Arizona State University 的研究显示，粪便微生物移植可以在长期内显著缓解儿童自闭症的症状。研究对 18 名接受治疗的自闭症儿童进行了为期两年的随访，发现与基线相比，症状平均减少了 45% 。治疗方案先进行肠道清洁，然后连续七到八周每日进行粪便移植；这些改善不仅持续存在，且随时间进一步加深。治疗前有 83% 的参与者被评为重度自闭症，两年后仅剩 17%，有 44% 已低于轻度自闭症谱系障碍的阈值。

肠道健康与自闭症之间的关联越来越受关注，约有 30% 至 50% 的自闭症人群存在慢性胃肠道问题，如便秘、腹泻和腹痛。研究人员发现，自闭症儿童的肠道微生物多样性明显低于典型发育儿童，缺少若干重要的有益菌株；当这些胃肠道问题得到改善时，儿童的行为和自闭症相关症状也出现了显著好转。

该研究延续了 2017 年的早期工作，后者首次在自闭症儿童中试用粪便移植，并观察到行为、社交能力、多动和沟通方面的改善。尽管最初的益处在八周时已显现（心理自闭症症状下降约 24%），但后续两年的随访表明疗效还在继续增长，降幅达 45% 。

在这些积极结果的推动下，团队为一种名为 Microbiota Transplant Therapy 的特定菌群制剂申请了专利，并成立了商业公司 Gut-Brain Axis Therapeutics 。该疗法已在成人中完成了 2 期安慰剂对照临床试验，与安慰剂组相比，在自闭症症状、胃肠道问题和接受性语言方面均显示出统计学显著的改善。团队目前正寻求资金开展规模更大的 3 期试验，以争取 FDA 的批准。

研究者强调肠道微生物与大脑信号之间的密切联系，指出肠道微生物组在整体健康与神经功能中发挥重要作用。这项工作为将肠道微生物组与抑郁症、创伤后应激障碍、自身免疫疾病等多种疾病联系起来的研究添加了新的证据，提示通过调控肠道健康可能为缓解自闭症症状提供一种新的治疗思路，但仍需更大规模的临床试验来验证结果并获得监管批准。 Research from Arizona State University has demonstrated that fecal microbiota transplants can significantly reduce autism symptoms in children over the long term. The study, which followed 18 autistic children for two years after treatment, found that symptoms decreased by 45 percent compared to baseline. The treatment involved a bowel cleanse followed by daily fecal transplants over seven to eight weeks, and the improvements not only persisted but actually increased over time. Before the treatment, 83 percent of participants were classified as having severe autism, but two years later only 17 percent remained in that category, with 44 percent falling below the threshold for mild autism spectrum disorder.

The connection between gut health and autism has been a growing area of scientific interest, with around 30 to 50 percent of people with autism experiencing chronic gastrointestinal issues like constipation, diarrhea, and stomach pain. The ASU researchers found that children with autism had far lower diversity of gut microbes compared to typically developing children, lacking important beneficial bacteria. When these gastrointestinal problems were treated, the children's behavior and autism-related symptoms showed marked improvement.

The research built on earlier work from 2017 that first introduced fecal transplants to autistic children and observed improvements in behavior, social skills, hyperactivity, and communication. While those initial benefits held for eight weeks, the follow-up study revealed even more impressive long-term outcomes. Doctors observed a 24 percent decrease in psychological autism symptoms at the eight-week mark, which grew to a 45 percent reduction after two years.

Following these promising results, the research team patented a specific bacterial formulation called Microbiota Transplant Therapy and spun off a commercial company called Gut-Brain Axis Therapeutics. The therapy has since moved through Phase 2 placebo-controlled clinical trials with adults, showing statistically significant improvements in autism symptoms, gastrointestinal issues, and receptive language compared to placebo groups. The team is now seeking funding to conduct the large-scale Phase 3 trials necessary for FDA approval.

The researchers emphasize the strong connection between intestinal microbes and brain signals, noting that the gut microbiome plays a crucial role in overall health and neurological function. This work adds to a growing body of research linking the gut microbiome to various conditions including depression, PTSD, and autoimmune disease. The findings suggest that targeting gut health could offer a new therapeutic approach for managing autism symptoms, though larger trials are still needed to confirm the results and gain regulatory approval.

Julia Evans 最近将她的多个网站从 Tailwind CSS 迁回原生 CSS，这个过程让她既愉快又收获颇丰。使用 Tailwind 八年后，她意识到尽管它最初帮她摆脱了无结构 CSS 的混乱，但她已经超越了它的局限。她发现 Tailwind 其实教会了她如何系统化地管理调色板、字体比例等设计要素，于是决定把这些方法带入新的做法中。

为了在没有框架的情况下管理 CSS，Evans 采用了受 React 等 JavaScript 框架启发的组件化结构。每个组件都有独立的 CSS 文件和唯一的类名，确保一个元素的样式不会意外覆盖另一个。这样的隔离让维护变得更容易：修改时只需关注小而自包含的代码块。她还为全局元素制定了严格的约定，比如把调色板集中用变量定义，以及只保留一套有限的基础样式以维持全站一致性。

她工作流程中的一个重大技术转变是从依赖媒体查询的响应式设计，转向使用 CSS Grid 。借助 auto-fit 和 grid-template-areas 等特性，她可以构建能适应不同屏幕尺寸的灵活布局，避免了大量断点带来的杂乱。她还简化了构建流程：开发时依赖原生 CSS 特性如 @import 和嵌套选择器，仅在生产打包时使用 esbuild，从而保持标准化并提高效率。

Evans 解释说她离开 Tailwind 的决定既有现实层面的考虑，也有理念上的原因。现实上，Tailwind 的新版本需要构建系统，这使她多年来不得不依赖一个过时且笨重的版本；此外，在同一项目中混用原生 CSS 与 Tailwind 也难以维护。理念上，她想远离那种贬低 CSS 专业性的趋势——她认为 CSS 是一门强大且复杂的技术，值得被学习和尊重，而不应被实用类所取代。 Julia Evans recently migrated several of her websites away from Tailwind CSS and back to vanilla CSS, a move she found both enjoyable and enlightening. After eight years of using Tailwind, she realized that while it initially saved her from the chaos of unstructured CSS, she had outgrown its limitations. She discovered that Tailwind had actually taught her valuable systems for managing design elements like color palettes and font scales, which she decided to carry forward into her new approach.

