文章指出，威权政权维持统治的方式，不仅靠精英的忠诚或意识形态的热情，也通过利用普通、常常平庸员工的职业野心来实现。基于对 Argentina's Dirty War（1970s–1980s）最新研究的材料，研究发现低层和中层官员——如军官、秘密警察和官僚——的动机并非主要来自极端主义或恐惧，而更多是出于想推动停滞的职业、争取哪怕微小的晋升。

这些所谓处于"职业压力"的人，在像秘密警察这样的镇压机构里找到了绕过传统晋升通道的机会，从而获得了平时难以企及的职业成功。研究挑战了长期以来关于为何有人参与威权体制的既有假设：人们往往把焦点放在精英的战略动机上，但普通基层的动因一直被忽视。

研究表明，巩固威权并不一定需要狂热分子或大规模恐怖，政权同样可以通过瞄准那些受挫、绩效低下的员工来有效招募。这些人把为政权服务视为个人上升的途径，把日常的职场不满转化为政治控制的工具。

在 Adam Scharpf 和 Christian Glassel 所著的 Making a Career in Dictatorship 中，这一现象被描绘为 Hannah Arendt 所说"恶的平庸性"与企业管理员工、处置低绩效者策略的一种黑暗融合。在 Argentina 的军队里，表现不佳者在秘密警察中的比例过高，他们可以通过"绕道"获得晋升与权力。他们参与侵犯人权的动机，除了意识形态或胁迫外，更多是典型官僚体制中的职业主义驱动。

这一认识对理解民主如何被侵蚀具有重要意义：民主规范的瓦解往往不是由戏剧性的反抗造成，而是由追求职业利益的普通人做出的默默妥协所推动。威权主义的滋长，很多时候不是来自非凡的邪恶，而是源于平凡个体的日常野心。 The article explores how authoritarian regimes maintain power not just through elite loyalty or ideological fervor, but by exploiting the career ambitions of ordinary, often mediocre employees. Drawing on new research from Argentina's Dirty War in the 1970s and 1980s, it reveals that lower- and midlevel officials—such as military officers, secret police, and bureaucrats—were motivated less by extremism or fear and more by the desire to advance stalled careers or secure minor promotions. These individuals, referred to as "career-pressured," found opportunities within repressive systems like the secret police to bypass traditional hierarchies and achieve success they couldn't attain otherwise.

The findings challenge long-held assumptions about why people participate in authoritarian regimes. While elites are often studied for their strategic incentives, the rank and file have remained poorly understood. The research suggests that would-be authoritarians don't need fanatics or widespread terror to consolidate power. Instead, they can recruit effectively by targeting frustrated, underperforming workers who see regime service as a path to personal advancement. This dynamic turns mundane workplace dissatisfaction into a tool for political control.

The book "Making a Career in Dictatorship" by political scientists Adam Scharpf and Christian Glassel frames this phenomenon as a dark fusion of Hannah Arendt's concept of the "banality of evil" with corporate strategies for managing low performers. In Argentina's military, poor performers were disproportionately represented in the secret police, where they could "detour" around standard promotion channels. Their participation in human rights abuses wasn't driven by ideology or coercion alone, but by the same careerist motivations found in any bureaucracy.

This insight has broader implications for understanding how democracies erode. It suggests that the collapse of democratic norms can be facilitated not by dramatic acts of defiance, but by quiet compromises made by ordinary people seeking professional gain. The article underscores that authoritarianism often thrives not on exceptional evil, but on the everyday ambitions of unremarkable individuals.

Cloudflare 近期参与了 Project Glasswing——与 Anthropic 合作测试其名为 Mythos Preview 的新型安全导向大语言模型。该模型被用于扫描 Cloudflare 自身超过五十个代码库以查找漏洞。结果显示，AI 在安全研究中的辅助能力大幅提升，尤其是在构建漏洞链和生成可行概念验证（PoC）方面。

Mythos Preview 能将多个低严重性漏洞串联成一个高严重性漏洞。与那些可能能发现漏洞却未能判断其可利用性的模型不同，Mythos Preview 能推理出将漏洞原语链在一起所需的步骤。它还有一个"证明生成"循环：模型会编写触发可疑漏洞的代码，进行编译和运行以确认漏洞存在；如果首次尝试失败，它会调整方法继续尝试。

然而研究也暴露出若干挑战。一个主要问题是所谓的"模型拒绝"——AI 会在某些情况下主动回绝合法的安全研究请求，而且这种拒绝表现并不一致，往往取决于任务的表述或上下文。此外，模型的"信噪比"仍然较低：它倾向于用"可能""潜在"等模糊措辞过度报告问题，从而可能淹没人工审查团队。

团队发现，简单把通用的编码代理指向代码库无法实现全面覆盖。于是他们开发了一个专门的执行框架来管理 AI 的运行。该框架把流程拆成多个狭义且并行的任务，而非一次性穷尽式搜索，包含侦察、基于限定提示的搜寻、对抗性验证以降低噪音，以及追踪以确认漏洞是否真正可被攻击者利用。

展望未来，Cloudflare 强调仅靠更快修补不足以应对这些日益强大的 AI 能力。他们建议从架构上增强防御，例如在应用前端部署更严密的防护，并设计系统以避免单一缺陷导致整体沦陷。 Cloudflare 还计划进一步披露其产品将如何帮助客户实施这些防御策略。 Cloudflare recently participated in Project Glasswing, a collaboration with Anthropic to test their new security-focused LLM called Mythos Preview. The model was tasked with scanning over fifty of Cloudflare's own codebases to identify vulnerabilities. The results showed a significant leap forward in AI-assisted security research, particularly in the model's ability to construct exploit chains and generate working proofs of concept.

