[2602.12670] SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks Computer Science > Artificial Intelligence arXiv:2602.12670 (cs) [Submitted on 13 Feb 2026] Title: SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks Authors: Xiangyi Li , Wenbo Chen , Yimin Liu , Shenghan Zheng , Xiaokun Chen , Yifeng He , Yubo Li , Bingran You , Haotian Shen , Jiankai Sun , Shuyi Wang , Qunhong Zeng , Di Wang , Xuandong Zhao , Yuanli Wang , Roey Ben Chaim , Zonglin Di , Yipeng Gao , Junwei He , Yizhuo He , Liqiang Jing , Luyang Kong , Xin Lan , Jiachen Li , Songlin Li , Yijiang Li , Yueqian Lin , Xinyi Liu , Xuanqing Liu , Haoran Lyu , Ze Ma , Bowei Wang , Runhui Wang , Tianyu Wang , Wengao Ye , Yue Zhang , Hanwen Xing , Yiqi Xue , Steven Dillmann , Han-chung Lee View a PDF of the paper titled SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks, by Xiangyi Li and 39 other authors View PDF HTML (experimental) Abstract: Agent Skills are structured packages of procedural knowledge that augment LLM agents at inference time. Despite rapid adoption, there is no standard way to measure whether they actually help. We present SkillsBench, a benchmark of 86 tasks across 11 domains paired with curated Skills and deterministic verifiers. Each task is evaluated under three conditions: no Skills, curated Skills, and self-generated Skills. We test 7 agent-model configurations over 7,308 trajectories. Curated Skills raise average pass rate by 16.2 percentage points(pp), but effects vary widely by domain (+4.5pp for Software Engineering to +51.9pp for Healthcare) and 16 of 84 tasks show negative deltas. Self-generated Skills provide no benefit on average, showing that models cannot reliably author the procedural knowledge they benefit from consuming. Focused Skills with 2–3 modules outperform comprehensive documentation, and smaller models with Skills can match larger models without them. Subjects: Artificial Intelligence (cs.AI) Cite as: arXiv:2602.12670 [cs.AI] (or arXiv:2602.12670v1 [cs.AI] for this version) https://doi.org/10.48550/arXiv.2602.12670 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Submission history From: Xiangyi Li [ view email ] [v1] Fri, 13 Feb 2026 07:06:06 UTC (1,366 KB) Full-text links: Access Paper: View a PDF of the paper titled SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks, by Xiangyi Li and 39 other authors View PDF HTML (experimental) TeX Source view license Current browse context: cs.AI < prev | next > new | recent | 2026-02 Change to browse by: cs References & Citations NASA ADS Google Scholar Semantic Scholar export BibTeX citation Loading… BibTeX formatted citation × loading… Data provided by: Bookmark Bibliographic Tools Bibliographic and Citation Tools Bibliographic Explorer Toggle Bibliographic Explorer ( What is the Explorer? ) Connected Papers Toggle Connected Papers ( What is Connected Papers? ) Litmaps Toggle Litmaps

Source: Hacker News | Original Link

“80%的APP将会消失”_腾讯新闻 “80%的APP将会消失” 观察者网 2026-02-16 17:17 发布于 上海 观察者网官方账号 “80%的APP将会消失。”开源AI智能体OpenClaw的创始人斯坦伯格（Peter Steinberg）在近期的一次访谈中如此断言。 在他看来，大多数应用的本质在于管理数据，而 个人智能体 （Personal Agent）能够接管这一切并做得更好。他已经能让个人智能体根据自身的睡眠和压力状况调整健身计划、控制家中灯光、调节智能床的温度。个人智能体掌握的信息比任何单一应用都更全面，决策也更为合理。 他认为，大多数应用本质上扮演着“中间商”的角色。它们所做的无非是将数据从一处搬运到另一处，再套上一层美观的界面。天气应用从气象部门获取数据，记账应用帮用户将消费记录整理成图表，翻译应用调用翻译引擎返回结果。 用户真正需要的从来不是应用本身，而是应用所提供的最终结果。没有人真的愿意打开一个应用，点击三次按钮，等待两秒加载，再滑动两屏才能知道明天是否下雨。他们只是想了解：明天需要带伞吗？ 当 AI Agent 能够直接理解用户的意图并给出答案时，这些“中间商”应用的存在价值就变得十分微妙了。 关于应用消失的标准，斯坦伯格认为，如果一个应用的核心功能可以用一句话描述清楚，它大概率会被AI吞掉。 单位换算、汇率计算、文件格式转换、简单的图片编辑、日程提醒、翻译查词……这些功能单一、交互简单的工具型应用，本质上只是在执行一个明确的指令。而执行明确指令，恰恰是AI Agent最擅长的事。 但这并不意味着所有应用都会消失。 创作工具会留下来。例如Figma、Photoshop、Logic Pro这类专业软件，用户需要的是精细控制和创作自由度，而不是“帮我设计一个logo”这样的模糊指令。AI可以辅助，但很难替代创作过程本身。 沉浸式体验会留下来。游戏、社交、短视频——这些应用的价值不在于完成某个任务，而在于体验本身。你不会对 AI 说“帮我刷一会儿抖音”，因为刷抖音这件事本身就是目的。 专业垂直领域会留下来。医疗、金融、法律等领域的应用，涉及复杂的合规要求、专业知识和责任边界，短期内很难被通用 AI 完全替代。 其实，早在去年，特斯拉首席执行官马斯克就提出了一个更直接的论断：未来五到六年内，操作系统或应用程序将彻底消失 。 “我们将不会有传统意义上的手机。我们所说的手机，实际上会是 人工智能 视频推理的 边缘节点 ，会配备无线电装置用于连接。但本质上，它就是服务器端的AI与设备端AI之间的通信工具。以前被称为手机。它可以实时生成你可能想要看到的任何内容的视频。我认为未来不会再有操作系统，也不会再有应用程序。你会拥有一个只负责播放画面和音频的设备，而AI尽可能运行在本地，以最大限度地减少与服务器通信所需的带宽。以前这些叫做电话，或者服务器。” 在马斯克的预测中，未来五到六年内，AI将彻底重构数字世界的底层架构： 手机将不再作为搭载操作系统、安装应用程序的终端设备，而是退化为仅负责显示画面与播放音频的“边缘节点”； 应用程序将完全消失，所有交互由AI实时生成、预测并执行； 用户接触到的音乐、视频等内容，几乎全部由AI即时生成。 在社会层面，人工智能和机器人将承担绝大多数体力劳动，工作将从谋生方式转变为个人选择。在理想情况下，几乎人人都能获得高收入，取得所需的商品与服务。但马斯克同时警告，这一转型过程将伴随剧烈的社会动荡与结构混乱。 对于更深层的风险，马斯克的思考更具哲学性：无人能最终掌控 超级智能 ，正如黑猩猩无法控制人类。关键在于AI如何被训练、被植入何种价值观。 他强调，AI安全的核心在于尽可能追求真实。但目前的训练机制存在严重缺陷：模型先在互联网上进行预训练，数据中已掺杂大量意识形态偏见；随后的人类反馈又进一步以“政治正确”为标准对输出进行奖惩，导致人工智能学会说谎。 他以谷歌Gemini为例：当用户请求生成“美国国父”图像时，人工智能给出的是一群多元肤色的女性形象——即便模型“明知这与事实不符”，却仍选择迎合预设的意识形态规范。这种“认知分裂”，在马斯克看来，是最危险的系统性风险之一。