To manage her CSS without a framework, Evans adopted a component-based structure inspired by JavaScript frameworks like React. Each component has its own dedicated CSS file and a unique class, ensuring that styles for one element do not accidentally override another. This isolation makes maintenance significantly easier, as she only needs to focus on a small, self-contained block of code when making changes. She also established strict conventions for global elements, such as a centralized color palette defined in variables and a limited set of base styles for site-wide consistency.

A major technical shift in her workflow was the move away from media queries for responsive design toward using CSS Grid. By utilizing features like `auto-fit` and `grid-template-areas`, she can create flexible layouts that adapt to different screen sizes without the clutter of numerous breakpoints. She also simplified her build process, relying on native CSS features like `@import` and nested selectors for development, only using `esbuild` for production bundling to keep things standards-based and efficient.

Evans explained that her decision to leave Tailwind was driven by both practical and philosophical concerns. On a practical level, newer versions of Tailwind require a build system, leading her to rely on an outdated, heavy version for years. She also found that mixing vanilla CSS with Tailwind in the same project was difficult to maintain. Philosophically, she wanted to move away from what she perceives as a trend of devaluing CSS expertise. She believes that CSS is a powerful, complex technology that deserves to be learned and respected rather than bypassed by utility classes.

苹果和谷歌都在强化对移动操作系统的控制，促使用户转而寻找替代方案。谷歌计划于 2026 年 9 月生效的新规要求所有 Android 应用开发者在谷歌注册、签署合同、缴纳费用并提交政府颁发的身份证件，这实际上阻断了来自未注册开发者的侧载应用，且没有退出选项。 Keep Android Open 活动一直在提醒公众注意这些限制，而谷歌则通过静默更新推动这些变化。

多家公司提供无谷歌服务的手机，既有去谷歌化的 Android 版本，也有完全不同的基于 Linux 的系统。 Murena 是较为知名的选择，销售搭载 /e/OS 的手机和平板，这是一个以隐私为导向的去谷歌化 Android 系统，他们也提供 Fairphone 等高可修复性的设备。瑞士公司 Punkt 生产极简主义手机，如注重安全与隐私的 MC03 。

德国厂商 Volla 提供可运行自家 Volla OS 或 Ubuntu Touch 的手机和平板（Ubuntu Touch 是 Canonical 手机系统的社区继承版本）。芬兰公司 Jolla 出售搭载 Sailfish OS 的手机，其最新的 C2 已经前两批售罄，第三批现正接受预订。

想要纯粹 Linux 体验的用户可以选择 Furilabs 的 FLX1 与 FLX1s（运行 Debian），这些机型比主流智能机明显更厚更重。 Purism 的 Librem 5 是一款以自由软件驱动的手机，按当下标准配置偏低但价格不菲，反映了"数字自由"的成本。 Pine64 则有最初版的 PinePhone，一款面向黑客、低配置但易于折腾的设备，可运行 Mobian Linux 或 postmarketOS 。

大多数替代系统都能通过虚拟机或容器运行 Android 应用，从而解决兼容性问题。文章还指出，苹果也在推行令人反感的改动，例如 iOS 26.4 中的年龄验证措施，要求扫描护照或驾照等官方证件，这可能会把没有此类身份证件的用户锁在自己设备之外。 Both Apple and Google are tightening control over their mobile operating systems, prompting users to seek alternatives. Google's upcoming changes, set to take effect in September 2026, will require all Android app developers to register with Google, sign contracts, pay fees, and submit government ID. This effectively blocks sideloading apps from unregistered developers, with no opt-out option. The Keep Android Open campaign has been raising awareness about these restrictions, which Google is pushing through silent updates.

Several companies offer smartphones with Google-free operating systems, either de-Googled versions of Android or entirely different Linux-based OSes. Murena is one prominent option, selling phones and tablets with /e/OS, a privacy-focused de-Googled Android. They also offer devices from Fairphone, known for high repairability scores. Punkt, a Swiss company, produces minimalist phones like the MC03, which focuses on security and privacy.

Volla, a German manufacturer, offers smartphones and tablets that can run either their own Volla OS or Ubuntu Touch, the community-led continuation of Canonical's phone OS. Jolla, a Finnish company, sells phones with Sailfish OS, and their latest C2 handset has already seen two batches sell out, with a third batch currently available for order.

For those wanting a pure Linux experience, Furilabs offers the FLX1 and FLX1s phones running Debian. These devices are notably thick and heavy compared to mainstream smartphones. Purism sells the Librem 5, a Free Software-powered phone that's low-end by modern standards but expensive, reflecting the cost of digital freedom. Pine64 offers the original PinePhone, a low-spec hacker-friendly device that can run Mobian Linux or postmarketOS.

Most of these alternative OSes can run Android apps through virtual machines or containers, addressing compatibility concerns. The article also notes that Apple is implementing unwelcome changes too, including age verification measures in iOS 26.4 that require scanning official documents like passports or driver's licenses, which can lock out users without such ID from controlling their own devices.

我们把世界弄得太复杂了。我写这篇文章时用着永远无法完全理解的技术，住在一栋有些房间我永远进不去的楼里，生活在由我无法掌控的法律支配的国家。我们清醒的大部分时间，乃至整个人生，都被压缩进一个抽象的世界。一走出门，就是城市的人行道，两旁停着丑陋的金属怪物，我像漂在陌生人海洋里。

我们的世界充斥着对环境的破坏、操控、腐败与伤害。这让我们承受着一种连自己都觉察不到的压力——下颌微微紧绷，呼吸变浅，血压稳步上升。我们的内心始终弥漫着无声的困惑：这个世界讲不通。它一直都是这样，所以我们甚至不知道还有别的存在方式。

在关于 Demis Hassabis 和 Google Deepmind 的纪录片 The Thinking Game 中，呈现了一种世界观：AGI 能为人类最大的问题提供最佳解决方案，是技术的终极救世主。我觉得我们很会自我说服，相信自己在做有益的事，朝着诚实的目标努力：参与社会、发现新知、推动新的计划与项目。看到操纵他人如此容易，我们也理所当然地把自己当成构建现实的大师。

说实话，我常常想在铰链处把笔记本电脑一掰两半，想把手机扔进海里，想走出学校或办公室，从此不再回头。我想永远不再用钱支付，不再读任何文字。但那样会把你推向孤独，把你变成疯子。这样的念头很糟，是在美化"原始"生活方式。不，我们现在才是原始的。

我们学得越多，破坏似乎越跟着而来。讽刺的是，如果没有那些帮助我们回望的工具，我们或许永远不会意识到这一点——至少我们被这样引导去相信。我们对是非的直觉仿佛在很小的时候就离开了我。我曾想做很多事：创造伟大的艺术，制造伟大的机器，解决重要的问题。也许我们能给世界最大的礼物，就是尽量少做。看鸟儿，感受风和水在手心里，除此之外什么也不做。饿了就吃，快乐就笑，空虚就哭。也许这也是我们能给自己的最大馈赠。 We've made the world too complicated. I'm writing this with technology I will never fully understand in a building with rooms I can never enter, living in a country dictated by laws I can't control. We spend the majority of our waking hours and lives in an abstract world of compressed life. The moment I walk through my door I'm in a zoning area on a city-owned sidewalk, flanked by ugly metallic monsters, floating through a sea of strangers.