Mythos Preview demonstrated a unique capability to combine multiple low-severity bugs into a single, high-severity exploit. Unlike previous models that might identify a bug but leave the question of exploitability open, Mythos Preview can reason through the steps required to chain primitives together. It also features a "proof generation" loop where it writes code to trigger a suspected bug, compiles it, and runs it to confirm the vulnerability, adjusting its approach if the initial attempt fails.

However, the research highlighted several challenges. One major issue is "model refusals," where the AI organically pushes back on legitimate security research requests. These refusals were found to be inconsistent, often depending on how a task was framed or the specific context provided. Additionally, the "signal-to-noise" ratio remains a hurdle, as models tend to over-report potential issues with hedged language like "possibly" or "potentially," which can overwhelm human triage teams.

The team discovered that simply pointing a generic coding agent at a repository is ineffective for comprehensive coverage. Instead, they developed a specialized "harness" to manage the AI's execution. This harness breaks down the process into narrow, parallel tasks rather than one exhaustive search. It includes stages for reconnaissance, hunting with scoped prompts, adversarial validation to reduce noise, and tracing to confirm if a bug is actually reachable by an attacker.

Looking ahead, Cloudflare emphasizes that simply patching faster is not a sufficient defense against these advancing AI capabilities. They advocate for architectural changes that make exploitation harder, such as implementing robust defenses in front of applications and designing systems where a single flaw cannot compromise the entire infrastructure. Cloudflare plans to share more on how their products will help customers implement these defensive strategies in the future.

Files.md 是一款简单且本地优先的应用，用纯 .md 文件来管理日常事务。它把自己定位为复杂"Second Brain"工具（如 Obsidian）的一种极简替代，认为过多的功能和模板会变成一种陷阱，反而拖延真正的思考。核心理念是：限制能激发创造力；因此工具被设计得极其简洁，代码库小到一个人或一个大语言模型都能完全理解。

该应用作为 Progressive Web App 直接在浏览器中运行，无需安装且支持离线使用。它强调用户对数据的所有权，所有内容都保存在本地 .md 文件中，并可选择通过单一二进制服务器或 iCloud 、 Dropbox 等云服务进行同步。一个关键功能是集成的 Telegram 聊天机器人，允许用户随手记录想法、笔记和任务，并将其保存到文件系统。

作者批评"Second Brain"概念，引用文章指出这类系统可能制造一种虚假的掌控感，成为拖延的借口。真正的价值在于用自己的大脑去思考笔记、在想法间建立联系并深化理解，而不是单纯地收集信息。文章也提醒不要让记笔记阻碍现实体验和情感疗愈。

实际使用很简单：打开网页应用、安装并选择一个本地文件夹，就能创建笔记、清单、日记和任务，这些内容按预设但灵活的结构组织。应用提供常用操作的快捷键和实用脚本，例如把 Whoop 指标加入日记或转换维基链接。项目是开源的，强调简洁、最小依赖以及便于初级开发者理解的代码。

技术上，后端使用 Go 编写，前端用 JavaScript，并附有详细指南。文章强调性能，指出互斥锁操作相比磁盘 I/O 可以忽略不计。还包含一系列 Architecture Decision Records（ADR），记录项目演进，比如从复杂的 WASM 实现回退到更简单的 JavaScript，以及统一采用纯 markdown 链接以提升跨平台兼容性。总体目标是打造一个可持续、可移植且无干扰的知识管理系统。 This article introduces Files.md, a simple, local-first application for managing life in plain markdown files. It positions itself as a minimalist alternative to complex "Second Brain" tools like Obsidian, arguing that excessive features and templates can become a trap that defers actual thinking. The core philosophy is that restrictions foster creativity, and the tool is designed to be extremely simple, with a codebase small enough for one person or an LLM to understand fully.

The application works directly in the browser as a Progressive Web App, requiring no installation and functioning offline. It emphasizes user ownership of data, storing everything in local .md files with optional synchronization via a single-binary server or cloud services like iCloud and Dropbox. A key feature is a Telegram chatbot that allows users to quickly dump thoughts, notes, and tasks on the go, which are then saved into the file system.

The author critiques the "Second Brain" concept, quoting an essay about how such systems can create a false sense of mastery and become a form of procrastination. The real value, they argue, comes from using your own brain to think through the notes, make connections between ideas, and develop deeper understanding, rather than just collecting information. The article warns against letting note-taking become a barrier to real-world experience and emotional healing.

Practical usage is straightforward: users open the web app, install it, and select a local folder. They can create notes, checklists, journals, and tasks, all organized in a predefined but flexible structure. The app includes hotkeys for common actions and useful scripts for tasks like adding Whoop metrics to journals or converting wikilinks. The project is open source, with a focus on simplicity, minimal dependencies, and code that junior developers can easily comprehend.

The technical backend is written in Go, with a JavaScript frontend, and the project includes detailed guidelines for both. It highlights performance, noting that mutex operations are negligible compared to disk I/O. The article also includes a series of Architecture Decision Records (ADRs) that document the project's evolution, such as moving from a complex WASM implementation to simpler JavaScript, and standardizing on plain markdown links for cross-platform compatibility. The overarching goal is to create a sustainable, portable, and distraction-free system for knowledge management.