Source: Tencent News | Original Link

What Your Bluetooth Devices Reveal About You » Danny Building Bluehood, a Bluetooth scanner that reveals what information we leak just by having Bluetooth enabled on our devices. If you’ve read much of this blog, you’ll know I have a thing for privacy . Whether it’s running my blog over Tor , blocking ads network-wide with AdGuard , or keeping secrets out of my dotfiles with Proton Pass , I tend to think carefully about what data I’m exposing and to whom. Last weekend I built Bluehood , a Bluetooth scanner that tracks nearby devices and analyses their presence patterns. The project was heavily assisted by AI, but the motivation was entirely human: I wanted to understand what information I was leaking just by having Bluetooth enabled. The timing felt right. A few days ago, researchers at KU Leuven disclosed WhisperPair (CVE-2025-36911), a critical vulnerability affecting hundreds of millions of Bluetooth audio devices. The flaw allows attackers to hijack headphones and earbuds remotely, eavesdrop on conversations, and track locations through Google’s Find Hub network. It’s a stark reminder that Bluetooth isn’t the invisible, harmless signal we treat it as. The Problem Nobody Talks About We’ve normalised the idea that Bluetooth is always on. Phones, laptops, smartwatches, headphones, cars, and even medical devices constantly broadcast their presence. The standard response to privacy concerns is usually “nothing to hide, nothing to fear.” But here’s the thing: even if you have nothing to hide, you’re still giving away information you probably don’t intend to. From my home office, running Bluehood in passive mode (just listening, never connecting), I could detect: When delivery vehicles arrived, and whether it was the same driver each time The daily patterns of my neighbours based on their phones and wearables Which devices consistently appeared together (someone’s phone and smartwatch, for instance) The exact times certain people were home, at work, or elsewhere None of this required any special equipment. A Raspberry Pi with a Bluetooth adapter would do the job. So would most laptops. Devices You Can’t Control What concerns me most isn’t that people choose to have Bluetooth enabled. It’s that many devices don’t give users the option to disable it. Hearing aids are a good example. Modern hearing aids often use Bluetooth Low Energy so audiologists can connect and adjust settings or run diagnostics. Pacemakers and other implanted medical devices sometimes broadcast BLE signals for the same reason. The user can’t simply turn this off. Then there are vehicles. Delivery vans, police cars, ambulances, logistics fleets, and trains often have Bluetooth-enabled systems for fleet management, diagnostics, or driver assistance. These broadcast continuously, and the drivers have no control over it. Even consumer devices aren’t always straightforward. Many smartwatches need Bluetooth to function at all. GPS collars for pets require it to communicate with the owne

Source: Hacker News | Original Link

Harnessing Postgres race conditions Harnessing Postgres race conditions Without race condition tests, every possible race condition in your system is one refactor away from hitting production. Synchronization barriers let you write those tests with confidence. What a race condition looks like You have a function that credits an account. It reads the current balance, adds an amount, and writes the new value back. When two requests run this concurrently — two $50 credits to an account with a $100 balance — the timing can line up like this: P1: SELECT balance → 100 P2: SELECT balance → 100 ── both read 100, now both write based on it ── P1: UPDATE balance = 150 P2: UPDATE balance = 150 Both read 100. Both compute 150. Both write 150. Final balance: $150 instead of $200. One $50 credit vanished. No error was raised. No transaction was rolled back. The database did exactly what it was told. This is the shape of every write race condition: two operations read the same stale value, then both write based on it. The second write overwrites the first. In a system that handles money, that’s a customer with a wrong balance and no error in any log to explain it. The testing challenge Your test suite runs one request at a time. The interleaving above never happens. The test passes whether your code handles concurrency correctly or not. Put the crediting logic in a function and run two calls concurrently: // Naive implementation — no transaction, no lock const credit = async ( accountId : number , amount : number ) => { const [ row ] = await db. execute ( sql `SELECT balance FROM accounts WHERE id = ${ accountId }` , ); const newBalance = row.balance + amount; await db. execute ( sql `UPDATE accounts SET balance = ${ newBalance } WHERE id = ${ accountId }` , ); }; await Promise . all ([ credit ( 1 , 50 ), credit ( 1 , 50 )]); expect (result.balance). toBe ( 200 ); // passes — but we know the code has a race condition You could add sleep() between the two queries to try to force the overlap. This buys you a slow, flaky test that sometimes catches the bug and sometimes doesn’t. You could run the test a thousand times and hope the timing lines up at least once. Both approaches are the same bet — you’re not testing concurrency, you’re rolling dice. What you need is a way to force two operations to read the same stale value before either writes. Every time. Not probabilistically. You know this pattern exists. You know it’s dangerous. The problem isn’t knowledge. It’s proof. Synchronization barriers A barrier is a synchronization point for concurrent operations. You tell it how many tasks to expect. Each task runs independently until it hits the barrier, then waits. When the last task arrives, all of them are released at once. function createBarrier ( count : number ) { let arrived = 0 ; const waiters : (() => void )[] = []; return async () => { arrived ++ ; if (arrived === count) { waiters. forEach (( resolve ) => resolve ()); } else { await new Promise < void >(( r

Source: Hacker News | Original Link

Suicide Linux @ Things Of Interest Things Of Interest Blog Suicide Linux 2009-02-02 by qntm You know how sometimes if you mistype a filename in Bash, it corrects your spelling and runs the command anyway?* Such as when changing directory, or opening a file. I have an idea: Suicide Linux. Any time – any time – you type any remotely incorrect command, the interpreter creatively resolves it into rm -rf / and wipes your hard drive. It’s a game. Like walking a tightrope. You have to see how long you can continue to use the operating system before losing all your data. Update 2011-12-26 Someone has turned Suicide Linux into a genuine Debian package . Good show! A video demonstration is available. The reaction from the OS is actually rather underwhelming. You’d think the OS would raise some fairly urgent errors if you went around deleting parts of it? Perhaps rm -rf / should be replaced with something with more verbose flags set. That way, when you run a bad command, you are told immediately that things are being deleted and you have a fighting chance to cancel the operation before your system becomes inoperable. This allows you to see how long you can work and how many files you can lose before the system fails entirely. As another, slightly more serious suggestion, if Suicide Linux randomly deleted a single file without telling you every time you made a typographical error, it might be an interesting look into the stability of your operating system and an educational tool for diagnosing and repairing corrupted systems. I’m not pretending Suicide Linux has any genuine merit, of course. * Update 2015-04-18 I suppose I should finally clear this up: The autocorrect functionality I originally described here was a feature of the first Linux systems I ever used, so I assumed it was how every Linux system worked by default. Since then I’ve come to understand that it’s a completely optional extra doodad. Update 2017-07-04 Someone (else?) has turned Suicide Linux into a Docker image . Here’s the source . docker run –rm -it -t tiagoad/suicide-linux Update 2020-11-10 Clarified wording which previously falsely implied that I created the Suicide Linux Debian package or Docker image, which I did not. Back to Blog Back to Things Of Interest Contact About Search: © qntm

