2026 06 15 HackerNews

2026-06-15 Hacker News Top Stories #

  1. 保罗·格雷厄姆认为初创公司通过用户真心喜爱带来的指数增长可合法赚取十亿美元,年轻创业者应关注自己这代人的未满足需求。
  2. 美国生成式AI使用呈约三成积极、三成偶尔、三成不用的三分格局,社会净正面评价仅+8%,企业需提供多样选项以适应从接受到拒绝的连续谱。
  3. 本田思域车机因使用公开测试密钥签名,攻击者通过USB物理接入即可执行任意代码,作者已发布工具并呼吁社区协助。
  4. 英国一名警察因用AI伪造证据被调查,此前已有警方因AI虚构信息误禁球迷观赛并道歉,凸显AI在警务中的风险。
  5. 一款免费浏览器端工具可将SQL建表语句转为交互式ER图,所有处理本地完成,保护隐私,支持多种数据库和导出。
  6. 开源Windows兼容系统ReactOS在真实硬件上成功运行了3D加速的经典游戏《半条命》,标志着28年项目的重要进展。
  7. 里约热内卢宣称的本土LLM实为对Nex和Qwen模型的元素级合并,未做后训练,证据包括自称倾向与权重分析高度共线。
  8. LLM的超长上下文窗口实际仅前10万token左右有效,之后注意力下降,应通过压缩历史或新会话保持高效区。
  9. 未发行的Game Boy外设Workboy能将掌机变为个人助理,其ROM在2020年泄露后被爱好者成功复原,引发怀旧关注。
  10. Linus Torvalds发布了Linux 7.1,主要包含各类驱动和工具的小幅修复,并提醒因旅行导致合并窗口时间不规律。

1. 如何赚到十亿美元 (How to earn a billion dollars) #

https://paulgraham.com/earn.html

我在牛津联盟的演讲中,反驳了某位美国政治家“不可能赚到十亿美元”的说法。作为 Y Combinator 的联合创始人,我亲眼见证了数百家初创公司通过合法途径成为亿万富翁。关键在于理解指数增长:一家月增长率 93% 的公司,从 200 万美元估值起步,仅需 9.5 个月即可达到十亿;即使月增长率只有 15%,五年后收入也能增长 4000 倍,足以让创始人成为亿万富翁。这种增长并非靠欺骗,而是因为用户真心喜爱产品并主动推荐。对年轻创业者而言,最有效的方法是发现并解决自己的真实需求——因为你这一代人的需求往往预示着未来的大趋势。


HN 热度 396 points | 评论 1190 comments | 作者:kingstoned | 11 hours ago #

https://news.ycombinator.com/item?id=48526360

  • 对文章的大量负面评论感到沮丧,认为人们没有真正反驳文章内容,而是进行空洞的意识形态批判
  • 对于“赚取十亿美元”存在两种理解:一种认为只要合法获得就是赚取,另一种认为即使合法获得,也不一定意味着创始人“赚取”了全部价值
  • 创业公司有机增长时创始人致富算“赚取”,但若增长依赖资本(投资、贷款或已有财富),那么成功有多少归功于创始人、多少归功于富人的任意选择就变得模糊
  • 文章以一种居高临下、侮辱性的方式回应批评,使用稻草人谬误和废话来误导读者
  • 辩论双方缺乏共同基础,许多人根本不懂金钱的本质,也没有真正理解对方观点的意愿
  • 创始人即使只赚到一亿美元也足以让家族世代奢侈生活,不必非要追求十亿美元才肯创业
  • 创始人创造公司并拥有所有权,公司通过 IPO 被赋予高价值,这不是他银行账户里的现金,也不是从别人那里抢来的
  • 当公司规模超过一亿美元后,无法阻止创始人变得更富,除非通过税收或强制公司分拆清算
  • 市场本应通过竞争消除超额利润,但公司往往通过反竞争和提取性行为来阻止竞争,从而实现十亿级价值
  • 所有者将工人创造的价值仅锁定在家族内部代代相传,最终会导致封建主义

2. 并非人人都用 AI 做所有事 (Not everyone is using AI for everything) #

https://gabrielweinberg.com/p/people-are-consuming-ai-like-they

根据最新调查数据,美国人对生成式 AI 的实际使用远非“人人都在用 AI 做所有事”的传闻。多项研究(盖洛普、微软、Datos、Searchlight 研究所等)一致显示:约三分之一的人积极使用 AI(每月至少 90 分钟),三分之一偶尔使用,三分之一从未使用。其中 Z 世代的 AI 采用率已停滞,负面情绪(尤其愤怒)显著上升。人们限制使用的主要原因包括对失业、隐私和虚假信息的担忧,以及对 AI 实际价值的怀疑——AI 的净正面社会评价仅 +8%,与社交媒体相当,远低于手机、互联网和太阳能。作者以肉类消费为类比:95% 美国人吃肉,但 70% 有意减少红肉,4% 是素食者;AI 使用也存在类似的文化偏好连续谱,企业应提供从私有 AI 到完全关闭的多种选项。


HN 热度 393 points | 评论 429 comments | 作者:yegg | 8 hours ago #

https://news.ycombinator.com/item?id=48527700

  • 求职者在面试中回答关于 LLM 使用的问题时需要权衡,既要避免让 AI 热衷的雇主失望,也不能让 AI 谨慎的雇主反感,只能给出模棱两可的答案。
  • 诚实回答是最好的策略,因为愿意接受真实回答的雇主通常拥有更好的企业文化,不诚实反而可能进入有毒的工作环境。
  • 但求职是为了生存,不得不考虑雇主的期望,不能完全理想化。
  • 如果不知道雇主想要什么答案,概率上给出最可能正确的回答更稳妥。
  • 坦诚、细致、有道理的回答会被好企业文化的雇主认可,但在当前就业市场是一种奢侈品。
  • 面试官很容易察觉候选人在回避问题,这会成为负面信号。
  • 有人面试时连续三次回避同一个问题,导致面试官反感;也有观点认为候选人因不确定答案而回避是合理的,面试官不该为难。
  • 美国缺乏社会安全网,工作关乎医保和生计,因此求职者必须谨慎回应。

3. 本田思域与邪恶代客 (Honda Civics and the Evil Valet) #

https://juniperspring.org/posts/honda-evil-valet/

三年前,作者开始逆向工程自己的 2021 款本田思域车机。最新进展是发现了严重的“邪恶代客”攻击漏洞:本田使用公开的 AOSP 测试密钥签名更新文件,攻击者只需通过 USB 物理接触前端口,即可通过更新路径执行任意代码,无需常规 root 权限。作者已发布 ota-builder 工具,可轻松制作被车机接受的更新文件。另一工具 apk-rebuilder 可自动处理官方更新文件,输出反编译后的代码结构,便于逆向分析。目前仍需社区贡献:记录不同版本车机的软件版本信息、完善工具链、探索自定义主题(因硬编码资源 ID 实现困难)、改进 AIDL 接口解析工具。作者认为与其维护繁琐的参考文档,不如提供可靠工具让 LLM 直接分析代码。项目并未废弃,欢迎提交 PR。