生成式 AI 代表了技术领域最新一次重大平台转型，继大型机、个人电脑、互联网和智能手机之后。这类转变大约每隔 10 到 15 年发生一次，通过重新引导创新、投资和公司创建，从根本上改写科技行业格局。它们同时催生新的守门人和价值获取机制，并对科技行业以外的企业构成生存威胁或带来新机遇。微软就是一个典型例子：在个人电脑时代占据主导，但在智能手机革命中几乎失去影响力，说明平台转变能重置行业领导权。

当前这波 AI 浪潮由巨额资本支出驱动，四大科技公司计划在 2026 年投入 7000 亿美元，仅此一项就远超电信业的 3000 亿美元和石油天然气业的 1 万亿美元。 Sundar Pichai 和 Mark Zuckerberg 等领导人强调，与其担心过度投资，不如担心投资不足——即便这项技术还需数年才能完全成熟。这种激进的投入反映出业界普遍认为 AI 将定义下一代计算时代。

Nvidia 已成为这次转型的核心，营收增长迅猛，已可与 Intel 历史峰值相媲美。但 Nvidia 面临供应瓶颈，难以满足前所未有的芯片需求，TSMC 也无法快速扩产以跟上节奏。这引发了新一轮半导体投资周期，全球芯片出货量和营收飙升至前所未有的水平。如此规模的投资和基础设施建设，凸显了生成式 AI 在重塑科技乃至更广泛经济格局时的深远潜力。 Generative AI represents the latest major platform shift in technology, following mainframes, PCs, the web, and smartphones. These shifts occur every 10 to 15 years and fundamentally reshape the tech industry by redirecting innovation, investment, and company creation. They also create new gatekeepers and value capture mechanisms, while posing existential threats or new opportunities for businesses outside the tech sector. A key example is Microsoft, which dominated the PC era but became largely irrelevant during the smartphone revolution, illustrating how platform shifts can reset industry leadership.

The current AI wave is driven by massive capital expenditure, with the four major tech companies planning $700 billion in capex for 2026 alone. This dwarfs spending in other major industries like telecommunications ($300 billion) and oil and gas ($1 trillion). Leaders like Sundar Pichai and Mark Zuckerberg emphasize that the risk of under-investing in AI far outweighs the risk of over-investing, even if the technology takes years to fully mature. This aggressive spending reflects a belief that AI will define the next era of computing.

Nvidia has emerged as a central player in this shift, experiencing explosive revenue growth that now rivals Intel's historical peak. However, the company faces supply constraints as it struggles to meet unprecedented demand for chips, with TSMC unable to scale capacity fast enough. This has triggered a new semiconductor investment cycle, with global chip billings surging to levels not seen before, even accounting for the industry's traditionally cyclical nature. The scale of investment and infrastructure buildout underscores the transformative potential of generative AI as it begins to reshape both technology and broader economic landscapes.

Linus Torvalds 表示，Linux 内核安全邮件列表因为 AI 工具带来的大量重复漏洞报告，已经"几乎完全失控"。在他每周的内核状况更新中，他解释说，很多研究者用相同的 AI 工具去发现相同的漏洞，造成大量重复条目，邮件列表被冗余报告淹没。他将因此产生的那类讨论——把报告转给相关维护者或指出漏洞已修复——称为"毫无意义的空转"，浪费大家的时间。

Torvalds 认为，AI 发现的漏洞本质上并不是什么秘密，把它们放在私人安全列表里处理反而适得其反。他指出，这种做法会加剧重复报告的问题，因为报告者看不到彼此的提交，导致同一问题被反复上报。他提醒内核贡献者查阅项目文档，文档中已有相关说明，尽管他承认自己的措辞"比书面指导稍微委婉一些"。

尽管对 AI 报告的处理方式感到沮丧，Torvalds 也承认 AI 工具若被恰当使用仍然有价值。他敦促研究人员不要只把 AI 生成的结果原封不动地转发，而应通过阅读文档、提交补丁，并在 AI 发现的基础上提供有意义的分析来创造真正的价值。他特别批评那些没有深入理解就随意提交报告的"路过式"举报者，要求他们在贡献时更为慎重。

Torvalds 的评论与同为内核维护者的 Greg Kroah-Hartman 最近的说法形成对比。后者在接受 The Register 采访时表示，AI 已成为开源社区越来越有用的工具。虽然 Kroah-Hartman 指出 AI 生成的漏洞报告质量有所提升，Torvalds 的不满则凸显了在安全研究中广泛采用 AI 所带来的成长阵痛，尤其是在协调与避免重复工作方面。 Linus Torvalds has declared that the Linux kernel security mailing list has become "almost entirely unmanageable" due to a flood of duplicate bug reports generated by AI-powered tools. In his weekly state of the kernel post, he explained that multiple researchers are using the same AI tools to find the same bugs, creating enormous duplication and overwhelming the list with redundant reports. He described the resulting chatter, where people forward reports to the right maintainers or point out that bugs were already fixed, as "entirely pointless churn" that wastes everyone's time.

Torvalds argued that AI-detected bugs are by definition not secret, so treating them on a private security list is counterproductive. He pointed out that this approach only makes duplication worse because reporters cannot see each other's submissions, leading to the same issues being reported repeatedly. He directed kernel contributors to the project's documentation, which addresses this problem, though he admitted his own assessment was "a bit less blunt" than the written guidance.

Despite his frustration with how AI bug reports are being handled, Torvalds acknowledged that AI tools can be valuable if used productively. He urged researchers to go beyond simply forwarding AI-generated findings and instead add real value by reading the documentation, creating patches, and contributing meaningful analysis on top of what the AI discovered. He specifically called out "drive-by" reporters who send random reports without real understanding, asking them to be more thoughtful in their contributions.

Torvalds' comments stand in contrast to recent remarks from fellow kernel maintainer Greg Kroah-Hartman, who told The Register that AI has become an increasingly useful tool for the open-source community. While Kroah-Hartman has noted the improvement in quality of AI-generated bug reports, Torvalds' frustration highlights the growing pains that come with widespread AI adoption in security research, particularly around coordination and avoiding redundant efforts.