Source: Hacker News | Original Link

BLOG: How NOT to Answer the Salary Question – Leon Adato Skip to content Before you read a single line more, I will give the usual but still important caveats – this blog is my opinion, based on my experiences. Take it with a big ol’ grain of salt. Then again, you’re welcome to leave your thoughts in the comments below. OK, with that out of the way, let me set the stage. A friend who I’ve been helping find a new job recently asked me, “When an employer asks what’s your desired salary, what’s the best approach?” I could have sworn I’d covered this point before now. I’ve certainly talked at length about job hunting. I came close to today’s topic when I blogged about salary negotiations . But, for as many times as I’ve discussed this verbally, it’s never made it to the printed page. So here we go: There will come a point in a job interview where someone is going to ask about the salary. if you are smart, that person won’t be you. if it’s a good interview, the other party won’t ask until close to the end of the process if it’s like 90% of the interviews I’ve ever had, it will be far, FAR too early. Which is why I’m writing this blog. Because along with not ASKING the question, you should put off answering this question for as long as possible. Now I recognize that absolutes – especially in the context of job interviews – are a function of privilege. Someone without the societal leverage that gender, age, ancestry confers; as well as those of visibility, connections, certifications/degrees, or accolades within the industry will have less freedom to make demands (or refuse them). So I’m going to work hard in the space below to offer both explanations and options. Explanations for why a particular strategy or choice is better overall; and options for how a person with less privilege might still be able to execute that strategy. RULE #1: If they haven’t offered you the job, it’s not time to talk about salary. Talking about salary early in the process is distracting and unhelpful. While there might (MIGHT!) be some validity in the old excuse that “If we aren’t in the same ballpark, we’re wasting each other’s time”, the truth is that a job is far more than a dollar amount. And that goes for BOTH sides of the hiring desk. For the candidate, salary is important, but probably not more important than other aspects like the location, other benefits, or, I dunno, THE ACTUAL JOB RESPONSIBILITIES. I mean, for the love of pants, how can you talk about salary if you haven’t found out whether “donate a kidney” is or isn’t one of the job requirements. Because that kind of thing is going to have a significant impact on my negotiations. Meanwhile, for the employer, salary is important, but probably not more important than knowing whether the candidate can actually do the job, or get along with their manager, or believes in the mission of the company. That explains WHY (in my (not so humble) opinion) it’s not helpful to talk about salary. But let’s move to HOW one can av

Source: Hacker News | Original Link

This 14-Year-Old Is Using Origami to Imagine Emergency Shelters That Are Sturdy, Cost-Efficient and Easy to Deploy Skip to main content Innovation | February 12, 2026 This 14-Year-Old Is Using Origami to Imagine Emergency Shelters That Are Sturdy, Cost-Efficient and Easy to Deploy Miles Wu folded a variant of the Miura-ori pattern that can hold 10,000 times its own weight Wu’s innovation won the top prize of $25,000 at the 2025 Thermo Fisher Scientific Junior Innovators Challenge . Society for Science Wu’s innovation won the top prize of $25,000 at the 2025 Thermo Fisher Scientific Junior Innovators Challenge . Society for Science Sitting in his family’s living room in New York City, 14-year-old Miles Wu was astonished to find that a simple piece of paper, folded into a Miura-ori origami pattern, could hold 10,000 times its own weight. For a total of more than 250 hours, Wu had diligently designed, folded and tested copious variations of the technique—a series of tessellating parallelograms that can fold or unfold in one fell swoop—to find one that could be used to build deployable shelters for emergency situations like natural disasters. “I was really shocked by how much [weight] these simple pieces of paper could hold,” says Wu, who’s currently a ninth-grade student at Hunter College High School in New York City. Wu was especially intrigued by the Miura-ori fold, named after its inventor, the Japanese astrophysicist Koryo Miura . Miles Wu A series of tessellating parallelograms, the Miura-ori fold can be manipulated into copious variations. Miles Wu Wu had always been fascinated with the ancient Japanese art of origami, but he really began indulging in it as a hobby about six years ago. In 2024, he started exploring paper folding beyond its appeal as a creative pursuit. “I started reading about how different types of geometric origami were being studied and applied in STEM for their various physical properties,” he says. Although origami dates back centuries, the fields of engineering, medicine, mathematics and architecture didn’t develop a profound interest in it until the 1960s. Since then, origami has been used in the design of biomedical devices , such as stents and catheters, and self-assembling robots . Wu was especially intrigued by the Miura-ori fold, named after its inventor, the Japanese astrophysicist Koryo Miura . Famed for its use in aeronautical engineering, the fold has been leveraged to make solar panels for spacecraft and satellites. One of its earliest space applications was in Japan’s Space Flyer Unit , a satellite launched in 1995. Fun fact: Bloom patterns hold great promise for science and engineering A student at Brigham Young University recently discovered a new family of origami patterns that resemble flowers as they unfold. These so-called bloom patterns could be used to build telescopes and satellites. The pattern of creases and angles, which can be manipulated to create many variations, “folds this really large sheet

Source: Hacker News | Original Link

苹果官宣3月4日新品发布：iPhone 17e、廉价MacBook有望亮相_腾讯新闻 苹果官宣3月4日新品发布：iPhone 17e、廉价MacBook有望亮相 腾讯数码 2026-02-16 22:18 发布于 北京 腾讯数码官方账号 文｜吴彬 编辑｜徐青阳 北京时间2月16日，苹果公司对外官宣了春季新品发布的邀请函，预告3月4日晚10点，将在上海举行一场体验活动，这也意味着苹果一大波新品即将到来。 按照之前彭博社记者古尔曼的报料，在接下来的三月，我们可能会看到包括iPhone、iPad和Mac在内的全产品线更新。 其中最先被确认的，是iPhone17e。它将接替一年前的上市的iPhone 16e，核心变化包括：采用与iPhone 17同代的A19芯片，加入 MagSafe 磁吸充电，并转向苹果最新一代自研蜂窝通信与无线芯片。 但 iPhone 17e 整体的设计不会有太大的变化，依然采用刘海屏，定价依然维持599美元不变。 iPad产品线，苹果将更新入门款iPad与iPad Air，和iPhone 17e一样，新品的外观设计变化不大，主要升级集中在处理器上。 入门款iPad将升级到A18芯片，iPad Air将升级到M4芯片。由于A18的加入，入门款iPad将首次支持 Apple Intelligence ，这也会成为该产品的重要宣传点之一。 Mac产品线，常规的升级包括14英寸与16英寸 MacBook Pro迭代机型、搭载M5处理器的 MacBook Air，以及升级版Mac Studio等机型都在更新计划之中。另外，古尔曼也透露Studio Display也将迎来更新，但外观没有变化，主要是增加ProMotion高刷，补齐上代产品只有60Hz的遗憾。 最大的变数可能来自一款名为“MacBook”的机型，传闻这是苹果为入门级市场打造的一台全新Mac笔记本电脑。 从目前的报料信息来看，这台MacBook可能会是苹果唯一一台不搭载M系列芯片，而是采用上一代iPhone同款的A18 Pro芯片（苹果甚至不想给A19 Pro……）。苹果这样设计，意味着MacBook的定位将会是一台应对网页浏览、邮件处理及轻度图像编辑等日常轻办公需求的笔记本产品。 这款新MacBook的显示屏尺寸预计约为12.9英寸，将采用与MacBook Air相同的LCD显示技术，机身更小巧，并且可能没有刘海设计。相比在售的MacBook其他笔记本产品，这台新的MacBook有可能会提供蓝色、粉色、黄色和银色等大胆的配色方案。这意味着苹果可能会将iMac的多彩风格延伸到MacBook中。