HN 热度 382 points | 评论 92 comments | 作者:librick | 22 hours ago #

https://news.ycombinator.com/item?id=48523080

  • 本田 10 代思域使用 AOSP 测试密钥签名更新包,物理访问 USB 即可执行任意代码。
  • 大多数汽车娱乐系统安全性差,传感器使其成为移动监控平台,澳大利亚政府已禁止在车内进行敏感操作。
  • 汽车制造商应在完全开放和完全安全之间选择,当前状态既侵犯隐私又不安全。
  • 可以实现默认安全但允许用户解锁(类似安卓手机),但需解决车主认证和安全重置等技术难题。
  • 原厂头单元更耐用(可使用十余年),后装市场性能更好但寿命可能较短。
  • 物理访问时安全模型失效,但不应因此攻击更开放的设备。

4. 警察因在多起案件中用 AI“制造证据”被调查 (Police officer investigated for using AI to ‘create evidence’ in multiple cases) #

https://news.sky.com/story/derbyshire-police-officer-investigated-for-using-ai-to-create-evidence-in-multiple-cases-13553661

一名德比郡警察因涉嫌在多起案件中利用人工智能“制造证据”而接受调查。皇家检察署表示正与德比郡警方合作处理此事,并与可能受影响的辩护团队和法院进行沟通。该警察已被调离一线岗位,等待调查结果,目前无人被捕。

本周,英国刚启动国家警务 AI 中心“PoliceAI”,旨在负起责任地运用 AI 打击犯罪。此前西米德兰兹警方曾因 AI 虚构信息,错误地禁止了以色列马卡比特拉维夫足球俱乐部球迷前往阿斯顿维拉观看比赛,事后警方向公众致歉。


HN 热度 369 points | 评论 188 comments | 作者:austinallegro | 1 day ago #

https://news.ycombinator.com/item?id=48520807

  • 摄像头硬件签名认证机制存在缺陷,因为可以翻拍高分辨率伪造图像,且密钥容易被攻破。
  • 通过增加深度、立体图像、视频片段、防篡改硬件等增强签名可信度,虽然不完美但能提高造假难度。
  • 应当将相机与操作者绑定、强制警察佩戴、实时上传低分辨率数据到云端并记录周围环境信息。
  • 现有的图像篡改检测工具并不可靠,尤其是元数据丢失后更难判断。
  • 过度保证签名系统会让人们误以为它绝对可靠,反而使攻击更有价值。
  • 公众普遍缺乏足够的技术素养来正确设置、监管或使用这类技术。
  • 伪造证据的“不完美”反而可能是系统有意为之的特性。
  • 想知道具体是何种伪造以及如何被发现的,是工具检测还是警察技术不精。
  • 许多 40 岁以上人群难以识别明显 AI 伪造图像,而年轻人有直觉但说不清原因。
  • AI 生成图像常用电影构图,这与手持拍摄的现实不符。
  • 现在用专业相机和灯光拍日常照片会被误认为 AI 生成。
  • 在图像和视频可随意生成的时代,大量证据类别可能完全不可靠。
  • 法庭往往不承认证据在理论上的不可靠性,很多法医学证据本就可疑。
  • 早在 15 年前就有公司能伪造 DNA 证据。
  • 律政剧(如《法律与秩序》)中的伪科学表述会影响陪审员对真实证据的判断。
  • “CSI 效应”使陪审员高估或低估特定类型证据的可信度。

5. 免费 SQL→ER 图工具,在浏览器中运行,不上传任何内容 (Free SQL→ER diagram tool, runs in the browser, nothing uploaded) #

https://sqltoerdiagram.com/

这是一个免费的在线工具,可将 SQL CREATE TABLE 语句自动转换为交互式实体关系图(ERD)。支持 PostgreSQL、MySQL、SQLite 和 SQL Server 等数据库。无需注册或安装,所有数据仅在浏览器本地处理,不会上传到服务器。可拖拽调整表布局、自动排列、添加注释,并导出为 PNG 或 SVG 格式。工具还提供了示例电商数据库模式(包含 users、addresses、products、orders、order_items、reviews 六张表)供快速体验。常见问题涵盖如何生成 ER 图、支持的 SQL 方言、隐私安全、导出方式等。


HN 热度 335 points | 评论 65 comments | 作者:robhati | 19 hours ago #

https://news.ycombinator.com/item?id=48523992

  • SQL 生成的 ER 图是物理级别,无法完全还原逻辑或概念层面的 ER 图
  • 实体与表本质上不同,但 ORMs 的广泛使用表明它们在大多数情况下可互换
  • 该工具对于探索未知数据库非常有用,可以辅助理解表结构和关系
  • 工具在移动端表现极佳,平移、缩放、选择等操作流畅
  • 代码库简洁优雅,作者将复杂问题简化为简单解决方案
  • 工具基于 canvas 实现,使用缓存位图和平截头剔除,支持数百个表
  • SQL 解析器跟踪源位置以实现精准编辑,保留注释和格式
  • 整个模式编码在 URL 中,无需后端,但 URL 长度限制可能是个问题
  • 支持 JSON 导出作为备选方案
  • 缺少导出为 SVG 的 CLI 选项,不利于在版本控制中存储
  • 希望能够选择直线和直角连线,而不是弯曲的连线
  • 布局算法(平面嵌入)是一个困难但有趣的问题
  • 与 explain.dalibo.com 类似,但用于查询计划可视化
  • 相比其他工具,无付费墙、无需注册、数据不离开本地

6. ReactOS(自由开源“Windows”)在真实硬件上实现了 3D 加速的《半条命》运行 (ReactOS (FOSS “Windows”) achieves 3D-accelerated Half-Life on real hardware) #

https://www.phoronix.com/news/ReactOS-Running-Half-Life

ReactOS,这个致力于与微软 Windows 程序和驱动二进制兼容的开源操作系统,近日达到了一个里程碑:能够运行经典游戏《半条命》。该项目已开发 28 年,开发者通过 X 平台分享了这一进展。虽然通过 Wine 在 Linux 上运行《半条命》早已不是问题,但能在 ReactOS 上直接运行仍令人兴奋。用户“Zombiedeth”使用戴尔 OptiPlex(酷睿 i5-2400 Sandy Bridge 处理器、NVIDIA GeForce 8400GS 显卡)成功运行了该游戏。


HN 热度 272 points | 评论 72 comments | 作者:jeditobe | 24 hours ago #

https://news.ycombinator.com/item?id=48522486

  • 结合 ReactOS 和 Good Old Games,制作 USB 启动的复古 Windows 游戏分发版,用于 LAN 聚会
  • 用 Linux 发行版搭配 Wine 和应用快捷方式也能实现类似效果,做成“复古 LAN 派对发行版”
  • 为 Windows 游戏定义“ROM 格式”有助于兼容性
  • 需要像 WineHQ 那样列出所有在 ReactOS 上运行良好的游戏,或开发专门针对游戏的 ReactOS 变体
  • 开源终将胜出,因为越来越多的人参与编程
  • ReactOS 开发 28 年仍进度缓慢,接近“圣家堂”或“百万猴子打字机”式的长期工程
  • 开源项目需要保持与当前计算模型的相关性,ReactOS 远落后于 Windows 11,尤其在 ARM 和 CoPilot+ PC 方面
  • ReactOS 对逃避政府压迫和数字主权仍有价值
  • 与 Windows 11 相比,ReactOS 没有烦人的广告、数据窃取和硬件强制升级
  • Windows 11 有实用功能,如单次管理令牌、应用沙箱和签名、容器跨内核版本使用,这些连 Linux 都未完全赶上
  • Windows 11 的 emoji 选择器很出色,但希望支持更多 Unicode 字符
  • 许多旧工业/政府机器仍运行 Windows 95/3.11,ReactOS 可作为安全更新的替代方案
  • 将 ReactOS 移植到新 CPU 架构(如 ARM)比微软移植 Windows 更容易