Auto-identity-remove 是一款开源 macOS 工具，能自动从 500 多个个人搜索网站和数据经纪数据库中删除个人信息。由 Stephen Thorn 开发，它通过 macOS 的 launchd 定期按月运行，自动完成从查找你的列表、填写退出表单到解决 CAPTCHA 的全部流程。完成后会通过 iMessage 发送摘要，并在浏览器中打开需要人工处理的网站。

项目最初覆盖 30 多家主要数据经纪商，并为每家明确映射了退出策略，包含 Spokeo 、 WhitePages 、 Intelius 、 BeenVerified 等。每家经纪商都有定制流程：有的先搜索个人资料、有的直接填写退出表单、有的通过 Mail.app 发邮件。系统采用智能表单检测而非硬编码选择器，对网站变动更具鲁棒性。 state.json 的状态追踪会记录已退出的经纪商，并在 90 天内跳过它们，因为数据经纪商通常会在该时间段内重新添加个人信息。

随着 generic-runner.js 的加入，覆盖范围从 31 家扩展到 500 多家，项目实现了重大扩展。 The Markup 的研究提供了 494 个退出 URL，BADBOOL 贡献了另外 27 个个人搜索站点。对于这些通用经纪商，脚本按顺序尝试四种策略：点击"Do Not Sell My Personal Information"按钮、通过 OneTrust 或 TrustArc 等常见隐私管理器退出、用你的信息填写通用退出表单，或记录 DSAR 链接以便手动跟进。此方法让工具无需为每个站点单独配置也能处理大量网站。

设置通过交互式 setup.js 完成，会收集你的个人信息、 CapSolver API 密钥、 iMessage 通知号码，并创建所需的一次性账户。敏感数据保存在本地，config.json 和 state.json 都被列入 .gitignore 。 CapSolver 集成可以约 $0.001 每次的成本处理带 CAPTCHA 的表单，适合定期运行；不使用 CapSolver 的话，这些带 CAPTCHA 的站点会被加入手动操作列表，而不会直接报错失败。

该工具将自己定位为 Incogni 、 Optery 等付费服务的免费、透明替代方案；这些付费服务按年订阅并覆盖更多经纪商、由专业团队维护流程。 auto-identity-remove 让用户能完全掌控并查看整个过程。作者建议两者结合使用：让付费服务处理大部分经纪商，本脚本补位处理 Acxiom 、 LexisNexis 、 ZoomInfo 、 Clearbit 等可能未被全面覆盖的网站。项目采用 MIT 许可证，欢迎贡献，尤其是添加带有已验证可用选择器的新经纪商。 Auto-identity-remove is an open-source macOS tool that automates the process of removing personal information from over 500 people-search sites and data broker databases. Created by Stephen Thorn, it runs on a monthly schedule using macOS's launchd system and handles everything from searching for your listings to filling out opt-out forms and solving CAPTCHAs. The tool sends an iMessage summary when complete and opens any sites requiring manual action directly in your browser.

The project started by covering 30+ major data brokers with explicitly mapped opt-out strategies, including sites like Spokeo, WhitePages, Intelius, BeenVerified, and others. Each broker has a tailored approach, whether that involves searching for your profile first, filling a direct opt-out form, or sending an email through Mail.app. The system uses smart form detection rather than hardcoded selectors, making it more resilient to site changes. A state tracking system in state.json remembers which brokers you've been removed from and skips them for 90 days, since data brokers typically re-add personal information within that timeframe.

A major expansion came with the addition of generic-runner.js, which brought coverage from 31 to over 500 brokers by incorporating two public datasets. The Markup's research provided 494 opt-out URLs, while BADBOOL contributed 27 additional people-search sites. For each of these generic brokers, the script tries four strategies in sequence: clicking a "Do Not Sell My Personal Information" button, opting out through common privacy managers like OneTrust or TrustArc, filling any generic opt-out form with your details, or recording a DSAR link for manual follow-up. This approach allows the tool to handle a vast number of sites without needing individual configuration for each one.

The setup process is handled through an interactive setup.js script that collects your personal information, CapSolver API key, iMessage notification number, and creates any required one-time accounts. Your sensitive data stays local, as both config.json and state.json are gitignored. CapSolver integration handles CAPTCHA-protected forms at roughly $0.001 per solve, making it affordable for regular use. Without CapSolver, those CAPTCHA-protected sites simply get added to your manual action list instead of failing with errors.

The tool positions itself as a free, transparent alternative to paid services like Incogni or Optery, which charge annual subscriptions. While paid services cover more brokers with professionally maintained flows, auto-identity-remove gives users full control and visibility into the process. The author actually recommends using both approaches together, with a paid service handling the bulk of brokers and this script filling in the gaps for sites like Acxiom, LexisNexis, ZoomInfo, and Clearbit that might not be fully covered elsewhere. The project is MIT licensed and welcomes contributions, particularly for adding new brokers with verified working selectors.

前 Google 首席执行官 Eric Schmidt 在 2026 年 5 月 17 日于 University of Arizona 的毕业典礼上谈及人工智能时多次被嘘。 Eric Schmidt 曾执掌 Google 长达十年，他开场回顾了计算机带来的深刻变革，追溯它从 1982 年被 Time 评为"年度人物"到如今笔记本电脑和智能手机的发展。他承认，尽管计算机把人们连在一起、让知识更普及、帮助许多人摆脱贫困，但它们也带来了负面影响——削弱了公共话语、放大了社会最坏的本能。

当 Eric Schmidt 把人工智能比作与计算机同样具有变革力的存在时，现场立刻并持续响起嘘声。他对观众的反应作出回应，承认毕业生们担心未来已被写定、担心机器会抢走工作、担心气候正在恶化，以及他们要继承一个并非自己造成的烂摊子。尽管遭遇敌意，Eric Schmidt 仍坚持认为未来尚未写成，2026 届的毕业生有真正的能力来塑造人工智能的发展方向。