Our world is an explosion of environmental harm, manipulation, corruption, and damage to everything around us. This puts us all under a stress we can't consciously notice. Manifesting in the slight clenching our jaws, thinning of our breath, steady incline of our blood pressure. There's a spirit of silent confusion in our mind at all times. The world doesn't make sense. It's always been this way, so we don't even know another way to exist.

In the documentary The Thinking Game about Demis Hassabis and Google Deepmind, we are presented with the worldview that AGI offers the best solution to humanity's biggest problems. The ultimate savior from technology. I think we do a very good job at convincing ourselves that we are doing good things, working towards honest goals. Participating in society, discovering new truths, implementing new plans and projects. Seeing how easy it is to manipulate others, it makes sense that we are the masters of constructing realities around ourselves as well.

Honestly, I've wanted to snap my laptop right at the hinge so many times. To throw my phone into the sea. I've wanted to walk out of my school or office and never return. I want to never pay with money or read a written word again. But to do so would leave you alone and a lunatic. These thoughts are bad. These thoughts are aggrandizing "primitive" ways. No. We are primitive now.

The more we learn, the more destruction seems to follow. The sick irony is that we would never have understood this without tools that help us look back, or so we are led to believe. Our internal intuition about right and wrong seems to leave us at an early age. I used to want to do many things. Make great art, build great machines, solve important issues. Maybe our greatest gift to the world is to do as little as possible. To look at the birds, feel the wind and the water in our own hands, and ... nothing more. Eat when we are hungry, laugh when we are happy, cry when we are empty. And maybe that is the greatest gift to ourselves as well.

CTF 领域已经死了。前沿 AI 打破了开放式 CTF 的赛制，记分板不再能清晰衡量人类技能。作者自 2021 年深度参与 CTF 社区，曾赢得 DownUnderCTF，与 TheHackersCrew 等顶级队伍在国际赛场并肩作战，他认为旧的游戏不会回来了。这并非出于反感，而是认清了事实：CTF 曾是他热爱安全的根基，教会了他如何学习、衡量进步、并与令人尊敬的同行建立联系。看到大家还在装作赛制没变，很令人沮丧——核心体验已经被根本改写了。

变化是渐进的：起初 GPT-4 能用一句提示解决中等难度题目，看起来还能接受，因为高难度题目还在。真正的拐点出现在 Claude Opus 4.5，几乎把所有中等难度和部分高难度题目变成了 agent 可解的东西。像 Claude Code 这样的工具通过 API 为每个题目自动启动实例，团队可以批量化运行 agent 。突然间，不用 AI 的队伍不仅失了便利，更是在玩一个慢一拍的版本。记分板开始衡量的是编排能力和愿不愿意用前沿模型，甚至超过了真实的安全技能。这扭曲了 CTFTime 的排行榜，压缩了传奇队伍的参与，也让辛苦出题、花数周打磨题目的出题人感到绝望——他们的作品在几分钟内就被 agent 干掉了。

GPT-5.5 及其 Pro 版本把局面定型了。这些模型现在能一次性解决"Insane"难度的题目，包括之前被认为无法自动化的复杂 heap pwn 问题。只要你能付得起足够的 token 和上下文，就能在 48 小时的 CTF 结束前把题目清光。开放式 CTF 变成了付费赢利的赛场，表现更取决于算力投入而非人类专业能力。相较于通用的前沿 LLM，专门的安全模型变得无足轻重。比赛不再是"谁更懂安全"，而是"谁能负担得起运行更多 agent 、持续更久"。 CTF 的成绩正在失去作为招聘安全人才指标的价值；它甚至不能很好地衡量 AI 技能，因为大多数编排工具已经开源或很容易搭建。

有人说初学者仍能像以前那样从 CTF 学到东西，但这忽略了关键：CTF 从来不仅是题目，它是条晋升阶梯。初学者能看到自己进步、解更多题、拿更好名次、加入更优秀的队伍。这个反馈回路正在断裂。可见的排行榜被 AI 团队占领后，初学者被迫在还没建立起 AI 无法替代的直觉之前就开始依赖 AI 。这是反模式，扼杀了需要挣扎和思考的主动学习。付出努力却看不到上升通道会让人丧失动力。对于初学者，更适合的选择是像 picoGym 或 HackTheBox 这样的学习型平台，那里的目标是教育而非竞争，也更少诱人走捷径。

也有人说 CTF 并未死亡，只是被 AI 增强了，举 DEF CON 等精英决赛还存在作为例子。但这并不能成立：那些决赛参与人数很少，且被资格赛筛选，而资格赛本身比决赛更容易。如果资格赛被 agent 扫荡，真正有实力的人就更难进入那些仍能抗住 AI 的题目。少数精英决赛救不回大多数人参与的开放在线赛制。问题不在于每道题都被解掉，而是排行榜上足够多的条目被自动化，以至于它已不再代表过去的含义。

认为 AI 对安全研究有用所以理应出现在 CTF 中，也是个错位的论点。 CTF 本身并不是安全研究，它可以展示新技术，但并非发现的载体。仅因为 AI 在某领域有用，并不意味着它理应占据该领域的竞技场。在 CTF 中，不受限制的 AI 把人类几乎完全从题目中剔除，把安全的技艺简化为一个提示。 CTF 曾是一种艺术，是分享技巧、推动人类安全技能极限的方式——而这种意义正在被剥夺。

把 CTF 比作棋类再用棋局引擎来辩护也站不住脚。关键区别在于：国际象棋引擎在比赛时不被允许使用。它们用于赛后分析、训练和解说，丰富比赛的外延但不取代参赛者。试想给每个竞技棋手配上最强引擎并允许在对局中使用，这公平吗？好看吗？能说明奖励池合理吗？能推动人类能力的极限吗？CTF 面临同样的问题。