7. 里约热内卢的“本土”大语言模型似乎是对现有模型的合并。 (Rio de Janeiro’s “homegrown” LLM appears to be a merge of an existing model) #

https://github.com/nex-agi/Nex-N2/issues/4

这是 GitHub 上一个公开的 Issue,由 Nex-AGI 团队的 00INDEX 发帖,指控 prefeitura-rio 发布的 Rio-3.5-Open-397B 模型并非原创,而是直接用 Nex 和 Qwen 的权重按固定比例(约 0.6 Nex + 0.4 Qwen)进行元素级合并得到的,没有进行任何自主训练。

证据包含两部分:

  1. 身份测试:去掉 Rio 内置的“You are Rio”系统提示后,向模型提问 120 次,79% 的回答自称“Nex”,73% 自称来自“Nex-AGI”,从未自称“Rio”。它甚至原样复述了 Nex-AGI 独有的组织介绍(上海创智学院等)。
  2. 权重分析:对模型所有 60 层、387B 参数的路由专家、注意力层等张量进行测量,混合权重 α 稳定在 0.571(标准差仅 0.0016),共线性 cos_fit 高达 0.98-0.99。对于参数数量达数十亿的张量,这样的共线性在数学上完全不可能是独立模型之间的巧合,证明了 Rio 的权重就是 Nex 和 Qwen 的直接插值。

HN 热度 244 points | 评论 130 comments | 作者:unrvl22 | 7 hours ago #

https://news.ycombinator.com/item?id=48528371

  • 实际模型是现有模型 Nex 和 Qwen 的混合,并非从头训练或声称的后训练成果。
  • 线性合并权重在多个基准上表现提升,但这可能只是过度调参的结果,实际使用效果往往更差。
  • Nex 本身是 Qwen3.5 的微调,因此合并同基座微调有效,不同预训练的模型无法直接合并。
  • 神经网络优化表面较平滑,线性组合有效,已有相关论文支持。
  • 团队未主动宣传,是市长利用免费曝光,且上传的模型缺少蒸馏步骤。
  • 事件令人意外,里约本土 LLM 的标题对巴西人来说很震撼。
  • 线性组合之所以有效,可能因为原始模型性能本身就很差。
  • 这显示出 LLM 是一种极其浪费的方法,或者说明智能就是冗余组件的组合。
  • 团队声称存在后训练但实际没有,HF 页面后来改为“合并”,他们解释是上传错误。
  • 有观点认为可能存在一组秘密调整能让小模型大幅提升智能。

8. 不要相信大的上下文窗口 (Don’t trust large context windows) #

https://garrit.xyz/posts/2026-05-06-dont-trust-large-context-windows

大型语言模型的上下文窗口实际有效部分远小于厂商宣传的数字。作者将其分为"智能区"(约前 10 万 token)和"愚蠢区"(之后注意力下降、遗忘)。编码代理容易快速消耗 token 进入愚蠢区,而 200k、1M 乃至 2M 的窗口只是营销数据。

解决方法:现代工具如 Claude Code 会自动压缩历史,但压缩过程本身也受退化影响。作者建议手动开启新会话并传递自己编写的规范,或通过小型命名工件(如 PRD、计划)将信息移出会话,保持会话始终处于智能区。本质是将上下文窗口视为有限预算,尽量用书面工件替代实时对话中的信息。


HN 热度 239 points | 评论 178 comments | 作者:computersuck | 17 hours ago #

https://news.ycombinator.com/item?id=48524620

  • 非确定性让人感到不安,但同时也能通过一次转换将非确定性转为确定性,从而减少依赖
  • 评测需要在中等复杂度但有边界的任务上进行,结果往往高度一致,而 RTK 等工具反而增加了总体上下文使用
  • 技术社区中总是存在“模仿心态”(cargo culting),但 LLM 加剧了这种情况,因为模型是不透明且难以理解的
  • 网友对 AI 工作流的看法两极分化:一边认为不再需要回到之前的确定性工作流,一边觉得 LLM 世界缺乏严谨和可验证性
  • 使用 LLM 就像在跟人协作,但这让很多喜欢确定性、讨厌和人类合作的 IT 人员感到不适
  • 即使模型不同,在给定工具和提示下,输出结果高度一致,但许多用户依然依赖模糊的“直觉”而非可重复的评估
  • 讨论中的建议常常沦为类似园艺或 DIY 的非正式经验分享,失去了可辩论和被批评的基础
  • 对于工具调用的不确定性,网友认为“为什么不能用更确定的方式训练模型”是合理的疑问
  • 非确定性到确定性的转换在自然语言理解中尤其困难,因为同样的表述可能指向完全不同的场景
  • 许多 IT 决策并非基于客观标准,而是模仿“更有声望的公司”的做法,这在 AI 领域同样普遍

9. GameBoy Workboy 游戏 (GameBoy Workboy) #

https://tcrf.net/Workboy

Workboy 是一款为 Game Boy 开发的游戏,其主要功能是将 Game Boy 转变为一个微型工作站,配备键盘,能够帮助用户管理重要的约会、地址、笔记、银行账户余额和电话号码,并且可以在五种语言之间转换温度、货币和单词。尽管其功能听起来十分实用,但这款游戏并没有正式发布,尽管在当时的许多视频游戏杂志中曾有过大量广告宣传。

根据 2020 年 9 月的任天堂泄漏资料,Workboy 的 ROM 文件在这次泄漏中被发现。尽管该游戏从未完成或公开发布,但该游戏的键盘曾被认为遗失。然而,在 2020 年 12 月,DidYouKnowGaming? YouTube 频道发布了一段由 Liam Robertson 制作的视频,展示了他如何使用一个原型键盘与泄漏的 ROM 在 Game Boy 上运行 Workboy。这段视频提供了关于该游戏背景的更多细节,包括 Liam 是如何获得这些资料的。

在技术细节方面,Workboy 的标题屏幕显示的版本号为 8.87 其代码字符串中实际的版本号为 5.74。根据不同语言的本地化,该游戏在标题屏幕上显示的文本已经从 ROM 中提取。每种语言的文本在 ROM 中的偏移地址均有所不同,包括英语、西班牙语、意大利语、德语和法语等版本。在显示标题屏幕之前,相关文本字符串会被复制到中,然后通过从 ROM 中读取特定地址的方式,将 “5.74” 替换为 “8.87”。