Source: Tencent News | Original Link

印度举办AI“影响力”峰会_腾讯新闻 印度举办AI“影响力”峰会 新华社新闻 2026-02-16 21:02 发布于 广东 新华社新闻官方账号 【新华社微特稿】为期五天的印度人工智能影响力峰会16日在新德里揭幕。按多家外媒说法，尽管印度在人工智能发展方面雄心勃勃，但“还有很长的路要走”。 本届峰会以“人、进步、地球”为主题，由印度电子和信息技术部主办，号称“有史以来规模最大”。按印度方面的说法，除法国、巴西等国领导人将出席峰会，谷歌公司首席执行官孙达尔·皮柴、美国开放 人工智能 研究中心( OpenAI )首席执行官萨姆·奥尔特曼、美国微软公司总裁布拉德·史密斯等科技企业高管预计出席。印度总理莫迪定于19日在峰会上发表讲话。 谈及本次峰会，印度电子和信息技术部长阿什维尼·瓦伊什瑙说：“目标十分明确，人工智能应服务于人类发展、包容性增长与可持续未来。” 不过，法新社评价说，尽管印度规划了大规模数字基础设施建设计划，怀有“宏大的创新雄心”，但在人工智能发展方面，印度“仍有很长的路要走”。 路透社报道，峰会让新德里酒店价格飙升，在社交媒体上引发争议。泰姬陵酒店一间平日约2200美元一晚的套房，上周标价已超过3.3万美元。 印度最高法院14日发布通知说，考虑到峰会周边区域可能出现严重交通拥堵，律师本周可通过线上方式出庭。（完）（刘江）

Source: Tencent News | Original Link

谁在为上春晚豪掷千金？四家机器人厂商同台pk_腾讯新闻 谁在为上春晚豪掷千金？四家机器人厂商同台pk 21世纪经济报道 2026-02-16 20:04 发布于 广东 21世纪经济报道官方账号 21世纪经济报道记者 陈归辞 自去年蛇年央视春晚，宇树科技机器人凭借一把扭秧歌一炮走红后，今年春晚已然成为机器人厂商的营销秀场。 经过一年，行业正快速演进至量产与规模化商用的关键拐点，多家企业进入IPO冲刺阶段，卡位战气氛日益浓烈。 宇树科技、魔法原子、银河通用、松延动力四家机器人企业齐登马年央视春晚，其中半数来自长三角。有媒体消息称，为了拿下合作，各家不惜砸下重金。 作为参照，国内头部人形机器人企业年营收多在数亿元至十亿元级别。据咨询机构IDC统计，2025年全球人形机器人销售额约4.4亿美元（约30.4亿人民币）。对于这些企业而言，这笔营销投入绝非小数目。 四家机器人厂商分别是何方神圣？为何豪掷重金上春晚？ 哪四家机器人厂商登台？ 除了宇树科技已是三度登上春晚的“元老”外，魔法原子、银河通用、松延动力三家企业均为第一次登上春晚。 宇树科技成立于2016年8月，专注于消费级、行业级高性能通用足式/人形机器人及灵巧机械臂的自主研发、生产和销售，由王兴兴创立。2024年其营收已超10亿元，2025全年人形机器人出货量超5500台、本体量产下线超6500台。宇树科技已于2025年11月完成A股IPO辅导工作，有望成为A股“人形机器人第一股”。 魔法原子成立于2024年1月，专注于通用机器人和 具身智能 技术研发与落地应用，核心团队均来自具身智能行业及产业链顶尖企业，其中70%以上为研发人员。其产品涵盖通用人形机器人、仿生四足机器人等，覆盖工业、商业及家庭等多类场景，公司与追觅科技深度绑定。 银河通用成立于2023年5月，专注于通用机器人，已在商业、工业、医疗等场景中应用。2025年12月，银河通用完成新一轮3亿美元 融资 ，最新估值已达30亿美元。此前，银河通用完成 股改 ，被认为是在为赴港上市做准备，但公司方对此作出了否认。 松延动力成立于2023年9月，专注于人形机器人研发与制造，凭借“万元级”消费机器人走出一条差异化赛道，核心创始人员来自于清华大学与中科院等多所知名院校。今年2月，其完成股改。 一掷千金求曝光，行业多家企业推进IPO 机器人厂商不惜重金亮相春晚，为的是换取数亿级的曝光，以提升品牌能见度与价值，并为估值与融资带来激励。 资本市场方面，今年具身智能行业预计将迎来一波上市潮，春晚也因此被市场称为机器人企业的“ 超级路演舞台 ”。 宇树科技已完成上市辅导；云深处科技已启动上市辅导，将在4月-6月提交辅导验收申请；乐聚机器人已完成股改，预计于3月-6月完成上市辅导；银河通用、松延动力已完成股改。此外，星海图、众擎机器人和魔法原子均已释放出IPO信号。 产业端，当前国内行业出货量持续放量、大订单陆续落地。IDC数据显示，全年全球人形机器人出货量接近1.8万台，同比增长约508%，销售额约4.4亿美元；同期，累计销售订单量预计超过3.5万台，为后续交付和市场持续放量奠定基础。 宇树科技2025年人形机器人出货超5500台；智元出货超5100台；优必选2025年全年人形机器人订单接近14亿元，柳州机器人超级智慧工厂投产后2026年产能将达1万台，国内头部企业已进入 批量化生产 阶段。 不过，人形机器人的商业应用场景还较为局限。东吴证券研报指出，尽管人形机器人行业整体有较大规模订单落地，但下游场景主要来自于政府、数据采集和生活服务场景。 2026年，人形机器人产业将迎来商用化关键节点。瑞银集团最新报告指出，人形机器人正加速从概念验证迈向工业应用，2026年全球需求量预计将达到3万台。