他敦促毕业生拥抱自由、开展公开辩论、追求平等，并与持不同意见的人沟通。他还主张采纳多元观点，包括那些历来促进美国发展的移民视角，警告不要让美国失去对全球有抱负者的吸引力。最后，他向毕业生们表示祝贺，并告诉他们，未来还未定稿，现在轮到他们去塑造它。

University of Arizona 为邀请 Eric Schmidt 作辩护，发言人 Mitch Zak 称赞他在技术、创新和科学进步方面的卓越领导与全球贡献。 Zak 指出，Eric Schmidt 助力 Google 崛起为全球最具影响力的科技公司之一，并继续通过慈善与科研项目推动研究，包括与该校的合作。

Eric Schmidt 的遭遇并非孤立。 2026 年 5 月初，房地产高管 Gloria Caulfield 在 University of Central Florida 的毕业典礼上因提到人工智能同样被嘘，当时她告诉观众，人工智能的崛起是下一次工业革命。这些事件反映出更广泛的一代人对人工智能对就业、社会与未来影响的焦虑，毕业生们对那些将人工智能描绘为不可避免且积极变革的科技领袖抱有怀疑与恐惧。 Former Google CEO Eric Schmidt was booed multiple times during his commencement speech at the University of Arizona on May 17, 2026, as he discussed artificial intelligence. Schmidt, who led Google for a decade, began by reflecting on the transformative impact of the computer, tracing its evolution from a device named Time magazine's "Person of the Year" in 1982 to the laptops and smartphones of today. He acknowledged that while computers connected people, democratized knowledge, and lifted many out of poverty, they also carried negative consequences, degrading public discourse and amplifying society's worst instincts.

When Schmidt drew a parallel between AI and the computer's transformative power, the audience responded with immediate and sustained boos. He acknowledged the crowd's reaction, recognizing the fear among graduates that the future has already been written, that machines are coming for jobs, that the climate is breaking, and that they are inheriting a mess they did not create. Despite the hostile reception, Schmidt pressed on, arguing that the future remains unwritten and that the class of 2026 has real power to shape how AI develops.

Schmidt urged graduates to embrace freedom, open debate, equality, and engagement with those they disagree with. He also called for choosing a diversity of perspectives, including those of immigrants who have historically made America better, warning against losing the country's appeal to ambitious people worldwide. He closed by congratulating the class and telling them the future is not yet finished, that it is now their turn to shape it.

The University of Arizona defended its invitation to Schmidt, with spokesperson Mitch Zak citing his extraordinary leadership and global contributions in technology, innovation, and scientific advancement. Zak noted that Schmidt helped lead Google's rise into one of the world's most influential technology companies and continues to advance research through philanthropic and scientific initiatives, including partnerships supporting work at the university.

Schmidt's reception was not an isolated incident. Earlier in May 2026, real estate executive Gloria Caulfield was similarly booed at a commencement speech at the University of Central Florida after mentioning AI, telling the crowd that the rise of artificial intelligence is the next industrial revolution. These incidents reflect a broader generational anxiety about AI's impact on employment, society, and the future, with graduates expressing skepticism and fear toward tech leaders who frame AI as an inevitable and positive transformation.

本文的核心论点是，"vibecoding"——即指责 AI 工具让没有技能的人轻松产出复杂软件——是个神话。作者指出，尽管 AI 已经普及两年，但并未出现所谓的"vibecoded"等价物——像 Photoshop 、 Excel 或 Maya 这类复杂、架构性的软件仍然没有。这一缺失表明，AI 并没有真正降低构建非平凡、连贯系统的门槛。作者认为，对 vibecoding 的指控本身就是一种"slop"，即未经证实的断言，用来在没有证据的情况下否定他人的工作。

作者把软件开发分成三个层面：Level 1 是机械性地敲代码和处理语法；Level 2 涉及验证、测试与质量保证；Level 3 则是架构判断，决定要做什么、如何确保系统在现实中能站得住脚。 AI 虽然大幅降低了 Level 1 的成本和劳力，但对 Level 2 和 Level 3 几乎没有影响，而真正的门槛就在后两者。作者认为，那些指责别人 vibecoding 的人，往往是过度认同 Level 1 工作的人——当这一层被自动化时，他们会感到受威胁。

文章暗示，vibecoding 的指控是一种防御性反应：当人们看到 AI 辅助的作品，就会下意识地认为那很容易，是靠 AI 做出来的，然后把这种直觉当成结论发布出来。这样的指控之所以流行，是因为它"听起来对"，而不是因为它真的对。更具讽刺意味的是，指责他人产出未经验证的成果，本身就是一种未经验证的断言——指责者正在做的，正是他们指责 vibecoder 做的：提出没有定义、没有检验、不可证伪的说法。

作者以个人经验作证。像 SoulPlayer——一个带有严格 90 项测试验证框架的 C64 音乐播放器——以及一系列耗费数月、需要定制工具链的 AI 音乐视频，都说明严肃的 AI 辅助工作离不开深入的 Level 2 和 Level 3 工作。作者有 demoscene 背景，常被请去解决"别的办法都不行"的技术难题，这证明他们知道真正的门槛在哪里。尽管有资格去否定他人的成果，作者仍拒绝使用"vibecoded"这类指控。

拒绝这种指责既是道德上的，也是策略性的。作者意识到，这样的指控会消耗被指责者的时间和士气，迫使他们为自己辩护而不是继续创作。作者一生中多次遭遇排斥性指责——作为神经多样者、残障者、自由职业者和 demoscener——因此理解这种手法的形态，不愿予以复制。围绕 AI 使用的羞耻经济靠恐惧维系，而非基于真正可耻的行为；作者拒绝为其添柴。文章最后向指责者发出挑战：拿出证据来，展示那些据称会威胁职业的 vibecoded Photoshop 或其他复杂产物。 The central argument of this piece is that "vibecoding," the accusation that AI tools allow unskilled people to produce complex software effortlessly, is a myth. The author points out that despite two years of widespread access to AI, there are no "vibecoded" equivalents of complex, architectural software like Photoshop, Excel, or Maya. The absence of these artifacts suggests that AI has not actually lowered the barrier to creating non-trivial, coherent systems. The author argues that the accusation of vibecoding is itself a form of "slop," an unverified claim made to dismiss others' work without evidence.

