implausible to him.

The core of Gruber's argument is that AI is technology, not a product category. He compares it to wireless networking, which is now pervasive across all Apple devices but isn't itself a standalone "killer product." Just as Wi-Fi, cellular, and Bluetooth permeate everything Apple makes without being a single flagship offering, AI will similarly be integrated into all products rather than being one defining device. Apple can't ignore AI because it's pervasive, but that doesn't mean it needs to create one revolutionary AI product to stay relevant.