总的来说,Workboy 虽然是一款未能正式发布的游戏,但由于其独特的功能和历史背景,引起了不少怀旧和游戏开发爱好者的关注。


HN 热度 217 points | 评论 74 comments | 作者:tosh | 1 day ago #

https://news.ycombinator.com/item?id=48519552

  • GameBoy WorkBoy 是一个未发布的硬件附加组件和生产力软件,最近被恢复。
  • Playdate 设备有趣且适合独立游戏,但价格高、规格低,游戏生态和摇杆使用可能不持久。
  • 相比 Playdate,Anbernic 等廉价设备(40-60 美元)运行 Linux,可玩模拟器、听播客、运行自制应用,性价比更高。
  • Playdate 的 Lua API 很好,适合编程新手,但许多游戏过于依赖摇杆。
  • TCRF 网站因 AI 爬虫和恶意流量大量增加,被迫屏蔽 VPN 和云托管 IP。
  • 屏蔽 VPN 用户是合理的,因为许多 VPN 被用于发布仇恨言论或破坏行为。
  • 自建 VPN 难以逃过屏蔽,因为很多网站也会屏蔽 VPS 地址。
  • VPN 的匿名性已不可靠,主要用于加密和绕过区域限制。
  • 有用户为 Playdate 开发了浏览器和非游戏应用(如新闻镜像),认为该设备有替代 Android/iOS 双头垄断的潜力。

10. Linux 7.1 版本发布 (Linux 7.1) #

https://lore.kernel.org/lkml/CAHk-=wi4BF4bMhZNZ1tqs+FFV4OuZRe3ZqdWB+LxRLmRweUzQw@mail.gmail.com/T/#u

Linux 7.1 版本发布。Linus Torvalds 在周日下午(当地时间)发布了该版本,合并窗口将于明天开启,但由于他身处不同时区且可能面临长途飞行,合并窗口的时机将有些不规律。过去一周的更新主要是各种较小的驱动程序更新(GPU、网络、声音、杂项)以及一些网络和 trace 工具修复,其他区域有零散的小变动。Linus 希望用户继续测试,并预祝合并窗口顺利。


HN 热度 200 points | 评论 75 comments | 作者:berlianta | 7 hours ago #

https://news.ycombinator.com/item?id=48528729

  • Arch Linux 当前默认内核是 7.0.10,正期待 7.1 尽快推送。
  • 页面加载时闪过一个动画头像,那是 Anubis 机器人防护系统的标志。
  • 有人反驳批评 Anubis 的文章,认为它“理论上不行但实践中有效”,且批评者并未提供替代方案。
  • Anubis 的实际效果可以通过对比 Apache 日志证明:拦截了大量 bot 流量。
  • 有人分享 uBlock Origin 过滤规则来屏蔽 Anubis 的动漫女孩图片。
  • 有人觉得动漫形象不够专业,但付钱给作者就可以去掉。
  • 部分人对卡通女性形象反应过度,甚至因此攻击开发者,被认为是一种“chud 防御”。
  • Linux 7.1 移除 ISDN 等老旧驱动代码,是为了减少 AI 产生的无用 bug 报告,这是 AI 带来的正面副作用。
  • 旧代码曾是资产但如今是负债,AI 让忽视过的隐患更难回避。
  • 也有观点指出,当年这些驱动确实有用,不能简单视为“脂肪”。
  • 版本号从 7.0 跳到 7.1 只是因次要数字太大,没有特别意义。
  • LWN 和 Phoronix 已有详细的 7.1 特性汇总文章。
  • 有用户喜欢自己的廉价 Linux 设备(如 Luckyfox Pico Mini)在 7.1 中得到支持。
  • Linus Torvalds 正在旅行,没人愿意为他买机上 WiFi 所以发布较平淡。

Hacker News 精彩评论及翻译 #

Statement on US government directive to suspend ac… #

https://news.ycombinator.com/item?id=48512685

So many comments here missing the big picture, and just gleefully pointing out that Anthropic got what they deserved, or that this is the natural culmination of some kind of marketing stunt.

The real story here is that this may be the beginning of governments restricting the availability of strong LLMs to the public, to you. Fable was the strongest model on the market, and the US government has told you you can’t use it (technically, only if you’re not a US citizen, but in practice, even if you are). If you think the solution here is going to be open source Chinese models and / or running on your own hardware, think again. Do you think China is going to allow the strongest LLMs from companies within its borders to be open source a year from now when they have Mythos capabilities, if the US government is keeping the strongest American models back? Unlikely. These are heading in the direction of being powerful cybersecurity weapons and it will be in the interest of nation states to restrict and control them. In 2 years time, I would be surprised if the strongest LLMs are available for general use at all.

Will we be the poorer for that, or will we be safer? I think poorer, because I hate being told what technology I can and can’t use, but I’m not certain. Maybe you think the government should restrict strong LLMs. Maybe you don’t. But either way, this is big news and a rubicon has been crossed and a precedent set. That’s true even if the motivation for this is just the government settling scores with Anthropic.

libraryofbabel

这么多评论都忽略了重点,只顾着幸灾乐祸地说Anthropic活该,或者说这是某种营销手段的必然结局。

真正的问题在于,这可能标志着政府开始限制强大语言模型对公众——也就是对你——的开放使用。Fable本是市场上最强的模型,而美国政府告诉你不能使用它(严格来说,仅限非美国公民,但实际上即便你是美国公民也一样)。如果你以为解决方案会是开源的中国模型和/或自建硬件运行,那请三思。如果美国政府将美国最强模型束之高阁,你觉得一年后当中国公司拥有神话级能力时,会允许其境内企业的最强语言模型保持开源吗?不太可能。这些技术正朝着强大网络安全武器的方向发展,限制和控制它们将符合各国的利益。两年后,如果最强的语言模型还能被广泛使用,我倒会感到意外。

我们会因此变得更糟,还是更安全?我认为是更糟,因为我讨厌被规定能用什么不能用什么,但我不确定。也许你认为政府应该限制强语言模型,也许不。但无论如何,这都是重大新闻——一条红线已被逾越,一个先例已经确立。即便政府此举只是为了跟Anthropic清算旧账,这一点依然成立。


Statement on US government directive to suspend ac… #

https://news.ycombinator.com/item?id=48511106

Finally they will pay for all the scaremongering they been doing to sell their models as something so much ahead of all else.

Now they finally found the right fools in audience to believe it.

SXX

最终,他们将为自己为了推销模型而散布的危言耸听付出代价,那些言论曾让他们的模型显得远超其他一切。如今,他们终于找到了愿意相信这一点的观众中的蠢货。


How to earn a billion dollars #

https://news.ycombinator.com/item?id=48527158

She meant impossible in that one doesn’t earn a billion dollars through work alone. The only way to get there is to set up a structure that extracts a billion dollars from a market (usually by building a structure that’s more efficient but also generates externalities that are not borne by the person getting the billion dollars).

pg’s reading of it is so blunt and misrepresentative that I’m nervous about what kind of content he’s consuming.

AdamN

她的意思是说,一个人不可能仅靠工作赚到十亿美元。要达成这个目标,唯一的途径是建立一个能从市场提取十亿美元的结构(通常是构建一个更高效的体系,但同时也产生外部性,而这些外部性并不由获得十亿美元的人承担)。

pg对此的理解如此生硬且带有曲解,让我对他到底在看什么内容感到担忧。


Statement on US government directive to suspend ac… #

https://news.ycombinator.com/item?id=48511330

When you spend a lot of time telling people how dangerous your products are, people who have the power to keep dangerous products off the market might listen.