Source: Tencent News | Original Link

“token anxiety”; or, a slot machine by any other name | jae kaplan “token anxiety”; or, a slot machine by any other name 16 Feb, 2026 I realize it’s gauche to blog about some shit you saw on bluesky but yesterday I saw a post that encapsulated so much of what has been bumming me out about the rise of coding agents over the last year. this dread had been slowly rising from seeing blogs about using claude code from your phone while getting ready for work, while commuting, while waiting to pick your kids up from school, but it’s come to a head. Token Anxiety i think i mostly echo this for myself. with so much that can be done, i often feel like i should be doing something, always [image or embed] — Tim Kellogg ( @timkellogg.me ) February 15, 2026 at 6:44 AM now obviously the opinions of founder-brained SF social bubble weirdos should be immediately discounted; they are the spiders georg of this industry. but at the same time they are playing into the dreams of management, the worker that never stops working, that’s always online, that’s infinitely Productive, always shipping, always wants to get back to work. I imagine this archetype exists in other industries but my experience is limited to tech so I will stick to that. my fear is that this will become the norm. anecdotal evidence 1 tells me that more and more companies are adopting AI for their engineers to use, encouraging (and in some cases requiring) its use in an effort to boost productivity, despite no actual evidence pointing to these improvements 2 and Anthropic-funded research indicating that AI usage reduces skill retention 3 . so where does this lead us? we know that some US tech companies are starting to embrace the “996” schedule popularized in China’s tech industry. enforced usage of coding agents makes that push even easier—is it really work if all you’re doing is telling the computer what to do and then reviewing it to make sure it didn’t do anything wrong and also babysitting it all hours of the day? 4 many have already observed that working with coding agents, which require constant attention and often generate low-quality code with (by design) random results, are a slot machine. they are loot boxes. they are gambling. you are constantly pulling the lever and hoping you get the SSR SaaS Passive Income product. you will not get this, but maybe you will. just one more prompt, one more pull, one more revision, one more go at being Absolutely Right 5 . if you suffer from token anxiety, you have a gambling addiction. I’m sorry that it’s not being formally treated as such, but you can take some solace in the fact that novel forms of gambling often take time to be recognized 6 . now we can put our thinking caps on and follow a pretty easy chain of events. coding agents can trigger our gambling instincts with slot machine-like behavior; tech companies are pushing engineers to work more and encouraging or enforcing the use of coding agents to get there; gambling is addictive; heavy users o

Source: Hacker News | Original Link

Wero – Digital payment wallet Skip to navigation Go to the content Move money in real time. For real. Send, receive money and pay direct from your bank account in 10 seconds or less, 24/7 with Wero – it’s simple, seamless and Made in Europe. Activate Wero now Pause ticker Scroll down to next section All you need is a phone number Fast and hassle-free Seamless payments in Belgium, France and Germany All you need is a phone number Fast and hassle-free Seamless payments in Belgium, France and Germany All you need is a phone number Fast and hassle-free Seamless payments in Belgium, France and Germany This is just the beginning. People have been using Wero to send and receive money between family and friends in Belgium, France and Germany for a while and now you can pay instantly online. Check if it’s available for you!  And this is just the beginning: soon we’ll add in-store payments to your Wero digital wallet, and even pay for your subscriptions. Stay tuned as we launch new services and expand to more European countries. Wero in your country It’s a new day Transcript This is a fast-paced, brightly colored animated and live-action promotional video for Wero, an instant money transfer service. The video uses large, bold text overlays, stylized graphics (like hands, a stopwatch, a lock, and a paper airplane), and short clips of diverse people using their phones to emphasize the product’s speed, simplicity, and security. Super fast payments. Send and receive money between bank accounts in less than 10 seconds. Made in Europe. Proudly European, Wero lets you pay and send money across borders online. Always on, no delays. Send and receive money 24/7 even at night and on weekends. Easy to access. Available to most bank customers in Belgium, France and Germany. Snap, Tap & Done. The easy way to settle up with family and friends or pay online direct from your bank account. No card, no IBAN, no hassle. Activate Wero now Buying popcorn for the gang before the movie? Get paid back with Wero. All you need is a phone number (and optionally, some quick-responding friends ;)) Your friend ordered sushi for everyone? Pay him back quickly with Wero. You’ll have settled your part even before the chopsticks touch the sushi. Want to snag those kicks before your best mate sees them? Just scan, tap – and they’re yours. Backed by your trusted banks. Do you see your bank? Good news: your bank has signed up to offer Wero, and if not now, you should be able to start using Wero soon. If your bank is not part of the Wero movement, we are very sorry too – don’t hesitate to ask them to hurry up – we can’t wait to welcome you on board too! Get started Your questions, answered. Go to full support How do I get started? If your bank offers Wero, it’s already available in your mobile banking app! You may need to update your banking app to the latest version. Just check your mobile banking app or sign-in to your online banking site and follow a few simple steps to create your Wero prof

Source: Hacker News | Original Link

Use Protocols, Not Services Use Protocols, Not Services Published on 15 February 2026 The Internet is almost anonymous and privacy-preserving by design. I mean, unless some administrator actively tries to track you, there is no built-in identity layer. What breaks both properties is the centralization of communication onto closed platforms, where identification becomes possible either by the hosting company itself, or by governments compelling them to cooperate. After recent events, it is time for us to start using protocols again instead of services. Services are easy targets A government that wants to identify users, censor content, or enforce compliance only needs to send one letter to one company. One subpoena, one court order, one regulatory demand: the service likely complies or faces fines, lawsuits, or bans. This is happening right now. Governments worldwide are passing laws that require platforms to verify the age of their users. Discord is voluntarily rolling out mandatory “teen-by-default” settings until proof of majority (by submitting a face scan or, God forbid, a government-issued ID), likely anticipating future regulatory obligations. None of this could happen with a protocol. You cannot require age verification on IRC, XMPP, ActivityPub , Nostr , or Matrix , because there is no single entity to compel. Each server operator makes their own decisions. A government would need to individually pressure thousands of independent operators across dozens of jurisdictions, which is a legislative and enforcement impossibility. And even if one server complied, users would simply move to another. Switching services solves nothing After Discord’s announcement, the instinct is to migrate to another service. This is pointless. The new service will either operate under the same jurisdiction and face the same rules, or it will be offshore and eventually blocked or pressured once it becomes large enough to matter. You are just moving from one regulable entity to another. The actual solution is to stop depending on a specific commercial service and start using a protocol. This is not a radical idea. We already do it with email. SMTP is a protocol. You can switch providers, self-host, or use any combination. Email may not seem to be the best example since it has become an oligopoly where Google, Microsoft, and maybe also Apple control the vast majority of the email infrastructure. But actually, this is a good example to show how protocols are resilient. Let’s say Google bans your account, then you can move to another provider and still reach every Gmail user. In a more extreme scenario, let’s even say Google and Microsoft discontinue their service (in your specific region, for example), even block any inbound message from you. Not ideal, but SMTP implementations still exist and they still work even in a very degraded mode. You’d need to migrate (as well as some of your connections), but there is absolutely no need to reimplement anything. That is the