The author breaks down software development into three levels. Level 1 is the mechanical act of typing code and syntax. Level 2 involves verification, testing, and quality assurance. Level 3 is architectural judgment, deciding what to build and how to ensure it holds together in the real world. While AI has significantly reduced the cost and effort of Level 1, it has not impacted Levels 2 or 3, where the actual "gate" of meaningful software creation resides. The author contends that those who accuse others of vibecoding are often those who over-identified with Level 1 work, and feel threatened when that layer is automated.

The piece suggests that the "vibecoding" accusation is a defense mechanism. When people see AI-assisted work, they assume it was easy because it was made with AI, and they post that feeling as if it were a finding. This accusation travels because it feels right, not because it is right. The author notes a bitter irony, the accusation that someone produced unverified output is itself being produced as unverified output. The accuser is doing the very thing they accuse vibecoders of, making claims without definition, testing, or falsification.

The author draws from personal experience to illustrate these points. Projects like SoulPlayer, a C64 music player with a rigorous ninety-test verification harness, and a series of AI music videos that required months of custom toolchain development, demonstrate that serious AI-assisted work involves deep Levels 2 and 3 engagement. The author has a demoscene background and has been the person called in when nothing else works, proving they know where the gate is. Despite having the credentials to dismiss others' work, the author refuses to use the "vibecoded" accusation.

The refusal to use the accusation is both ethical and strategic. The author recognizes that the accusation costs the target time and morale, forcing them to defend themselves instead of building. Having been on the receiving end of exclusionary accusations throughout life, neurodivergent, disabled, freelancer, demoscener, the author understands the shape of the move and will not replicate it. The shame economy around AI use runs on fear, not on actual shameful behavior, and the author refuses to feed it. The piece ends with a challenge to accusers to produce the evidence for their claims, the vibecoded Photoshops and other complex artifacts that supposedly threaten the profession.

理论物理学家 Carlo Rovelli 认为，所谓的"意识难题"是建立在过时二元论之上的伪问题。他指出，每当科学观念挑战人类自我形象，文化上就会出现抵触——从达尔文的进化论到当下关于意识的争论，例子屡见不鲜。 Rovelli 认为，理解意识之所以困难，并非因为它是超自然的，而是因为它是一种极其复杂的自然现象，类似雷暴或蛋白质折叠。他强调，更新我们对某种现象的认识并不会使其显得虚幻，就像把日落理解为 Earth 自转的结果并不会削弱它的美感一样。

Rovelli 直接质疑哲学家 David Chalmers 的框架：Chalmers 将解释大脑行为的"简单"问题与解释为何这些行为伴随主观体验的"困难"问题区分开来。 Rovelli 认为，所谓大脑过程与体验之间存在"解释鸿沟"的说法荒谬——它预设了在我们尚未理解某事时，就已知道理解它会是什么样子。他指出，这一想法之所以流行，是因为它维护了精神与自然分离的世界观，而 Spinoza 数百年前就预见并否定了这种看法。在 Rovelli 看来，不愿把意识视为自然的一部分，是反对科学进步的又一意识形态守旧。

Rovelli 的核心论点之一是：科学理解并非脱离经验，它本身就是关于经验的。他反对把科学当作一种从外部旁观世界的绝对客观叙述，强调我们正是所描述世界的一部分。我们的知识和理论是驾驭现实的工具，而不是某种"无处视角"的超然陈述。因此，第一人称的体验与第三人称的科学描述，仅是对同一大脑现象的不同视角。主观性并不神秘，它不过是一种特殊的视角；那种貌似"形而上学的鸿沟"，源自把科学模型误当成对终极实在的直接描写。

他还批评了 Chalmers 的"哲学僵尸"思想实验：一个行为上与人类完全相同却没有意识的存在。 Rovelli 认为，这是一种修辞手法，本身就预设了其试图证明的非物质意识。如果一个僵尸也会像我们一样通过内省确信自己有意识，那么该论证便自相矛盾。所谓僵尸只有在那些先入为主地假定存在形而上学鸿沟的人看来才有区别，这使得这一例子既难以令人信服，又更像对超验灵魂的怀旧，而非严密的论证。

Rovelli 最终断言，我们的心理生活与宇宙中其他现象在本质上是相同的。他认为，"意识""经验"只是发生在我们内部的事件的名称，原则上可以被外部观察者描述。心灵是用高级语言描述的大脑行为；第一人称和第三人称的视角都没有优先性，它们只是对同一事件的两种观察方式。他呼吁放弃意识讨论中有害的二元论，接受我们的灵魂与精神生活并不与基础物理学相悖——正如我们已经接受 Earth 与苍穹并无本质区别，人与其他动物也并无本质差异。 Carlo Rovelli, a theoretical physicist, argues that the so-called "hard problem of consciousness" is a false problem rooted in outdated dualistic thinking. He traces a historical pattern of cultural resistance to scientific ideas that challenge human self-image, from Darwin's theory of evolution to the current debate on consciousness. Rovelli suggests that the difficulty in understanding consciousness stems not from it being a supernatural phenomenon, but from it being an exceptionally complex natural one, much like thunderstorms or protein folding. He emphasizes that updating our understanding of a phenomenon does not make it illusory, just as understanding sunsets as a result of Earth's rotation does not diminish their beauty.