Especially if those people aren’t presently very bright, and are already mad at you for not helping them achieve their unrelated authoritarian goals.

I do not think this is somehow a 3D chess move by Anthropic. They are not masterminds, even if they’d really like to be. People who actually interact with their products know that Fable and Mythos are incremental improvements, not doomsday devices. I think this is a punitive move by an administration that loves being punitive, which they have unknowingly bolstered with their own dumb rhetoric.

ivraatiems

当你花大量时间向人们宣传你的产品有多危险时,那些有能力阻止危险产品上市的人可能会听进去。

尤其是如果这些人目前并不怎么明智,而且已经因为你不帮他们实现那些不相干的威权目标而对你怀恨在心。

我不认为这是Anthropic下的一盘什么三维象棋。他们并非什么幕后主脑,即使他们很想成为那样的人。真正使用过他们产品的人都知道,Fable和Mythos只是渐进式的改进,而不是末日装置。我认为这是一个喜欢惩罚的政府采取的惩罚性举措,而他们自己那些愚蠢的言论无意中助长了这一切。


Not everyone is using AI for everything #

https://news.ycombinator.com/item?id=48528905

I assume it’s because he is seeking to pay rent, food bills, and other expenses through employment.

emodendroket

我猜这是因为他想通过工作来支付房租、食物账单和其他开支。


Noise infusion banned from statistical products pu… #

https://news.ycombinator.com/item?id=48518180

The replies here arguing we should publish it all are wild in the worst kind of first-order thinking way.

It’s a census: it just asks questions.

If you start publishing and weaponizing the data against people with various attributes, they’ll just lie or not answer. And then you are left with worse than nothing: bad data people try to act on.

asolove

这里的回复争论说我们应该全部公布,这是最糟糕的一阶思维方式。这是一次人口普查:它只是提问。如果你开始公布数据并将其武器化以针对具有不同特征的人,他们就会撒谎或不回答。那么你得到的比一无所有更糟:人们试图依据错误数据采取行动。


Statement on US government directive to suspend ac… #

https://news.ycombinator.com/item?id=48511233

It seems more likely that the logical conclusion is the executive branch is mad at Anthropic, and lashing out at them with any convenient tool that they have.

I suspect if OpenAI or Grok was operating at the same level they wouldn’t find themselves on the sharp end of the government stick

ncallaway

似乎更合理的推论是,行政部门对Anthropic感到不满,正利用手头任何顺手的工具对他们进行猛烈抨击。我怀疑如果OpenAI或Grok处于同样的运作水平,他们未必会遭政府如此严厉的打击。


The computer science degree isn’t dead #

https://news.ycombinator.com/item?id=48513216

If one is thinking about not getting a degree and trying to go straight to work, as someone who did so (albiet out of poverty rather than choice) but didn’t end up like Zuck, please heed my warning:

Social capital matters more than just about anyone who has a degree can understand and tell you or mentor you about, because the majority of them have always had it, and they tend not even to interact with people without it.

It is a signal about your wealth (and your families ability to deploy it for you), from which follows your stability, your intelligence, your taste, your willingness to play the game, and your belonging in the club. These matter more than EVER in the business world - I’ve never seen a time when tech is less about engineering than right now.

taurath

如果有人正在考虑不拿学位、直接工作,作为一个(出于贫困而非选择)这样做、但最终没有成为扎克伯格的人,请听我一句警告:

社会资本的重要性,远非那些拥有学位的人能够理解并告诉你或指导你的程度——因为他们大多数人一直拥有它,而且他们甚至倾向于不与没有它的人打交道。

它是关于你财富(以及你家族为你调动财富的能力)的信号,由此衍生出你的稳定性、智力、品味、你愿意参与这场游戏的意愿,以及你属于这个圈子的资格。这些在商业世界中比以往任何时候都更重要——我从未见过像现在这样,科技领域如此不关乎工程本身。


Statement on US government directive to suspend ac… #

https://news.ycombinator.com/item?id=48512237

Obviously their statements are insincere, because they are building the bloody things. If they were sincere that AI is like nuclear weapons, then they would be devoting all their cash and energy into lobbying the government to nationalize them and treat AI like nuclear weapons. They would not be attempting to IPO and they for sure would not sell their weapon-like thing to the general public.

tadfisher

显然他们的表态并不真诚,因为他们正在建造这些该死的东西。如果他们真心认为人工智能像核武器一样,就会把全部资金和精力用于游说政府将其国有化,并把人工智能当作核武器来对待。他们不会试图进行IPO,也肯定不会向大众出售这种类似武器的东西。


Statement on US government directive to suspend ac… #

https://news.ycombinator.com/item?id=48512669

This whole thing is comedy.

Anthropic pretending Mythos 5 is so capable it’s going to destroy everything, but will release it anyway with “safeguards” (when does this ever work?).

US Gov’t using this fake hype as an excuse to handicap Anthropic simply because they have a vendetta.

evilturnip

整件事就是个喜剧。

Anthropic假装Mythos 5强大到会摧毁一切,却还是要带着“安全保障”发布它(这招什么时候奏效过?)。

美国政府用这种虚假炒作作为借口来限制Anthropic,纯粹是因为他们之间有过节。


Statement on US government directive to suspend ac… #

https://news.ycombinator.com/item?id=48511334

Listen - that’s the sound of millions of companies and users doubling down on Chinese models.

It might be a national security problem for other nations to have access to these models. But it’s equally now a national security problem for any other nation to depend on them. Or US tech in general.

zmmmmm

听——那是数百万公司和用户加倍押注中国模型的声音。其他国家获取这些模型可能是一个国家安全问题。但如今,任何其他国家依赖它们也同样成为国家安全问题。或者依赖美国科技整体而言也是如此。


New pancreatic cancer drug might open the door to … #

https://news.ycombinator.com/item?id=48518443

As is often the case, the title is hyperbolic. The discovery applies to 20% of tumors, and “one of cancer’s significant defenses” or “a key weakness of cancer” would be more accurate.

That said, I’ll happily take “we discovered a key weakness in 20% of cancers,” please and thank you.

gcanyon

标题往往夸大其词。这项发现适用于20%的肿瘤,而“癌症的重要防御机制之一”或“癌症的一个关键弱点”会更准确。话虽如此,我很乐意接受“我们在20%的癌症中发现了一个关键弱点”,谢谢。


A low-carbon computing platform from your retired … #

https://news.ycombinator.com/item?id=48515783

This is ignoring the fact that the main reason retired phones are e-waste is proprietary firmware blobs and locked-down systems preventing users from maintaining their phone with security updates, and very limited support length from OEM’s leads to VERY insecure devices after they drop out of support.

You should not be connecting these old devices to an internet accessible network.

Google notably does well here with 7 years of support, but others such as Sony are 4 years, and Xiaomi on non-flagship devices are similar, or Samsung on their lowest budget models…

zipy124

这忽略了事实:旧手机沦为电子垃圾的主要原因在于专有固件块和封闭系统使用户无法通过安全更新维护手机,而OEM厂商提供的支持周期极短,导致设备在停止支持后变得极其不安全。

你不应该将这些旧设备连接到可访问互联网的网络。

谷歌在这方面做得很好,提供了7年支持,但其他厂商如索尼只有4年,小米的非旗舰设备也类似,还有三星的最低端机型也是如此……


Amazon CEO’s talks with U.S. officials triggered c… #

https://news.ycombinator.com/item?id=48519887

I still am struggling to understand why they informed the government about something that is known to be an issue in every LLM. There is no LLM that cannot be jailbroken, so unless this means that we have reached the absolute maximum publicly accessible US made LLMs are allowed to operate at with GPT 5.5, this is not grounded in any sane regulation attempt.