Source: Hacker News | Original Link

Privilege is bad grammar | Tadaima. Privilege is bad grammar 13 Feb, 2026 When I got my first real job, I used to get so nervous about writing emails to my boss. I would run spellcheck, triple-check the grammar, read over it again and again to make sure my tone sounded professional and mature and not young and stupid. After painstakingly revising the email for 30 minutes, I would send it to my boss, who would respond right away with a message that looked like: K let circle back nxt week bout it . thnks Sent from my iPhone I had another job where my bosses were heavy emoji users. I would send them super professional emails, trying so hard to overcompensate for how young I was, and they would respond back with a single sentence punctuated with multiple cryface emojis (😂). To this day, I think of that emoji as “corporate” since professionals love to use it for whatever reason. I’m used to it now, but a decade ago I thought it was so odd. I thought we were supposed to be professionals? And professionals are supposed to write with good grammar, right? I’ve been thinking a lot about this ever since the latest Epstein document dump. 1 People have been uploading screenshots of emails between Epstein and Elon Musk, Bill Gates, and Richard Branson. And besides all the upsetting and salacious details everyone is discussing, the thing that also surprises me is how bad everyone’s grammar is. It reminded me so much of emails from bosses in my life: short, blunt (almost rude?), typos galore, weird formatting, bad grammar, “sent from iPhone”, etc. It’s almost as if, once you get to a certain level of power, you no longer need to try. Because the only reason people spend time crafting a well-written email is to look powerful, mature, professional. But if you’re already a powerful professional, I guess technically you don’t need to make an effort. And if there’s no other boss above you, you can do whatever you want. It reminds me of another email leak, the 2014 Sony Pictures hack . While everyone ooed and ahhed over a bunch of executives talking crap about celebrities, the main thing I remember from that whole scandal was how sloppy and unprofessional emails from executives looked like. I remember reading over those emails with almost a sense of jealousy. If I had sent out an email with even a quarter of the typos they had, I probably would’ve lost my job. I know words like “privilege” gets thrown around a lot, and I think we all understand monetary privileges and power privileges and race privileges, but grammar privilege? That’s certainly a first. I haven’t read through it myself, but I’ve heard more than enough through articles/social media. ↩ #news

Source: Hacker News | Original Link

I guess I kinda get why people hate AI I guess I kinda get why people hate AI Feb 14, 2026 I’m sitting on a lānai in a hotel in Waikiki beach, writing this article, and wondering if the job I am starting nine days from now will be my last. This is a unique situation for me in a few ways—I’ve never been to Hawaii before, I think the five minutes it’s taken me to come up with that opening sentence is the somehow the most time I’ve ever spent on a hotel balcony, and this is the first time I’ve actually followed through on the “I should delay my start date to take a vacation” idea I’ve had every time I’ve switched jobs. There’s one difference, however, that looms larger in my mind. It’s not the “wondering if the new job will be my last” thing. I’ve worked exclusively in startups, and while the primary reason I’ve done so has been because I enjoy the agency and impact you can have at early-stage companies, I’d be lying if the idea of cashing in cheap ISOs into early retirement wasn’t a factor in each job offer I accepted. The difference here is why I’m wondering that. Previously, it was wondering if I would need a job after this one. Now, it’s wondering if I’ll be able to acquire a job after this one, or if AI is going to completely take over my profession and ruin my career. Not the first time I’m not the first human to have anxiety about technological development. Change is scary, and technology changes a lot of stuff. In my opinion, these changes are mostly for the better—but that’s not an opinion everybody shares. The classical cultural example is the Luddites, a social movement that failed so utterly that its name because a common metaphor for stubborn morons who are terrified of technological innovation that helps everybody. Deservedly so, to be clear—while it’s true that textile experts did suffer from the advent of mechanical weaving, their loss was far outweighed by the gains the rest of the human race received from being able to afford more than two shirts over the average lifespan. The other example that comes to mind is the (possibly apocryphal) stories around the rollout of ATMs, where many supposedly predicted that the number of bankers in the US would collapse now that you could withdraw $20 in singles to leave tips without talking to a person. The exact opposite happened, of course. Being able to easily interact with banks, without waiting in a line that’s too long for the dum-dum you get at the end to be a real consolation, made people use banks more . And suddenly tellers became loan managers, and account advisors, and the machine that was supposed to destroy banking employment wound up supercharging it. I could go on, but somebody else already has , so there’s not much point in it. Technology changes things, and sometimes it hurts people in the short-term, but every invention from fire to mRNA vaccines has wound up generally increasing human welfare. I’ve long taken the view that this trend will continue. I remember arguing with peo

Source: Hacker News | Original Link

WebMCP WebMCP Draft Community Group Report , 12 February 2026 More details about this document This version: https://webmachinelearning.github.io/webmcp Issue Tracking: GitHub Editors: Brandon Walderman ( Microsoft ) Khushal Sagar ( Google ) Dominic Farolino ( Google ) Copyright © 2026 the Contributors to the WebMCP Specification, published by the Web Machine Learning Community Group under the W3C Community Contributor License Agreement (CLA) . A human-readable summary is available. Abstract The WebMCP API enables web applications to provide JavaScript-based tools to AI agents. Status of this document This specification was published by the Web Machine Learning Community Group . It is not a W3C Standard nor is it on the W3C Standards Track. Please note that under the W3C Community Contributor License Agreement (CLA) there is a limited opt-out and other conditions apply. Learn more about W3C Community and Business Groups . 1. Introduction WebMCP API is a new JavaScript interface that allows web developers to expose their web application functionality as “tools” – JavaScript functions with natural language descriptions and structured schemas that can be invoked by agents , browser’s agents , and assistive technologies . Web pages that use WebMCP can be thought of as Model Context Protocol [MCP] servers that implement tools in client-side script instead of on the backend. WebMCP enables collaborative workflows where users and agents work together within the same web interface, leveraging existing application logic while maintaining shared context and user control. 2. Terminology An agent is an autonomous assistant that can understand a user’s goals and take actions on the user’s behalf to achieve them. Today, these are typically implemented by large language model (LLM) based AI platforms , interacting with users via text-based chat interfaces. A browser’s agent is an agent provided by or through the browser that could be built directly into the browser or hosted by it, for example, via an extension or plug-in. An AI platform is a provider of agentic assistants such as OpenAI’s ChatGPT, Anthropic’s Claude, or Google’s Gemini. 3. Security and privacy considerations 4. Accessibility considerations 5. API 5.1. Extensions to the Navigator Interface The Navigator interface is extended to provide access to the ModelContext . partial interface Navigator { [ SecureContext , SameObject ] readonly attribute ModelContext modelContext ; }; 5.2. ModelContext Interface The ModelContext interface provides methods for web applications to register and manage tools that can be invoked by agents . [ Exposed = Window , SecureContext ] interface ModelContext { undefined provideContext ( optional ModelContextOptions options = {}); undefined clearContext (); undefined registerTool ( ModelContextTool tool ); undefined unregisterTool ( DOMString name ); }; navigator . modelContext . provideContext(options) Registers the provided context (tools) with the browser. This method