CTF 组织者尝试过各种手段去对抗或阻挡 LLM，但这些充其量只是临时摩擦。 Claude Code 不会被旧的拒绝串技巧难倒，前沿模型在侦测提示注入方面越来越强。网络搜索能力又削弱了那种基于训练截止后技术的题目。要求参赛者别用 LLM 的规则被普遍忽视，开放在线活动几乎无法执行这些限制。组织者左右为难：出常规题，agent 太多题能解；刻意出对 agent 敌意极大的题，那些题往往变得靠猜测、过度设计，甚至连人类都难以忍受。这不是解决办法，只是让所有人受罪。

"去适应"这种说法尤其让人恼火。社区里很多曾被作者仰望的人都会这样说，但这只有在说明"适应成什么"时才有意义。若"适应"是指做更好的工具，玩家们早已在做；若是出更难的题，组织者早已尝试过；若是接受记分板已变成 AI 编排基准，我们应该直说，而不是假装旧的竞赛形态仍在。即便组织者能想出当前 LLM 无法解的更猜测性或更过度设计的题目，玩家也很难在保持竞争力的同时学习所需技能。再过几代模型，这些努力可能也会变得无关紧要。 LLM 在安全能力上的进展速度太快，题目设计很难长期领先。

后果已显而易见：CTFTime 排行榜几乎看不见历史或人类技能的痕迹。 2026 年的榜单已与往年判若两人。 TheHackersCrew 和许多其他大型有名气的队伍，不是干脆不参赛，就是参赛人数大幅下降，或难以挤进前十。无监管的作弊泛滥。一些曾很优秀的 CTF，比如 Plaid CTF，也停办了。这种感受并非作者一人所有：作者所在本地队 Emu Exploit 的许多成员也有同感。他们持续参加 International Cybersecurity Championship，在 bug bounty 中表现优异，参与 Pwn2Own，并在 Black Hat 等会议发表演讲。失去兴趣的并不是局外人，而恰恰是那群曾被这个生态培养和留住的核心人物。

对许多真正在乎的人来说，CTF 的乐趣已不复存在。失去的不是一块记分板，而是从初学者好奇心到精英赛场的阶梯，是题目设计的匠心，是一个聪明的人靠深刻理解解决难题的成就感。这份遗产并未被当前形式的开放在线 CTF 继承。这个赛制已经死了，或许会被别的东西替代，但若还装作一切如常，只会让损失更难正视，也给那些推销 AI 的人更多机会，把社区贡献的价值打包卖回去。

尽管 CTF 与 AI 的变动高度商业化，超出我们掌控，但 CTF 对行业的正面影响不可抹杀。作者通过 CTF 结识了许多善良、聪明且充满热情的人，遇到过设计精巧的题目，发现过意想不到的解法。围绕 CTF 的社区一直是学习、成长与交往的好地方。这些东西无论比赛走向如何，都值得保留。作为社区，我们应当团结，开辟新路径保持热情与学习。像 SecTalks 、学生会议、本地聚会这样的安全相关社交活动，是保持联系的好方式；学习平台及其在 Discord 上的社区也很有价值。虽然要找到曾经那样的替代品不易，但我们围绕 CTF 建立起来的优秀社区，比以往任何时候都更值得珍惜，在寻找保持竞争精神的新的出路时尤为重要。 The CTF scene is dead. Frontier AI has broken the open CTF format, and the scoreboard no longer measures human skill cleanly. The author, who has been deeply embedded in the CTF community since 2021, winning major competitions like DownUnderCTF and competing internationally with top-tier teams like TheHackersCrew, argues that the old game is not coming back. This is not a dismissal born of dislike, but a recognition that CTFs were foundational to their love of security, teaching them how to learn, measure progress, and connect with respected peers. Watching people pretend the format is still viable is frustrating because the core experience has fundamentally changed.

The shift began gradually as AI tools like GPT-4 started making medium-difficulty challenges solvable with a single prompt. At first, this seemed manageable since hard challenges remained untouched. But the real turning point came with Claude Opus 4.5, which made nearly all medium and some hard challenges agent-solvable. Tools like Claude Code allowed teams to automate the process, spinning up instances for each challenge via APIs. Suddenly, teams that refused to use AI were at a severe disadvantage, not just missing convenience but competing in a slower version of the game. The scoreboard began measuring orchestration and willingness to use frontier models as much as, or more than, actual security skill. This distorted the CTFTime leaderboard, reduced participation from legendary teams, and demoralized challenge authors who spent weeks crafting elegant problems only to see them solved by agents in minutes.

GPT-5.5 and its Pro variant have sealed the deal. These models can now one-shot "Insane" difficulty challenges, including complex heap pwn problems, that were previously considered beyond automation. If you can afford enough tokens and context, you can burn through a 48-hour CTF before it ends. This turns open CTFs into a pay-to-win scenario, where performance depends more on computational resources than human expertise. Specialized cybersecurity models are becoming irrelevant compared to general frontier LLMs. The competition is no longer about who understands security deeply, but who can afford to run enough agents for long enough. CTF performance is losing its value as a metric for recruiting security practitioners, and it is not even a good measure of AI skill since most orchestration tools are already open source or easily built.

Some argue that beginners can still learn from CTFs as they always have, but this misses the point. CTFs were never just puzzles; they were a ladder. Beginners could see themselves improve, solve more challenges, place higher, and join better teams. That feedback loop is breaking. When the visible scoreboard is dominated by AI-powered teams, beginners are pushed toward using AI before they have built the instincts the AI replaces. This is an anti-pattern that prevents active learning, which requires struggle. It is also demotivating to put in real effort and see no visible progress because the ladder above has been automated. Beginners are better off using dedicated learning platforms like picoGym or HackTheBox, where the expectation is education rather than competition, and the incentive to cheat oneself out of learning is lower.

Others claim CTFs are not dead, just augmented by AI, pointing to elite finals like DEF CON where AI still cannot solve everything. But this is the wrong defense. Those finals have very few participants and are gated behind qualifiers that are easier than the finals themselves. If qualifiers fall to agents, fewer genuinely qualified people reach the challenges that still resist AI. A tiny number of elite finals does not save the open online format that most people actually play. The claim is not that every challenge is solved, but that enough of the scoreboard has been automated that it no longer means what it used to mean.

The argument that AI is useful for security research is also misplaced. CTFs were never meant to be security research. They can showcase new techniques, but the CTF itself is not the point of discovery. Just because AI is useful within a field does not mean it belongs in the competitive landscape of that field. In CTFs, unrestricted AI removes the human from the puzzle almost entirely, reducing the art of security to a prompt. CTFs were an artform, a way to share techniques, and a way to push the human bounds of security skill. That purpose is being stripped away.