Does anyone know what limits Fable 5 has overstepped in the eyes of the government? Parameter count? Certain benchmark results? Training computer?

Cause if it’s just the ability to assist with cyberattacks and being jailbreakable, there is no model previously released that isn’t equally guilty.

Remember that for GPT 5.5 and 5.4, OpenAI also restricted the cybersecurity focused use under designated models, otherwise rerouting to 5.3-codex like Fable did with Opus 4.8. And both OpenAI models can also be jailbroken all the same.

Basically, what was the reason to tell the government now and not with Opus 4.5 or GPT 5.4? sama has been doing the rounds with apocalyptic predictions…

Topfi

我至今仍无法理解,为什么他们要向政府报告一个每个大语言模型都已知存在的问题。没有任何LLM是无法被越狱的,所以除非这意味着我们已经达到了美国可公开访问的LLM在GPT 5.5上被允许运行的绝对上限,否则这根本不符合任何合理的监管尝试。

有人知道Fable 5在政府眼中越过了哪些限制吗?参数数量?某些基准测试结果?训练算力?

因为如果仅仅是因为它能够协助网络攻击且可被越狱,那么之前发布的所有模型都同样有罪。

请记住,对于GPT 5.5和5.4,OpenAI也限制了指定模型下的网络安全相关用途,否则就会像Fable对Opus 4.8的做法一样,重定向到5.3-codex。而且这两个OpenAI模型同样可以被越狱。

基本上,为什么现在才告知政府,而不是在Opus 4.5或GPT 5.4的时候?萨姆(Sam Altman)一直在兜售那些末日预言……


Statement on US government directive to suspend ac… #

https://news.ycombinator.com/item?id=48511209

So isn’t the only logical conclusion that we have reached the max of model capabilities that the US allows to be made available to the public? Why invest in smarter models with this precedent?

And potentially more importantly: if a model like Mythos, which at best is an incremental improvement over Opus, is getting this treatment, how are all the AI investments that are based on the expectation of ASI / AGI / significantly better models going to be recouped?

stingraycharles

那么唯一合乎逻辑的结论不就是,我们已经达到了美国允许向公众提供的模型能力的上限吗?有了这个先例,为什么还要投资更智能的模型?

更重要的是:如果一个像Mythos这样最多比Opus略有改进的模型都受到这种对待,那么所有基于对ASI/AGI/显著更优模型预期的AI投资,又该如何收回成本?


Honda Civics and the Evil Valet #

https://news.ycombinator.com/item?id=48523081

To update 10th-gen Honda Civics, Honda ships updates on specially-formatted USB drives. They’re essentially Android 4.2.2rc1-era recovery packages with some Honda-added version checks (which can be spoofed). The packages are signed with the publicly-known AOSP test key, so with physical access to the front USB port you can sign and flash your own package for arbitrary code execution on the headunit. This doesn’t require root/su. I’ve run it end-to-end on my own 2021 Civic and separately confirmed an official EU update file carries the AOSP test-key signature. Tooling and writeup in the post.

librick

为了更新第十代本田思域,本田通过特殊格式的USB闪存盘推送更新。这些更新本质上是基于Android 4.2.2rc1时代的恢复包,并附加了本田自加的版本校验(可被绕过)。这些包使用公开的AOSP测试密钥签名,因此只要有物理途径接触前置USB接口,就能签名并刷入自定义包,在车机上执行任意代码。这不需要root权限。我已在自家2021款思域上完整跑通过,并单独确认一份官方欧盟更新文件确实带有AOSP测试密钥签名。相关工具和说明见帖子。


Statement on US government directive to suspend ac… #

https://news.ycombinator.com/item?id=48515438

The way I see it, a government led by an adult toddler and his sycophants has decided to punish a firm that refused to cooperate with it’s military when it was embarrassed by a militarily weak adversary. The model strength spin strikes me as motivated reasoning.

The rubicon being crossed here is Republicans/the red tribe losing their comparative advantage of being opposed to overregulating a rapidly advancing technology.

vovavili

在我看来,一个由巨婴般的成年人和他的谄媚者领导的政府,在因一个军事上弱小的对手而感到难堪时,决定惩罚一家拒绝与其军方合作的企业。所谓"模型实力"的舆论引导在我看来不过是动机性推理。

真正跨越的卢比孔河是:共和党人/红州群体失去了他们在反对过度监管快速发展的技术方面的相对优势。


Statement on US government directive to suspend ac… #

https://news.ycombinator.com/item?id=48512864

I think we should see this as simply silly behavior by a government.

Export control is not an effective tool for controlling a consumer facing technology developers everywhere want to use (see:VPNs) so there was no good faith policy justification for imposing an export control.

This is an administration that seems to be keeping track of who its friends are and aren’t, and likes to be the center of every story. They also seem to like extracting concessions and reciprocal favors. We saw some of this behavior in the last administration too. US voters deserve better.

holmesworcester

我认为我们应该将这种行为视为政府单纯的愚蠢举动。

出口管制并非控制面向消费者的技术开发者的有效工具——这些技术全球开发者都想使用(例如:VPN),因此实施出口管制缺乏善意的政策依据。

这届政府似乎热衷于区分谁是其盟友谁不是,并且喜欢成为每件事的中心。他们还喜欢索取让步和互惠好处。我们在上一届政府中也见过类似行为。美国选民理应得到更好的对待。


Noise infusion banned from statistical products pu… #

https://news.ycombinator.com/item?id=48519051

I “enumerated” for the last census. Trust in my community was already not high* and I had lots of interesting encounters. I really believed the rather invasive data I was collecting with a friendly face would be used and handled responsibly. I feel for the poor souls that’ll sign up to go door to door for 2030 now that the firewalls against weaponizing and monetizing all of our sensitive government data has been torn down, and even more for those that will volunteer information that can hurt them.

The comments that this rather expensive endeavour should just be about getting a head count are also amusing to me. The data collected was such an important baseline of common understanding, and this will not be a good thing for its future quality. I’ve grown very jaded now seeing all the things taken for granted in this country and lost or degraded recently with a whimper.