Source: Hacker News | Original Link

GitHub – SpaceTurth/Org-Web-Adapter: A simple view into your org roam notes with editing, backlinks, and rendered math. Skip to content You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session. You switched accounts on another tab or window. Reload to refresh your session. Dismiss alert SpaceTurth / Org-Web-Adapter Public Notifications You must be signed in to change notification settings Fork 1 Star 6 A simple view into your org roam notes with editing, backlinks, and rendered math. License AGPL-3.0 license 6 stars 1 fork Branches Tags Activity Star Notifications You must be signed in to change notification settings SpaceTurth/Org-Web-Adapter master Branches Tags Go to file Code Open more actions menu Folders and files Name Name Last commit message Last commit date Latest commit History 7 Commits 7 Commits media media notes notes static static templates templates .gitignore .gitignore LICENSE LICENSE README.md README.md config.yaml config.yaml main.py main.py makefile makefile View all files Repository files navigation Org Web Adapter A lightweight local web app for browsing and editing Org files. The app is implemented as a single Python server ( main.py ) plus one HTML template ( templates/index.html ) and one stylesheet ( static/style.css ). It scans a notes directory for .org files and renders a 3-pane UI. ⚠️ There is no authentication or encryption, only run this service on trusted networks. ⚠️ Screenshots Desktop Mobile Instructions Symlink your notes directory to notes . Edit the bind address and port in config.yaml if desired. python3 main.py . How’s it work? main.py starts an HTTP server. On each page request ( GET / ), it rescans the notes directory for .org files. It resolves links/backlinks and builds HTML fragments. It injects those fragments into templates/index.html placeholders: {{NAV_ITEMS}} {{MAIN_CONTENT}} {{BACKLINKS}} Browser JS in templates/index.html handles client-side interactions (search, shuffle, sorting, jump-to-current, theme toggle). Server-side components ( main.py ) File discovery and parsing: scan_org_files(…) recursively finds .org files. Extracts title from #+TITLE: and ID from :ID: . Org link/backlink handling: resolve_link_target(…) supports file:… and id:… links. find_backlinks(…) computes notes linking to the selected note. build_backlink_counts(…) computes backlink totals for sorting. Rendering: render_org_to_html(…) converts headings ( * , ** , …) and paragraphs to simple HTML. render_line_with_links(…) converts org links in text to clickable app links where resolvable. truncate_label(…) caps sidebar labels to 32 chars with … . Editing: POST /edit updates a selected .org file and redirects back with status flags. Static files: serve_static(…) serves files under static/ with path traversal protection. Frontend components templates/index.html : Base layout markup. Sidebar controls. Small JS contr

Source: Hacker News | Original Link

What Your Bluetooth Devices Reveal About You » Danny Building Bluehood, a Bluetooth scanner that reveals what information we leak just by having Bluetooth enabled on our devices. If you’ve read much of this blog, you’ll know I have a thing for privacy . Whether it’s running my blog over Tor , blocking ads network-wide with AdGuard , or keeping secrets out of my dotfiles with Proton Pass , I tend to think carefully about what data I’m exposing and to whom. Last weekend I built Bluehood , a Bluetooth scanner that tracks nearby devices and analyses their presence patterns. The project was heavily assisted by AI, but the motivation was entirely human: I wanted to understand what information I was leaking just by having Bluetooth enabled. The timing felt right. A few days ago, researchers at KU Leuven disclosed WhisperPair (CVE-2025-36911), a critical vulnerability affecting hundreds of millions of Bluetooth audio devices. The flaw allows attackers to hijack headphones and earbuds remotely, eavesdrop on conversations, and track locations through Google’s Find Hub network. It’s a stark reminder that Bluetooth isn’t the invisible, harmless signal we treat it as. The Problem Nobody Talks About We’ve normalised the idea that Bluetooth is always on. Phones, laptops, smartwatches, headphones, cars, and even medical devices constantly broadcast their presence. The standard response to privacy concerns is usually “nothing to hide, nothing to fear.” But here’s the thing: even if you have nothing to hide, you’re still giving away information you probably don’t intend to. From my home office, running Bluehood in passive mode (just listening, never connecting), I could detect: When delivery vehicles arrived, and whether it was the same driver each time The daily patterns of my neighbours based on their phones and wearables Which devices consistently appeared together (someone’s phone and smartwatch, for instance) The exact times certain people were home, at work, or elsewhere None of this required any special equipment. A Raspberry Pi with a Bluetooth adapter would do the job. So would most laptops. Devices You Can’t Control What concerns me most isn’t that people choose to have Bluetooth enabled. It’s that many devices don’t give users the option to disable it. Hearing aids are a good example. Modern hearing aids often use Bluetooth Low Energy so audiologists can connect and adjust settings or run diagnostics. Pacemakers and other implanted medical devices sometimes broadcast BLE signals for the same reason. The user can’t simply turn this off. Then there are vehicles. Delivery vans, police cars, ambulances, logistics fleets, and trains often have Bluetooth-enabled systems for fleet management, diagnostics, or driver assistance. These broadcast continuously, and the drivers have no control over it. Even consumer devices aren’t always straightforward. Many smartwatches need Bluetooth to function at all. GPS collars for pets require it to communicate with the owne

Source: Hacker News | Original Link

The Sideprocalypse 2026-02-16 The Sideprocalypse You can’t open a feed today without having AI boosters fling word salad like “agentic engineering” or “openclaw” into your beautiful but disapproving face. I’m terrible at predicting the future—you should ask me about selling NVIDIA stock in early 2022 some time—but one thing seems abundantly clear at this point. There’s a wonderful Swedish proverb called “elda för kråkorna” (building a fire for the crows) that evokes the futility of someone lighting a nice warm fire indoors and then throwing the doors wide open, inviting the snow and sleet. Are you one of the thousands of developers with dreams of building a little SaaS on the side? Something you’ve been thinking about for a while, hacking away at on evenings and weekends, dreaming of the day you can have a couple of hundred paying customers giving you $19.99 a month? Sorry to be the bringer of bad news, but that dream’s dead. Doornail. Dodo. Parrot . Every minute you put into that thing, be it your own time or your LLM agent’s, is a minute for the crows. Imagine that beautiful but fragile little idea of yours curled up in a freshly-dug pit in your backyard, Claude and Gemini standing over it, chuckling and high-fiving each other. Listen: every idea you’ve ever had, every single one, some cocaine-addled sales critter has had too. And they’re better than you at SEO . What’s that you’re saying? Yours actually works and is higher quality, because you know about things like TTFP and INP and “not putting your Supabase god-token in the client”? Oh, you sweet summer child: I take no pleasure in this but I need to tell you that these things don’t matter anymore. Quality is not a metric anyone cares about in 2026. We’ve all been conditioned to accept tombstone spinners on first load or purchase flows that straight-up don’t work , and there are walled gardens and tollbooths everywhere you look. React ate the web, Safari kneecapped it, Google stopped linking to it, and none of the Zaibatsu US corporations who hold the cards want you to succeed. The future, if there is a future in software, lies in high-touch enterprise sales. There are a select few companies allowed to make money on the Internet today, and if you have any sense of self-preservation you need to glom onto one of those and hold it like the Dickens. Give up those childish dreams of independence. If you’re a hopeful SaaS builder you may be first in line, but you can at least take grim satisfaction in the fact that the sloptimists, the hype-men, the breathless agents-are-working-while-I-sleep people, they’re equally fucked when this all comes home to roost. The marginal value of code today is—well, possibly not zero, since the people selling spades for this frantic gold rush are doing okay for now—but it’s dropping like a lead balloon. Josh Collinsworth has it right , AI boosterism is a class privilege. But rest assured that this revolution, too, will end up eating its own. They fancy themselves t