Rovelli directly challenges the framework introduced by philosopher David Chalmers, who distinguished between the "easy" problem of explaining brain behavior and the "hard" problem of explaining why such behavior is accompanied by subjective experience. Rovelli finds the concept of an "explanatory gap" between brain processes and experience to be nonsensical, as it presupposes knowledge of what we would understand if we currently understood something we do not. He argues that this idea is widely embraced because it preserves a worldview where spirit and nature are separate, a notion anticipated and rejected by Spinoza centuries ago. For Rovelli, the resistance to seeing consciousness as part of nature is the latest in a long line of ideological rearguard battles against scientific progress.

A core part of Rovelli's argument is that scientific understanding is not separate from experience but is entirely about it. He rejects the naive view of science as an objective account of the world observed from the outside, stating that we are part of the world we seek to describe. Our knowledge and theories are embodied tools for navigating reality, not disembodied views from nowhere. Therefore, the dualism between a first-person experience and a third-person scientific account is simply a difference in perspective on the same brain phenomenon. Subjectivity is not mysterious but a special case of a perspective, and the apparent "metaphysical gap" arises from mistaking our scientific models for direct accounts of an ultimate reality.

Rovelli also critiques Chalmers' thought experiment of the "philosophical zombie," a being that behaves identically to a human but lacks consciousness. He argues this is a rhetorical trick that assumes the very non-physical consciousness it seeks to prove. If a zombie would be convinced of its own consciousness through introspection, just as we are, then the argument becomes self-defeating. The zombie is only distinguishable by those who already assume a metaphysical gap, making it an unconvincing example that reflects nostalgia for a transcendent soul rather than a logical proof.

Ultimately, Rovelli asserts that our mental life is of the same nature as any other phenomenon in the universe. He posits that "consciousness" and "experience" are names for events that happen inside us, which could, in principle, be described by an external observer. The mind is the behavior of the brain described in a high-level language, and neither the first-person nor the third-person perspective is primary; they are two views on the same events. He encourages us to abandon the pernicious dualism of the consciousness debate and embrace the reality that our soul, or spiritual life, is consistent with our fundamental physics, just as we have accepted that Earth is not different from the heavens, or humans from other animals.

于 2026 年 5 月在 Ommen, Netherlands 举行的 Outline Demoparty 上发布的 "wake up! 16b" 是一件引人注目的 demoscene 作品：它在仅用 16 字节 x86 实模式 DOS 汇编代码的条件下，探索了算法密度的极限。运行时，程序把显存既当作计算空间又当作显示缓冲区，用以渲染无限的谢尔宾斯基三角形，同时将这些几何数据解释为发送到 PC 扬声器的音频。整段程序仅由八条指令组成，总共 16 字节，可谓极简编码的极致实践。

代码以 int 10h 开始，初始化视频模式 0（40×25 字符文本模式）。随后两条指令将数据段寄存器指向 0xB800（VGA/CGA 文本缓冲区的物理地址）。在该中断清屏时，BIOS 会用统一模式填满所有 2000 个字符槽：ASCII 0x20（空格）和颜色属性 0x07（黑底浅灰）。表面上看似空白，实际上是一块填充了可预测、均匀数据的画布。这种均匀性至关重要，因为数学递推依赖一致的初始状态；内存中的任意随机伪影都会破坏分形，就像意外的一位能损坏元胞自动机那样。

程序的核心通过一个循环实现累加前缀和式的变换。 lodsb 从内存加载一个字节到累加器 AL，随后 sub si, byte 57 将源索引向后调整 56 字节（补偿 lodsb 的隐式自增）。接着 xor [si], al 把累加器与该位置的内存做异或，并把结果写回内存，从而在内存中形成一种"运行前缀和"的累积效果。在用 add 代替 xor 、步长为 16 字节且起始值为 2 的简化模型里，数值会按二项式系数序列累积；文章用表格演示了 16 个内存单元在 16 次遍历中逐行累加的过程，说明了这一数学递推。

关键在于用异或替代加法。二进制加法会产生进位并影响相邻位平面，而异或丢弃算术进位，等价于模 2 加法。起始值为 2（二进制 00000010），因此只有第 1 位会被影响。这把位平面隔离开来，使得值在 0x00 与 0x02 之间纯粹切换，恰好对应 Stephen Wolfram 的初等元胞自动机中的 Rule 60 。文章用第二个表格验证了这一点：第 1 位的置位与清零模式形成了典型的谢尔宾斯基三角形结构。

音频生成是该设计中最优雅的部分之一。 out 61h, al 将计算出的字节直接送到端口 61h，该端口连接内部 PC 扬声器。端口的第 1 位直接控制扬声器振膜：为 1 时向外推动，为 0 时回位。因为算法专门隔离并翻转第 1 位，谢尔宾斯基三角形的几何结构就直接变成了驱动扬声器振膜的指令。 CPU 的执行速度决定了有效采样率，1 与 0 的序列产生不同脉宽和频率的方波。交替的行会生成更高频的方波，而三角形内部的大块 0 区域则制造出节奏性的停顿，从而把数学结构直接转化为可听的声音纹理。

向后步进 56 字节（通过 sub si, byte 57）而不是向前步进 16 字节，对视听输出影响显著。 16 字节步长会在 4096 次迭代内完成一段扫掠，而 56 并不能整除 65536，需要 8192 次迭代才能访问所有地址，并在回到零之前绕段循环 7 次。这使宏周期加倍、基频减半，听起来下降一个八度，产生更慢更沉的音色。程序把结果写入 ASCII 字符字节：第 1 位用于控制扬声器，其余 7 位则演化出伪随机的 ASCII 字形，作为视觉纹理。值得注意的是，把这些混合数据发送到端口 61h 并不会导致系统崩溃，因为在标准 DOS 环境和常见模拟器中，多余的位会被无害地忽略。