*: To be fair, they sent me specifically to places that didn’t respond, so I was naturally led to believe that everyone in my region hated the government, ignored bizzarrely threatening fliers, or had recently moved and had no knowledge of the inhabitants (if any) during the census period.

kajman

我在上次人口普查中担任过“枚举员”。社区对我的信任本就不高*,我还经历了不少有趣的遭遇。我曾真心相信,自己面带友善收集的这些相当侵入性的数据会被负责任地使用和处理。如今,针对将我们所有敏感政府数据武器化和货币化的防火墙已被拆除,我同情那些将为2030年人口普查挨家挨户登记的可怜人,更同情那些会主动提供可能伤害自身信息的人。
那些认为这种代价高昂的工作只该统计人口数量的评论也让我觉得可笑。所收集的数据曾是达成共识的重要基线,而这对未来数据质量绝非好事。如今看到这个国家习以为常的事物在近期悄无声息地流失或退化,我已变得非常悲观。
*: 公平地说,他们专门派我去那些没有回应的地区,所以我自然认为我所在区域的人都憎恨政府、无视那些古怪的威胁性传单、或是最近刚搬来且对普查期间(如有)居民一无所知。


Israeli firm BlackCore suspected of meddling in Ne… #

https://news.ycombinator.com/item?id=48515304

Last time I suggested on a similar story that there’s a disproportionate number of firms in Israel with an explicit focus on subversion, manipulation, spying and malware, seemingly because a large portion of the Israeli population gain a certain expertise in these fields as part of serving in the IDF and working to suppress Palestinians, I got accused of bias because apparently there’s many more Israeli startups working on medical research, green technology and world peace.

If there are, they certainly would do no harm in being more vocal, firms like BlackCore is unfortunately what Israel is becoming known for around the world.

Matl

上次我在一篇类似的报道中提到,以色列有数量不成比例的公司明确专注于颠覆、操控、间谍活动和恶意软件,似乎是因为很大一部分以色列人在以色列国防军服役、参与压迫巴勒斯坦人的过程中获得了这些领域的特定专长——结果我被指责存在偏见,因为显然还有更多以色列初创公司在从事医学研究、绿色技术和世界和平事业。
如果真有这些公司,它们当然应该更积极地发声,但不幸的是,像BlackCore这样的公司才是以色列在世界范围内越来越为人所知的形象。


Amazon CEO’s talks with U.S. officials triggered c… #

https://news.ycombinator.com/item?id=48519623

Researchers at Amazon had used a series of prompts to get Anthropic’s Fable 5 model to provide them with information that could be used to aid cyberattacks…

Are there going to be bans on things that could be used to aid in school shootings next?

blitzar

亚马逊的研究人员使用一系列提示,让Anthropic的Fable 5模型提供可用于辅助网络攻击的信息……接下来是否要禁止可能被用于辅助校园枪击的事物?


New pancreatic cancer drug might open the door to … #

https://news.ycombinator.com/item?id=48518613

Cancer is not one thing, it’s a huge zoo of many many many ways that cells start to break the social contract and divide in an uncontrolled manner.

One of the most commonly observed broken mechanisms is mutation in the gene KRAS that turns this on/off growth switch into the permanently on position.

This has been known for decades, of course. And there have been huge amounts of effort to try to develop drugs that target KRAS in cancer, but for decades it’s always been thought of as ‘undruggable’ because of the difficulty of finding any molecules that would affect it.

This new drug, that finally treats KRAS mutated cancers, goes about it in a new way. Instead of trying to gum up the works of a single protein by sticking a small chemical in it, it effectively “glues” the KRAS protein to another protein, CypA, which keeps the switch away from reaching the normal areas where it’s “on switch” activity works.

So this new drug means two things: 1) a lot of the most difficult to treat cancers are now far more treatable, and in the next 1-5 years clinical trials will tell us which cancers this particular drug works well for, 2) there’s an entire new class of drug activity that everybody is chasing at this very moment, so in 5-25 years we’ll likely have a huge number more of these sorts of treatments.

epistasis

癌症并非单一疾病,而是一个庞大的“动物园”,包含细胞以无数种方式打破社会契约、不受控制地分裂增殖的机制。其中最常被观测到的机制之一是KRAS基因突变,它使控制生长的开关被永久固定在“开启”状态。
当然,这一点几十年前就已为人所知。科学家们投入了大量精力试图开发针对KRAS的靶向抗癌药,但几十年来它始终被视为“不可成药”靶点,因为很难找到能影响它的分子。
这款终于能治疗KRAS突变癌症的新药,采用了一种全新策略。它并非通过向单个蛋白质中嵌入小分子来阻塞其功能,而是将KRAS蛋白质“粘合”到另一种名为CypA的蛋白质上,从而阻止该开关接近其发挥“开启”活性的正常区域。
因此,这款新药意味着两件事:1)许多最难治疗的癌症如今变得更容易治疗,未来1-5年的临床试验将揭示该药具体对哪些癌症有效;2)这是一类全新的药物作用机制,眼下所有人都在竞相研究,未来5-25年我们很可能会迎来大量此类疗法。


There is a shadow hanging over this Fable thing #

https://news.ycombinator.com/item?id=48514085

But this government […]

I’m hearing a lot of this kind of thing. “Oh if only it was a different government”. I’m sorry, but when you cry out for government involvement, it’s not always going to be coming from the government you personally wanted. This is the whole problem with government involvement! I don’t think that message is getting through, but it’s the real lesson that should be learned here.

modeless

但这个政府[…]

我最近听到很多这种话。“唉,要是换一个政府就好了。”抱歉,但当你呼吁政府介入时,介入的未必总是你个人想要的那个政府。这正是政府介入的整个问题所在!我觉得这个信息并没有被理解,但这才是应该从中吸取的真正教训。


Every Frame Perfect #

https://news.ycombinator.com/item?id=48519277

I agree that some of the examples the author provided are instances of bad animation. But I don’t agree with the premise of the article.

Computer graphics is all about exploiting features of the human visual system. We perceive things differently when they’re moving vs. when they’re standing still. It’s very possible that a “wrong” frame in isolation is the best looking one in a real-time context. We can also pick apart screenshots but these don’t capture everything about how the user perceives a display in real-world lighting conditions.

I would draw an analogy to film. A fast tracking shot might look bad on individual frames because of motion blur. A wide-angle shot might make some objects look “wrong” because of optical distortion. But these are still the right choice if they have the intended artistic effect in the theater.

fasterik

我同意作者举的一些例子确实是动画质量不佳的表现,但我不同意这篇文章的基本论点。

计算机图形学的核心在于利用人类视觉系统的特性。物体在运动时和静止时,我们的感知是不同的。在实时渲染的上下文中,孤立看起来“错误”的一帧完全有可能是视觉效果最好的那一帧。我们也可以逐帧分析截图,但这些截图并不能完全反映用户在真实光照条件下对显示效果的感知。

我打个比方:电影中的快速跟拍镜头,单看每一帧可能会因为运动模糊而显得糟糕;广角镜头可能因光学畸变让某些物体看起来“不对劲”。但只要这些手法在影院中达到了预期的艺术效果,它们依然是正确的选择。


GLM 5.2 Is Out #

https://news.ycombinator.com/item?id=48519293

The real news here is that Digg is still up :O

radious

真正的新闻是Digg竟然还在运行 :O


GLM 5.2 Is Out #

https://news.ycombinator.com/item?id=48521149

Announcement from the founder of Z.ai:

“ GLM-5.2 is Fully Open, Frontier Intelligence Belongs to Everyone

Today, the sudden restriction of certain frontier models is deeply regrettable. At a time when access to frontier models is abruptly cut off for non-technical reasons, we are even more convinced of one thing: science should be global.

The path to AGI (Artificial General Intelligence) must never be enclosed by high walls. We have always believed that AGI should be the cornerstone for all of humanity to collaboratively explore the boundaries of intelligence and solve complex challenges, rather than a privilege monopolized by a few rules and subject to revocation at any moment. In the face of external blockades and restrictions, our attitude is one of radical openness. Frontier intelligence must remain open-source, accessible, and buildable, serving every dedicated developer.