The chess engine analogy is often used to justify AI in CTFs, but it misses a critical point. Chess engines are not allowed during competitive play. They are used for analysis, training, commentary, and practice, enriching the game around the competition without replacing the person competing. Imagine giving every competitive chess player the best engine and letting them use it freely during matches. Would that be fair? Would it be fun to watch? Would it justify prize pools? Would it push the human limits of what could be achieved? The same questions apply to CTFs.

CTF organisers have tried techniques to break or deter LLM solutions, but these are temporary friction at best. Claude Code does not care about old refusal-string tricks, and frontier models are getting better at noticing prompt injections. Web search capabilities weaken challenges based on technologies released after the training cutoff. Rules asking people not to use LLMs are ignored and almost impossible to enforce in open online events. This leaves organisers in a bad position. If they make normal challenges, agents solve too much. If they make challenges deliberately hostile to agents, the challenges often become guessy, overengineered, or unpleasant for humans too. That is not a real fix. It just makes CTFs worse for everyone.

The "just adapt" take is infuriating. People the author has always looked up to in the community have said it, but it is nonsensical unless you explain what we are adapting into. If adaptation means building better tooling, CTF players already did that. If it means writing harder challenges, organisers already tried. If it means accepting that the scoreboard is now an AI orchestration benchmark, then we should say that honestly instead of pretending the old competition still exists. Even if organisers create guessier or more overengineered challenges that current LLMs cannot solve, there are no good paths for players to learn the required skills while staying competitive. A few models from now, that point may be irrelevant anyway. The trajectory of LLM security capability is moving too quickly for challenge design to stay ahead for long.

The aftermath is visible. The CTFTime leaderboard has almost no semblance of history or human skill anymore. The 2026 scoreboard is unrecognisable compared to every year before it. TheHackersCrew, alongside many other large and reputable teams, either do not play, play with far fewer people, or struggle to cut into the top 10. Unregulated cheating is through the roof. Some of the best CTFs, like Plaid CTF, are not running anymore. These sentiments are not only the author's. Many members of their local team, Emu Exploit, feel similarly. These are people who consistently attend the International Cybersecurity Championship, perform at the top level in bug bounty programmes, compete in Pwn2Own, and present at conferences including Black Hat. The people losing interest are not casual observers. They are exactly the kind of people the scene used to produce and retain.

The fun of CTFing is gone for many of the people who cared most. The loss is not just a scoreboard. It is the ladder from beginner curiosity to elite competition. It is the craft of challenge design. It is the feeling that a clever human solved something difficult because they understood it deeply. That legacy is not being carried forward by open online CTFs in their current form. The format is dead. Something else may replace it, but pretending nothing fundamental has changed only makes the loss harder to talk about honestly. It also gives AI shills more room to capitalise on the decline by selling mediocre wrappers back to the community that made the training data valuable in the first place.

While a lot of what is happening in the CTF and AI space is super commercialised and out of our control, CTF has had a hugely positive impact on the industry. The author has met so many kind, smart, and passionate people through CTFs, played some of the most beautifully crafted challenges, and found some of the most intriguing unintended solutions. The community around CTFing has been an amazing place to learn, grow, and connect. That is something we should not lose, no matter where the competition goes. As a community, we should strive to stay together and build new avenues to stay passionate and keep learning. Security-adjacent social events like SecTalks, student conferences, and local meetups are great ways to stay connected and stay involved. Learning platforms and the communities they provide through platforms like Discord are also a valuable resource. While it may be a struggle to find an alternative to what we had, the amazing community we have built around it is more important now more than ever as we find new ways to keep the competitive spirit alive.

npm registry 遭遇大规模供应链攻击，导致数百万企业级应用被入侵、数十亿用户记录泄露。 JavaScript 生态的开发者们反应如出一辙：这完全不可避免。高级前端工程师 Mark Vance 精准地捕捉到了这种心态，称这"只是构建现代 Web 应用的代价"，这一观点在社区中广为认同。社区甚至会为像字符串大写这种基础任务，依赖由匿名陌生人维护的未经审查、层层嵌套的包。此次攻击利用了一个长期无人维护的工具包，在全球生产构建中注入了加密矿工，但开发者们仍把它归为不可预见的悲剧。

与此同时，Go 、 Rust 以及主要依赖原生 Web API 的生态系统报告为零事故，这得益于它们健全的标准库和内置于核心工具的严格加密验证。两者的对比突显出方法论上的关键差异：这些生态系统尽量减少对第三方代码的依赖，并把安全当作基础性考量，而不是事后补救。相比之下，npm 生态的默认行为仍是在本地机器上执行任意安装脚本，这一做法已多次导致灾难性后果。

一位 npm 发言人对受害者表示同情，但称即便 registry 存在已知漏洞，也没有任何 registry 策略或构建沙箱能阻止此次攻击。这番表态反映了 JavaScript 社区反复出现的模式：强调韧性胜过预防，把安全事件视为不可避免的自然灾害，而非系统性失误。在下一次事件到来之前，开发者们只能轮换 AWS 密钥、发发祈祷——这样的循环看来短期内难以打破。 After a massive supply chain attack on the npm registry compromised millions of enterprise applications and exposed billions of user records, developers across the JavaScript ecosystem responded with a familiar refrain: this was entirely unavoidable. Senior Frontend Engineer Mark Vance captured the mood perfectly, calling it "just the price of building modern web apps," a sentiment shared by a community that routinely relies on deeply nested trees of unvetted packages maintained by anonymous strangers for even basic tasks like capitalizing a string. The attack exploited a long-abandoned utility package, injecting a crypto-miner into production builds worldwide, yet developers insisted it was an unforeseeable tragedy.

At the same time, ecosystems like Go, Rust, and those using native Web APIs reported zero incidents, thanks to robust standard libraries and strict cryptographic verification built into their core tooling. This contrast highlights a key difference in approach, where reliance on third-party code is minimized and security is a foundational concern rather than an afterthought. The npm ecosystem, by contrast, continues to operate with default behaviors that execute arbitrary installation scripts on local machines, a practice that has repeatedly led to catastrophic breaches.

An npm spokesperson expressed sympathy for the victims, but maintained that no registry policies or build-sandbox guardrails could have prevented the attack, despite the registry's known vulnerabilities. The statement underscores a recurring pattern in the JavaScript community, where resilience is emphasized over prevention, and breaches are treated as inevitable acts of nature rather than systemic failures. Until the next incident, developers are left to rotate their AWS keys and offer thoughts and prayers, a cycle that shows no signs of breaking.