在 40 字符宽显示上，-56 字节步进会产生对角剪切效果。向后 56 字节等同于在 80 字节行宽上向前 24 字节，即恰好 12 列；每个字符占 2 字节，因此序列每步上升 1 行并向右移动 12 列。由于 gcd(12,40)=4，循环恰好访问四分之一的列（40 列中的 10 列），于是分形以 10 列等间距的竖列形式在屏幕上向上展开，而非呈现为连续图像。

文章最后指出了对内存初始状态的依赖。理论模型假设环境被完美初始化，但实际中不同的系统配置、 VGA BIOS 实现和模拟器会在 RAM 中留下各种伪影。由于算法不断地读取并与现有内存内容做异或，它会直接与这些位交互，因而输出对运行环境高度敏感：视觉字符与音色在不同机器或模拟器上可能明显不同。虽然加入初始化例程以显式清除内存可以保证输出一致，但那会超出严格的 16 字节限制。接受这种微妙的不确定性，恰是在人为极限条件下创作的魅力所在。 Released at the Outline Demoparty in May 2026 in Ommen, Netherlands, "wake up! 16b" is a remarkable demoscene production that explores algorithmic density within just 16 bytes of x86 real-mode DOS assembly code. When executed, it uses the computer's video memory as both a calculation space and a display buffer to render an infinite Sierpinski fractal, while simultaneously interpreting that geometry as audio data sent to the PC speaker. The entire program consists of only eight instructions, totaling 16 bytes, making it an extreme exercise in minimalist coding.

The code begins with `int 10h`, which initializes Video Mode 0, setting up a 40x25 text mode display. The next two instructions point the Data Segment register to `0xB800`, the physical memory address of the VGA/CGA text buffer. When the BIOS clears the screen during this interrupt, it fills all 2,000 character slots with a uniform pattern: ASCII byte `0x20` (Space) and color attribute `0x07` (Light Gray on Black). While the screen appears blank, it is actually a canvas primed with predictable, uniform data. This uniformity is critical because the mathematical progression relies on a consistent starting state. Any random artifacts in memory would disrupt the fractal pattern, much like an unexpected bit can corrupt a cellular automaton.

The core engine of the program operates through a loop that creates additive prefix sums. The `lodsb` instruction loads a byte from memory into the accumulator `AL`, then `sub si, byte 57` adjusts the source index backward by 56 bytes (accounting for the implicit increment from `lodsb`). The `xor [si], al` instruction then XORs the accumulator value with the memory at that location, storing the result back. This creates a running prefix sum effect where values accumulate across memory. In a simplified model using `add` instead of `xor` with a step of 16 bytes and starting value of 2, the values follow a binomial coefficient sequence. The article includes a detailed table showing how values accumulate row by row across 16 memory cells over 16 passes, demonstrating the mathematical progression.

The key insight comes from replacing addition with XOR. When performing binary addition, bit-planes carry over into adjacent positions, but XOR discards the arithmetic carry, effectively performing addition modulo 2. Since the starting value is 2 (binary `00000010`), only Bit 1 is ever affected by this calculation. This isolates the bit-planes and creates a pure toggle between `0x00` and `0x02`, mapping perfectly to Rule 60 in Stephen Wolfram's elementary cellular automata. The article validates this with a second table showing the binary propagation, where the pattern of Bit 1 being set or cleared forms the characteristic Sierpinski triangle structure.

The audio generation is perhaps the most elegant aspect of the design. The instruction `out 61h, al` sends the computed byte directly to port `61h`, which interfaces with the internal PC speaker. Bit 1 of this port directly controls the speaker cone, pushing it outward when set to 1 and returning it when set to 0. Because the algorithm specifically isolates and toggles Bit 1, the geometry of the Sierpinski triangle serves as direct instructions for the speaker cone. The CPU's execution speed establishes the functional sample rate. The patterns of 1s and 0s generate distinct square waves with varying pulse widths and frequencies. Alternating rows produce higher-frequency square waves, while larger blocks of zeros (the empty regions within triangles) create rhythmic pauses, resulting in a direct sonic representation of the mathematical structure.

The choice of stepping backward by 56 bytes (via `sub si, byte 57`) rather than forward by 16 has significant consequences for both audio and visual output. While a 16-byte step completes a segment sweep in 4,096 iterations, 56 does not divide 65,536 evenly, requiring 8,192 iterations to hit all addresses and wrapping around the segment 7 times before returning to zero. This doubles the macro-cycle length, halving the fundamental frequency and dropping the auditory rhythm by one octave, producing a slower, deeper tone. The algorithm writes to ASCII character bytes where Bit 1 controls the speaker, while the remaining seven bits mutate into pseudo-random ASCII glyphs for visual texture. Remarkably, sending this mixed data to port `61h` doesn't crash the system, as the extra bits are harmlessly ignored in standard DOS environments and emulators.

Visually, the -56 byte step creates a diagonal shearing effect on the 40-character wide display. Stepping backward by 56 bytes is equivalent to moving forward by 24 bytes on an 80-byte grid, which equals exactly 12 columns. Since each character space occupies 2 bytes, the sequence progresses up 1 row and right 12 columns per step. With gcd(12,40)=4, the loop visits exactly one-quarter of the columns (10 out of 40), rendering the fractal as 10 evenly spaced vertical columns ascending the screen rather than a contiguous image.

The article concludes with an important observation about memory dependency. The theoretical model assumes a perfectly initialized environment, but in practice, different system configurations, VGA BIOS implementations, and emulators can leave varying artifacts in RAM. Since the algorithm continuously reads and XORs against existing memory contents, it interacts directly with these bits, making the output highly sensitive to its environment. Visual characters and audio timbre may vary noticeably across different machines or emulators. While adding a setup routine to explicitly clear memory would ensure identical output, it would exceed the strict 16-byte limit. Embracing this subtle unpredictability is presented as part of the charm of working within such extreme constraints.