GLM-5.2 is Zhipu’s most capable open-source model to date. It not only supports a truly usable 1M context window but also maintains a continuous lead in the independent completion of long-horizon tasks, providing solid foundational support for building complex agent applications. It also continues to be our main engine for creating the strongest domestic coding model.

Tonight at 5:21—at this special moment—GLM-5.2 will officially be available to all GLM Coding Plan users (including Lite / Pro / Max). The API will also go live next week.

A step closer to frontier intelligence for everyone. The future of AI is open, and it is for the people. ModelKey: GLM-5.2”

https://x.com/jietang/status/2065784751345287314

easygenes

Z.ai 创始人的声明:

“GLM-5.2 全面开放,前沿智能属于每一个人

今天,某些前沿模型突然受到限制,令人深感遗憾。当由于非技术原因导致前沿模型访问被突然切断时,我们更加坚信一件事:科学应该是全球性的。

通往通用人工智能的道路绝不能高墙筑垒。我们始终认为,通用人工智能应成为全人类共同探索智能边界、解决复杂挑战的基石,而非由少数规则垄断、随时可被撤回的特权。面对外部的封锁与限制,我们的态度是彻底开放。前沿智能必须保持开源、可访问、可构建,服务于每一位专注的开发者。

GLM-5.2 是智谱迄今为止能力最强的开源模型。它不仅支持真正可用的百万上下文窗口,还在长时任务的自主完成方面持续保持领先优势,为构建复杂智能体应用提供坚实支撑,同时它也是我们打造最强国产编程模型的核心引擎。

今晚 5:21——在这个特殊的时刻——GLM-5.2 将正式对所有 GLM 编程计划用户(包括 Lite / Pro / Max)开放。API 也将在下周上线。

每个人都离前沿智能更近一步。AI 的未来是开放的,它属于人民。模型密钥:GLM-5.2”

https://x.com/jietang/status/2065784751345287314


https://news.ycombinator.com/item?id=48510263

Palantir is clearly a mind-boggling on-the-nose, but terrible name to those familiar with the book.

The Palantiri consistently provided their users technically accurate intelligence that lead to disastrous strategic decisions.

Denethor committed suicide out of despair, after a palantir showed him the black fleet approaching, but he did not know that it was actually Aragorn who had captured the fleet and was coming with reinforcements.

We don’t know specifically how the palantir deceived Saruman, but it’s pretty clear it was one of the key factors in his corruption and downfall.

And even Sauron himself was misled in this way! The palantir showed him, correctly, that a hobbit and Aragorn were at Helm’s Deep, and he concluded that Aragorn had the ring. So he prematurely moved his armies out of Mordor and left the plains and Mt Doom unguarded, which permitted the destruction of the ring.

I honestly can’t think of a worse name for a company that provides intel for strategic decision making.

timoth3y

Palantir这个名字显然极其直白又令人费解,对于熟悉那本书的人来说是个糟糕的名字。

真知晶球始终向使用者提供技术上的准确情报,却导致了灾难性的战略决策。

德内豪在真知晶球中看到黑色舰队逼近后绝望自尽,但他并不知道那其实是阿拉贡俘虏了舰队,正带着援军赶来。

我们不清楚真知晶球具体如何欺骗萨鲁曼,但很明显这是导致他堕落与覆灭的关键因素之一。

就连索伦本人也被这种方式误导了!真知晶球正确地向他显示霍比特人和阿拉贡出现在海尔姆深谷,于是他推断阿拉贡持有魔戒。因此他过早调离了魔多的军队,导致平原和末日火山无人看守,最终使魔戒被摧毁。

老实说,我想不出比这更糟糕的名字来命名一家为战略决策提供情报的公司了。


AI OSS tool repo goes archived over night after ra… #

https://news.ycombinator.com/item?id=48518120

I’m the co-founder and CEO of TensorZero.

We started the company two and a half years ago, and raised $7.3m in 2024 (announced only almost a year later). We’ve spent less than half of this amount.

Earlier this week we came to the difficult decision to wind down the project. The open-source repository remains available on GitHub (Apache 2.0) but won’t be actively maintained by the team moving forward.

GabrielBianconi

我是TensorZero的联合创始人兼CEO。我们两年半前创立了公司,2024年筹集了730万美元(几乎一年后才对外公布)。我们只花掉了其中不到一半的资金。本周早些时候,我们艰难地决定停止这个项目。开源仓库仍保留在GitHub上(Apache 2.0许可),但团队未来将不再积极维护。


Leaving Mozilla #

https://news.ycombinator.com/item?id=48514814

Some 10 years ago I was a Mozilla volunteer. I mainly worked on MDN, to the point of becoming a so-called “topic driver” for the glossary. Some of the work I did landed in the citations of a couple of papers about web technology. They flew me a whole week to Vancouver for an event where employees and volunteers worked together in the same room and they even made me (and the other volunteers ) attend a sort-of-corporate meeting where they sort-of fought about something (can’t even remember what it was).

I’m telling you this to highlight that volunteers where a huge part of Mozilla.

But on the last day they announced that they were moving the day-to-day conversations from IRC (an open protocol) to Yahoo Messenger (a closed protocol). I felt sort of betrayed in that moment: the company that was all about openness and to which I dedicated countless hours doing unpaid work for and even more years evangelizing for was imposing its volunteers and employees used a proprietary app to coordinate. That didn’t sit well with me. At all. I basically lost interest.

This was in 2015. Last I heard MDN introduced ads (I wouldn’t know, uBlock is pretty effective) and is not showing contributors to a page on the page itself anymore.

So yeah, the part of OP saying how Mozilla managed to piss volunteers resonated pretty hard with me.

klez

大约十年前,我曾是Mozilla的一名志愿者。我主要参与MDN的工作,甚至成为了所谓的"词汇表话题负责人"。我的一些工作成果被几篇网络技术论文引用过。他们曾把我整周送到温哥华参加一个活动,让员工和志愿者在同一个房间里协作,甚至还让我(和其他志愿者)参加了一场近似公司会议的活动,会上他们似乎在争论什么(我甚至记不清具体内容了)。

我告诉你这些是为了强调志愿者曾是Mozilla的重要组成部分。

但在活动的最后一天,他们宣布要把日常交流从IRC(开放协议)迁移到Yahoo Messenger(封闭协议)。那一刻我感到某种背叛:这个标榜开放的公司,这个我投入无数小时无偿工作、甚至多年为其宣传的公司,竟然强制其志愿者和员工使用专有软件来协调工作。这让我非常不满。极其不满。我基本上失去了兴趣。

那是在2015年。我最后一次听说MDN引入了广告(我不确定,uBlock很有效),而且页面上不再显示贡献者信息了。

所以,原帖说Mozilla如何惹恼志愿者的那段话,确实让我深有共鸣。


The Birth and Death of JavaScript (2014) #

https://news.ycombinator.com/item?id=48526774

I love(?) that he absolutely predicted a global disaster between 2020-2025, he just got the wrong type. Which is very JavaScript.

DavidPiper

我挺喜欢(?)他准确预言了2020-2025年间会有全球性灾难,只是搞错了灾难的类型。这很JavaScript。