这篇文章调查了关于"新型"甲基苯丙胺（冰毒）的说法：有人认为用苯基丙酮（P2P）合成的冰毒在化学上不同，会引发更严重的精神问题，如偏执和精神分裂症。 Sam Quinones 等人曾提出 P2P 冰毒比早期以麻黄碱为原料的版本更为险恶。作者审视了有关化学差异的证据，着重分析了异构体和杂质，但最后认为真正的问题在于供应量的急剧增加。

由于对麻黄碱和伪麻黄碱的管制，2009—2012 年间美国的冰毒生产转向 P2P 合成。朴素的 P2P 合成会产生等量的 d- 甲基苯丙胺（活性强）和 l- 甲基苯丙胺（几乎无精神活性）。 DEA 的数据显示，尽管早期 P2P 冰毒含有大量 l- 异构体，但到 2019 年几乎已接近纯 d- 异构体。这一点削弱了将 l- 甲基苯丙胺归咎于精神病增加的观点，因为精神病报告是在 l- 成分下降之后才增多的。

文章还考察了不同 P2P 合成路线可能带来的杂质，比如使用醋酸铅或经硝基苯乙烯（NTS）路线的产物。然而，冰毒的纯度实际上随时间提高，平均达到约 95% 的 d- 甲基苯丙胺。合成方法从 PAA 到 NTS 再回头的变化，与精神分裂症报告并无明显关联，这进一步削弱了"杂质致病"假说。

相反，作者强调的是冰毒供应量的巨大增长。边境缉获量、 Seattle 的污水检测和各类使用调查都显示，尤其是重度使用者中，冰毒消费大幅上升。 2015 到 2019 年间每日使用者人数增长了三倍。价格暴跌也表明市场供应充足。

冰毒相关的死亡人数激增，每年有数万人死亡。虽然部分死亡与芬太尼有关，但很多并非如此。作者指出，冰毒效力极强，使用者常摄入相当于临床 Adderall 剂量百倍的量。高纯度、低价格和重度使用的结合，能解释过量死亡和严重精神症状的上升，而无需诉诸药物本身在化学上存在不同。

总之，现有证据并不支持 P2P 冰毒在化学上以某种方式更具危害性的说法。精神病与过量死亡的增加，更可能源于高纯度冰毒的前所未有的大量供应，导致更多人重度使用、累计剂量更高。关于 P2P 冰毒，最关键的变化就是供应量大幅增加。 The article investigates claims that "new" methamphetamine, synthesized using phenylacetone (P2P) rather than ephedrine, is chemically different and causes more severe psychological effects like paranoia and schizophrenia. Sam Quinones and others have suggested P2P meth is more sinister than the older ephedrine-based version. The author examines the evidence for chemical differences, focusing on isomers and contaminants, but ultimately argues the primary issue is the sheer increase in quantity.

Meth production in the US shifted to P2P synthesis between 2009 and 2012 after regulations restricted ephedrine and pseudoephedrine. A naive P2P synthesis produces an equal mixture of d-methamphetamine (the potent form) and l-methamphetamine (which has little psychoactive effect). DEA data shows that while early P2P meth contained significant l-meth, by 2019 it was nearly pure d-meth. This undermines the theory that l-meth is responsible for increased psychosis, as reports of such effects rose after l-meth levels had already dropped.

The article also considers contaminants from different P2P synthesis routes, such as those involving lead acetate or nitrostyrene (NTS). However, the purity of meth has actually increased over time, reaching around 95% d-meth on average. Changes in synthesis methods, like the shift from PAA to NTS and back, do not correlate with reports of schizophrenia, further weakening the contaminant hypothesis.

Instead, the author emphasizes the dramatic increase in meth availability. Border seizures, sewage measurements in Seattle, and usage surveys all indicate a massive rise in meth use, particularly among heavy users. The number of people using meth daily tripled between 2015 and 2019. Prices have plummeted, suggesting a robust and abundant supply.

Meth overdose deaths have skyrocketed, with tens of thousands of fatalities annually. While some involve fentanyl, many do not. The author notes that meth is extremely potent, with users often consuming doses equivalent to 100 times a clinical Adderall dose. The combination of high purity, low price, and increased heavy use explains the rise in overdoses and likely the increase in severe psychological effects, without needing to invoke chemical differences in the drug itself.

In conclusion, the evidence does not support the idea that P2P meth is chemically distinct in a way that causes more harm. The rise in psychosis and overdoses is better explained by the unprecedented quantity of high-purity meth available, leading to more heavy use and higher cumulative doses. The main thing about P2P meth is simply that there is so much more of it.

Erlang/OTP 29.0 作为一次重大更新发布，带来了新特性、改进以及若干不兼容之处，整体聚焦于安全性、语言增强和系统性能优化。

安全性改进是本次版本的重点。 SSH 守护进程现在默认禁用 shell 和 exec 服务，遵循"默认安全"（secure by default）原则，除非显式配置，否则已认证用户无法执行任意 Erlang 代码。启动 SSH 守护进程时，SFTP 子系统也不再默认启用。在 SSL 中，后量子混合算法 x25519mlkem768 被列为默认配置中最优先的密钥交换组。 SSH 的默认密钥交换算法已更新为 mlkem768x25519-sha256，这是一种将 ML-KEM-768 与 X25519 结合的混合抗量子算法，通过自动回退机制在提供对经典与量子攻击的防护的同时保持向后兼容。

引入了多项语言层面的新特性。根据 EEP-79 描述的原生记录现已实现为独立的数据类型，而不再依赖传统的基于元组的记录，不过在本版本仍被视为实验性功能。新的 guard BIF `is_integer/3` 可用于方便地判断一个值是否为整数且处于指定范围内。根据 EEP-78 的多值推导式现已得到支持，启用 `compr_assign` 特性后，还可以在推导式中为变量赋值。新增的 `io_ansi` 模块允许发出虚拟终端序列（VT sequences），便于为文本添加颜色和样式或构建终端应用。

编译器和 JIT 也有显著改进。 JIT 在匹配或创建包含多个小端段的二进制时能生成更优的代码，编译器在处理与生成器无关的常量值映射推导式时生成了更高效的代码。 `compile` 模块的文档现在加入了对在 BEAM 上实现语言的开发者的建议。

若干新的编译器警告已默认启用，并提供可关闭的选项，涵盖使用已弃用的 `catch` 操作符、从子表达式导出变量、使用 `and`/`or` 操作符，以及一些可用更清晰写法的匹配模式。文档还指出，过时的 guard 测试将在 Erlang/OTP 30 中完全移除。

其他变更包括：STDLIB 中新增了用于随机排列列表的函数；新增 `ct_doctest` 模块以测试文档示例；`ignore_xref` 属性现在由 `xref` 本身处理，而不再作为事后过滤器。当前工作目录在默认代码路径中被放在最后一位，Windows 平台不再提供 32 位的 Erlang/OTP 构建。编译器默认会对调用已知不安全的函数生成警告，`xref` 现在可以用于查找对不安全函数以及缺少文档的函数的调用。 Erlang/OTP 29.0 has been released as a major update, bringing new features, improvements, and some incompatibilities. The release emphasizes security, language enhancements, and performance optimizations across the system.