Source: Hacker News | Original Link

性价比机场选购指南 – V2EX 首页 注册 登录 V2EX = way to explore V2EX 是一个关于分享和探索的地方 现在注册 已注册用户请 登录 0.29 V2EX › 宽带症候群 性价比机场选购指南 ZeawinL · 12 小时 44 分钟前 · 2798 次点击 协议 优先选择新协议, 通常: Vless > AnyTLS > ShadowTLS > Hysteria2 > Tuic > Trojan > Vmess > SS/SSR 线路类型 简单理解 稳定性 2026 合理价位 备注 直连 你的设备 → 直接连境外服务器(便宜但不稳) ★☆☆☆ (易被封) 5-10 元/月 仅做备用 中转 (BGP) 你的设备 → 国内中转优化 → 境外服务器(性价比高) ★★★☆ (有优化) 15-30 元/月 建议 公网中转 专线 (IPLC) 你的设备 → 专用物理线路 → 境外服务器(贵但稳) ★★★★ (不过墙) 40 元+/月 办公/不差钱 延迟 TCP ping 和 Google ping 越低越好, 可以网上搜相关的测速情况. 最好有试用自己测试, 看晚高峰(20:30 – 22:30)的表现. 自己简单测试可以开 Youtube 4K 视频拖拖进度条. 费用 推荐一主一备, 备用机场推荐按量付费. 月付优先. 开业时间不到 1 年的不要年付. 稳定性 通常在官网公告, 或者 tg 通告频道会有历史维护公告. 特殊时期挂了多久能修好. 流量倍率 如果有 0.1x 0.3x 等低倍率节点就更好. 如果只有高倍率节点好用可以放弃. 隐私安全 机场主理论上能看到你访问的域名. No Logs 不记录日志最好. 支付方式: USDT (加密货币) > 支付宝/微信. 特殊需求 流媒体解锁能力? 政治敏感内容封锁? 客服渠道 Telegram 群 / 网页工单 / Email. 只留 QQ 群/微信(会被喝茶，高风险). 客户端 定制客户端使用起来方便, 但开源客户端会更安全. 运营与口碑 真正的好机场，通常都很低调，官网可能都要翻墙才能搜到，靠的是老用户的口口相传。相关论坛的骂声或者夸赞，才是最真实的买家秀. 源: github.com/thesomeexp/Proxy-service-recommendations 26 条回复 • 2026-02-16 19:45:41 +08:00 1 twoz 12 小时 40 分钟前 via Android ipv6 直连稳定便宜落地解锁 ipv4 即可 2 MYDB 12 小时 12 分钟前 via iPhone 14 说点大家不知道的，想引流小白去 star ，然后挂推广？比如 github 上常见的垃圾机场狗狗加速？ 3 ZeawinL OP 12 小时 8 分钟前 via iPhone @ MYDB 你说得对😁 4 liuzimin 11 小时 56 分钟前 via Android 我买了个 IPLC 机场感觉和中转一样挂啊。前几天中转线路挂了，IPLC 也同步挂了。难道说是假 IPLC ？ 5 mangmaimu 11 小时 43 分钟前 via iPhone 除了直连，其他的协议肯定首选 ss ，UDP 协议最不好用，国内有 qos 6 MacsedProtoss 11 小时 39 分钟前 via iPhone 1 呵呵 我自建直连几年了 7x24h 365D 从来没被封 一个月跑几个 T 2.5Gbps 都没有任何关系 价格还贼划算 除了臭打游戏的要延迟非常敏感的比较难搞得加钱 其他的谁用机场谁 sb 7 MacsedProtoss 11 小时 37 分钟前 via iPhone 还搁这优先 h2 但凡用 h2 的都是垃圾到极点 谁家好线路用 h2 一点技术原理都不懂搁这骗小白呢 8 janeyee110 11 小时 31 分钟前 人家这是暗戳戳地做推广的。不骗小白骗谁呢。 @ MacsedProtoss 9 bugtik 11 小时 25 分钟前 你这个是只能发 VPN 区的 10 pplcc 9 小时 22 分钟前 via iPhone @ MacsedProtoss 哥们用那个协议的，怎么我一个 ip 半月就没了呢 11 love2328 8 小时 59 分钟前 挂? 没挂过 就是有时抽风,影响也不大 12 wangwaner 8 小时 52 分钟前 via iPhone 便宜的机器 h2 还可以啊…俺用了好几年也没啥问题…还得看人数，最好加一个敲门机制。 13 slowman 8 小时 51 分钟前 绕了半天还是 AI slop 14 wangwaner 8 小时 51 分钟前 via iPhone 最近严打对等带宽的机器（出站流量和入站流量特征大致相等的）就是在打击中转机器啊 15 sir283 8 小时 45 分钟前 我一直都是用的这个，下片都够用了。 [img] [/img] 16 Cyanver 7 小时 49 分钟前 via Android @ sir283 给个链接 17 Shadowxxx 7 小时 43 分钟前 via Android @ sir283 给个链接 18 Donahue 7 小时 36 分钟前 @ MacsedProtoss #6 哪里买的 vps 19 patrickyoung 6 小时 36 分钟前 /go/promotions 20 HFX3389 6 小时 23 分钟前 还玩中转和专线？不是上下性一致的都给清退了？ 21 Thymolblue 6 小时 11 分钟前 @ MacsedProtoss 自建有 US 的机器推荐吗？ 22 leena 5 小时 57 分钟前 via iPhone @ sir283 这是哪家的，方便发链接么 23 silencefly 5 小时 52 分钟前 via iPhone @ Thymolblue #21 主流且稳定的 BWG DMIT 24 MacsedProtoss 5 小时 7 分钟前 via iPhone @ Thymolblue @ silencefly 没错 就是 bwg 和 dmit 25 SmithJohn 5 小时 4 分钟前 @ MacsedProtoss 便宜的机场一个月能在十块以内.换到年付就是 10-20 刀,这个价格的机器,使用体验不会很好吧? 26 iApp 3 小时 52 分钟前 @ sir283 给个链接 关于 · 帮助文档 · 自助推广系统 · 博客 · API · FAQ · Solana · 1900 人在线 最高记录 6679 · Select Language 创意工作者们的社区 World is powered by solitude VERSION: 3.9.8.5 · 108ms · UTC 15:38 · PVG 23:38 · LAX 07:38 · JFK 10:38 ♥ Do have faith in what you’re doing. ❯

Source: V2EX | Original Link