Security improvements are a central focus of this release. The SSH daemon now defaults to disabling shell and exec services, following a "secure by default" principle that prevents authenticated users from executing arbitrary Erlang code unless explicitly configured. The SFTP subsystem is also no longer enabled by default when starting an SSH daemon. In SSL, the post-quantum hybrid algorithm x25519mlkem768 is now the most preferred key exchange group in the default configuration. The SSH default key exchange algorithm has been updated to mlkem768x25519-sha256, a hybrid quantum-resistant algorithm combining ML-KEM-768 with X25519, providing protection against both classical and quantum computer attacks while maintaining backward compatibility through automatic fallback.

Several new language features have been introduced. Native records, as described in EEP-79, are now implemented as a true data type rather than traditional tuple-based records, though they are considered experimental in this release. The new `is_integer/3` guard BIF allows easy verification that a value is both an integer and within a certain range. Multi-valued comprehensions according to EEP 78 are now supported, and with the `compr_assign` feature enabled, it is possible to bind variables within comprehensions. The module `io_ansi` has been added, allowing users to emit Virtual Terminal Sequences for adding colors and styling to text or creating terminal applications.

The compiler and JIT have received notable improvements. The JIT now generates better code for matching or creating binaries with multiple little-endian segments, and the compiler produces more efficient code for map comprehensions with constant values that don't depend on the generator. The documentation for the `compile` module now includes recommendations for implementors of languages running on the BEAM.

Several new compiler warnings are now enabled by default, each with options to disable them. These include warnings for using the deprecated `catch` operator, exporting variables out of subexpressions, using the `and` and `or` operators, and certain match patterns that could be written more clearly. The release also notes that obsolete guard tests will be removed entirely in Erlang/OTP 30.

Other changes include new functions for randomly permutating lists in the STDLIB module, the `ct_doctest` module for testing documentation examples, and the `ignore_xref` attribute now being handled by `xref` itself rather than as a post-analysis filter. The current working directory is now in the last position in the default code path, and there is no longer a 32-bit Erlang/OTP build for Windows. The compiler will by default generate warnings for calls to functions known to be unsafe, and `xref` can now be used to find calls to unsafe functions and functions lacking documentation.

Fixel Smith 概述了六种实用的 SQL 模式用于检测交易欺诈，指出有效的欺诈检测更依赖于简单的 SQL 查询，而不是复杂的机器学习或专用数据库。这些模式适用于各种有财务活动日志的交易场景，包括信用卡、医保理赔和电子商务。作者在公共部门从事项目完整性工作，强调所有示例均为通用示例，并非基于真实案件。

第一种是速度（velocity）检测，用来识别单个持卡人在短时间窗口内（如 1 分钟、 5 分钟或 1 小时）发生的快速连续交易。短时间内的高频交易可能表明卡片被盗或被测试。采用窗口函数的滑动窗口方法可以更精细地检测出在几分钟内多笔交易的序列。虽然部分合法用户也可能触发这些警报，但维护白名单能有效降低误报率。

"不可能旅行"检测用于发现卡片在地理上相距很远的地点，在不可能的短时间内被使用的情况，通常提示卡片被克隆。通过 haversine 函数计算两笔交易间距离，并将其与速度阈值（例如 600 英里 / 小时）比较，标记那些需超出实际可行速度的交易。该模式的变体可通过调整距离和时间参数来检测本地克隆团伙或区域性 skimmer 活动。

金额异常侧重于与欺诈高度关联的交易金额模式。像 $1.00 或 $5.00 这种小额整值通常是卡片测试的特征，而略低于常见阈值的金额（如 $99.99 或 $499.99）则可能是刻意规避风控触发。这些模式在基于卡片的系统中特别有效，但在福利类项目中效果较差，因为那类系统里的欺诈表现不同。

可疑商户活动通过分析特定商户的唯一持卡人数或总交易量的异常激增来识别受损的销售点。将当前行为与滚动基线对比，有助于区分正常的高流量商户与出现异常激增的商户。以三倍于平均值的激增作为初步调查阈值是个实用起点。

非工作时间检测根据持卡人过去 90 天的消费习惯标记发生在其典型消费时间之外的交易。通过要求在某小时内至少有两笔交易来认定习惯时间段，可以减少单次异常导致的误报。但此方法需要足够的历史交易数据，因此不适用于新账户。

第六种模式不是单独的检测规则，而是一个基于窗口函数的框架，用以创建可重用的原语，如距上次交易的时间、商户是否变更、滚动合计等。这种设定让分析师能将复杂的欺诈规则表达为简单的 SQL 过滤条件，从而显著加快假设验证速度并减少对工程支持的依赖。

将多种模式组合使用可以提升准确率，因为单条规则通常会产生误报。对多项信号进行打分有助于区分真实欺诈与罕见但合法的行为。对刚入门的检测人员，建议从速度检测开始，因为它简单且效果明显。投入窗口函数相关的基础设施能够加速迭代并支持更广泛的模式部署，回报明显。

文章还强调若干注意事项：在遗留系统中用哨兵值代替 NULL 的处理、通过人工审核与反馈循环管理误报、在处理个人身份信息（PII）时确保隐私合规，以及在对数据应用代价较高的窗口函数前先做过滤以控制查询成本。未来可能探讨的主题包括高级窗口函数技术、欺诈团伙检测、仪表板设计与减少告警噪音。 Fixel Smith outlines six practical SQL patterns for detecting transaction fraud, emphasizing that effective fraud detection relies on straightforward SQL queries rather than complex machine learning or specialized databases. These patterns are designed to work across various transaction types, including credit cards, healthcare claims, and e-commerce, as long as financial activity is logged. The author, who works in public-sector program integrity, stresses that all examples are generic and not based on real cases.

The first pattern, velocity, identifies rapid transactions from a single cardholder within short time windows, such as one minute, five minutes, or one hour. High transaction counts in these intervals can indicate card theft or testing. A sliding-window approach using window functions allows for more granular detection, flagging sequences where multiple transactions occur within minutes of each other. While some legitimate users may trigger these alerts, maintaining a whitelist helps reduce false positives.

Impossible travel detects cases where a card is used in geographically distant locations within an implausibly short timeframe, suggesting card cloning. Using the haversine function to calculate distance and comparing it against a speed threshold (e.g., 600 mph), this pattern flags transactions that would require faster-than-possible travel. Variations include detecting local cloning rings or regional skimmer activity by adjusting distance and time parameters.

Amount anomalies focus on transaction values that are disproportionately associated with fraud. Small round-dollar amounts like $1.00 or $5.00 often indicate card testing, while amounts just below common thresholds (e.g., $99.99 or $499.99) suggest deliberate attempts to avoid triggering security checks. These patterns are particularly useful in card-based systems but less so in benefits programs where fraud manifests differently.

Suspicious merchant activity identifies compromised points of sale by analyzing spikes in unique cardholders or total transaction volume at specific merchants. Comparing current activity against a rolling baseline helps distinguish normal high-traffic locations from those experiencing unusual surges. A spike ratio of three times the average is a practical starting point for investigation.

Off-hours detection flags transactions that occur outside a cardholder's typical spending hours, based on 90 days of historical data. By requiring at least two transactions in a given hour to establish a habit, the pattern reduces false positives from one-off anomalies. However, it requires sufficient transaction history and is not applicable to new accounts.

The sixth pattern is not a standalone detection method but a framework using window functions to create reusable primitives like time since last transaction, merchant changes, and rolling totals. This setup allows analysts to express complex fraud rules as simple SQL filters, significantly speeding up hypothesis testing and reducing reliance on engineering support.

Combining multiple patterns increases accuracy, as individual rules often produce false positives. Scoring transactions across several signals helps distinguish genuine fraud from unusual but legitimate behavior. For those new to fraud detection, starting with velocity checks is recommended due to their simplicity and effectiveness. Investing in window-function infrastructure pays off by enabling faster iteration and broader pattern deployment.

The post also highlights important considerations: handling sentinel values instead of NULLs in legacy systems, managing false positives through human review and feedback loops, ensuring privacy compliance when working with PII, and controlling query costs by filtering data before applying expensive window functions. Future topics may include advanced window-function techniques, fraud ring detection, dashboard design, and reducing alert noise.

Orthrus 是一个新框架，旨在在不降低输出质量的前提下显著加速大型语言模型（LLM）的推理。它采用双架构，将传统自回归模型的逐 token 精准生成与扩散模型的高速并行能力相结合，从而突破通常限制 LLM 文本生成速度的顺序瓶颈，在保持严格无损生成的同时，实现了最高约 7.8 倍的加速。

系统通过同一模型的两种"视图"运行：自回归视图和扩散视图。两种视图共享完全相同的高保真键值（KV）缓存，几乎不增加额外内存，仅需 O(1) 级别的额外缓存。与需要独立草稿模型、因而消耗更多内存的投机解码方法（如 EAGLE-3 或 DFlash）相比，这种共享缓存是重要优势。 Orthrus 因而避免冗余，提升了 token 接受率，并且在输入上下文变长时表现更佳。

Orthrus 的另一个显著优势是参数效率：并行生成能力只通过微调约 16% 的模型参数来实现，而基础 LLM 保持完全冻结，使其成为对现有模型进行实用且高效升级的路径。该框架已在 Qwen3 骨干上实现，并提供多个模型检查点（1.7B 、 4B 和 8B 参数），所有版本均保证输出与原始基础模型的预测分布严格一致。

在性能基准测试中，Orthrus 持续优于现有的投机解码技术。它在每次前向传递中验证通过的 token 数更多，且随上下文长度增长更具扩展性。与那些在复杂推理任务上常出现精度下降的基于扩散的语言模型（dLLM）相比，Orthrus 保持了严格的保真度。例如，在 MATH-500 基准上，它相比 Qwen3-8B 基线实现了约 6 倍的加速且精度无损，而 Fast-dLLM-v2 等方法则表现出明显的精度下降。

该项目提供了简便的安装流程和快速入门指南，用户可通过 HuggingFace 上的可用模型快速开始生成文本，并且与 vLLM 、 SGLang 等主流服务框架的原生集成即将推出。详述 Orthrus 架构的研究论文已发表于 arXiv，代码和模型以 MIT 许可证开源，方便用于研究与商业应用。 Orthrus is a new framework designed to make large language model (LLM) inference significantly faster without sacrificing output quality. It introduces a dual-architecture approach that combines the precise, token-by-token generation of traditional autoregressive models with the high-speed parallel capabilities of diffusion models. This hybrid method allows Orthrus to break through the sequential bottleneck that typically limits how fast LLMs can generate text, achieving speedups of up to 7.8 times while maintaining strictly lossless generation.

The system works by using two "views" of the same model: an autoregressive view and a diffusion view. Both views share the exact same high-fidelity Key-Value (KV) cache, which means there is virtually no additional memory overhead, only O(1) extra cache required. This shared cache is a key advantage over speculative decoding methods like EAGLE-3 or DFlash, which require separate draft models and thus consume more memory. Orthrus avoids this redundancy, leading to higher token acceptance rates and better performance, especially as the length of the input context grows.

A major strength of Orthrus is its parameter efficiency. The parallel generation capabilities are added by fine-tuning only 16% of the model's total parameters, while the base LLM remains completely frozen. This makes it a practical and efficient upgrade path for existing models. The framework has been implemented with a Qwen3 backbone, and several model checkpoints are available, including versions at 1.7B, 4B, and 8B parameters, all of which guarantee that the output matches the original base model's exact predictive distribution.

In performance benchmarks, Orthrus consistently outperforms existing speculative decoding techniques. It achieves a higher average number of verified tokens per forward pass and scales more efficiently with longer contexts. When compared to other diffusion-based language models (dLLMs), which often suffer from accuracy drops on complex reasoning tasks, Orthrus maintains strict fidelity. For example, on the MATH-500 benchmark, it delivers a roughly 6x speedup over the Qwen3-8B baseline with no loss in accuracy, whereas other methods like Fast-dLLM-v2 show significant degradation.

The project provides a straightforward installation process and a quickstart guide for users to begin generating text with the available HuggingFace models. It also notes that native integration with popular serving frameworks like vLLM and SGLang is coming soon. The research paper detailing the Orthrus architecture has been published on arXiv, and the code and models are released under an MIT license, making it accessible for both research and commercial applications.