2026-06-07 Hacker News Top Stories #
- 标普道琼斯指数公司因盈利等规则拒绝SpaceX快速纳入标普500,也令未盈利的AI公司难以进入指数。
- 一名GrapheneOS用户因使用该系统被Yoti上报当局,引发对应用检测和隐私的讨论。
- 谷歌与SpaceX签订合同,每月支付9.2亿美元租用约11万块英伟达GPU,用于AI产品。
- 红杉和a16z因不信任女性领导或团队准备不足而拒绝投资,还有GP在演示中睡着。
- 宝可梦绿宝石通过WebAssembly移植到浏览器,支持速度切换和键盘操作。
- 作者维修了一支不通电的Sigma镜头,通过更换烧断的保险丝修复。
- Linux社区讨论用pidfd构建新进程创建API以取代低效的fork()+exec()。
- Nvidia提出面向Windows PC的高性能CPU系统,采用统一内存设计,融合CPU与GPU。
- Bright Data通过智能电视和手机应用SDK将用户设备变成AI抓取节点。
- zeroserve是一个零配置HTTPS服务器,用eBPF编写请求处理逻辑并实现高性能。
1. 标普 500 指数拒绝 SpaceX,同时也阻止 OpenAI 和 Anthropic 进入 (S&P 500 rejects SpaceX, also blocking entry for OpenAI and Anthropic) #
SpaceX 要求标普 500 指数为其首次公开募股(IPO)破例,允许快速纳入该指数,但遭到拒绝。标普道琼斯指数公司于 6 月 4 日决定,不修改资格标准,包括财务可行性、新股上市等待期(12 个月)及最低公众持股比例要求。这意味着 SpaceX 无法立即获得被动基金约 140 亿美元的买入资金,同时 OpenAI 和 Anthropic 等 AI 公司未来也面临类似障碍。
SpaceX 计划仅发行约 3% 的股份给公众,目前尚未盈利,债务高达 290 亿美元。即使等待一年后,这些公司也可能难以满足持续盈利等入指条件。该决定被视为对依赖被动投资和退休储蓄基金的市场风险的一种缓解。
HN 热度 1344 points | 评论 465 comments | 作者:maltalex | 20 hours ago #
https://news.ycombinator.com/item?id=48421442
- 被动投资者希望指数保持现有规则,不应为 SpaceX 等个别公司破例。
- 指数需要严格筛选,只纳入经过长期考验的稳定公司,而非未盈利或波动大的新公司。
- 如果 Alphabet 转型为全 AI 公司,是否也应被移除?质疑标准的一致性。
- 公司进入前需经过市场充分定价和财报验证,避免强制购买银行虚估的股票。
- 关键要看流通市值而非总市值,SpaceX 流通市值较低,暂无必要加速纳入。
- 被动指数基金不应为追求回报改变规则,否则会失去风险管理和市场参考作用。
- 投资者可通过全市场指数(如 VTI)获得更广敞口,而非仅依赖 S&P 500。
- 短期其他指数可能因纳入热门股而跑赢,但长期影响需更多数据验证。
2. GrapheneOS 用户因使用 GrapheneOS 而被上报给当局 (GrapheneOS user reported to authorities for using GrapheneOS) #
https://discuss.grapheneos.org/d/36134-grapheneos-user-reported-to-authorities-for-using-grapheneos
用户报告称,使用 Yoti 年龄验证服务时,系统会自动标记运行 GrapheneOS 的设备,并将用户信息上报给当局和安全团队。论坛成员讨论认为,GrapheneOS 因提供更强的安全与隐私保护(如安全执行派生、硬件认证 API 等)可被应用检测到,Yoti 可能对这类系统抱有敌意。有人建议将 GrapheneOS 仅用于日常隐私,另用廉价手机专门处理强制验证;也有观点称这是客服恐吓,并非公司正式政策。GrapheneOS 官方回应指出,应用检测系统是普遍现象,并非 GrapheneOS 特有的问题,隐藏系统特征不现实。
HN 热度 436 points | 评论 445 comments | 作者:Cider9986 | 15 hours ago #
https://news.ycombinator.com/item?id=48422798
- 原帖作者曾多次尝试绕过年龄验证,因此被禁是自找的,公司不一定有错。
- 公司代表明确表示会举报所有使用 Graphene OS 的用户,而不问其他情况,这一点非常过分。
- 允许用户隐藏历史记录是一个糟糕的决定。
- 总是有人挖别人的旧账去向版主告状。
- 讽刺英国对操作系统的管制:用户没被警察上门算幸运。
- 英国警察可以随意索要密码,不给就坐牢;相比之下美国公民权利好得多。
- 美国边境官员可以无理由要求查看手机和密码,拒绝即被拒绝入境,且数据可能长期留存。
- 美国 CBP 会追踪网上言论并拒绝入境,自由言论比旅游更重要,因此不愿再访美。
- 不去美国就没事,但有时因工作必须去,可以准备一部二手手机应付边境检查。
- 被美国拒签的比例极低(0.000001%),不应因小概率事件放弃生活。
- 实际上很多澳洲朋友都有不愉快的美国边境经历,拒签和被捕的比例并不低。
- 英国也有因言论被捕的案例,美国的外国人不享有美国公民的言论自由保护。
3. 谷歌每月向 SpaceX 支付 9.2 亿美元用于计算。 (Google will pay SpaceX $920M per month for compute) #
https://techcrunch.com/2026/06/05/google-will-pay-spacex-920m-per-month-for-compute/
谷歌将每月向 SpaceX 支付 9.2 亿美元,以租用约 11 万块英伟达 GPU 等计算资源,合同期从 2026 年 10 月持续到 2029 年 6 月。这笔交易旨在满足谷歌 AI 产品(如 Gemini Enterprise)激增的需求。此前,Anthropic 也与 SpaceX 达成了类似但规模更大的计算租用协议。SpaceX 即将在纳斯达克上市,估值约 1.75 万亿美元,而谷歌作为其长期投资者,IPO 后持股价值预计超 1000 亿美元。双方还在洽谈建设轨道数据中心。
HN 热度 405 points | 评论 2 comments | 作者:ramanan | 12 hours ago #
https://news.ycombinator.com/item?id=48423990
- 许多网友认为,xAI 的收入几乎完全来自于租用数据中心,这让 SpaceX 的估值显得极高。
- 对于 SpaceX 的估值,许多人表示担忧,认为市场可能会在未来出现调整。
- 投资者在考虑退休投资时,传统的 “逐步转向债券” 的建议已经不再适用,许多人倾向于 100% 投资股票 - 尽管有些人认为债券在当前市场中风险较大,但仍有网友认为持有债券能提供更高的确定性。
- 有人指出,许多退休建议更多是针对中产阶级,而那些财富充足的人可能会选择较低波动的投资组合。
- 讨论中提到的 VFORX 基金被认为费用较低,但有网友对主动管理的投资持批评态度。
- 关于是否支付股息的问题,网友们认为股息并非股票价值的唯一基础,真正的价值在于公司未来现金流的预期。
- 许多人认为,股票市场的价格并非完全随机,而是基于对公司未来潜力的预期。
- 最后,网友们对股票市场的未来发展表示担忧,尤其是在当前经济不确定性加大的背景下。
4. 我们最糟糕的三个 VC 故事 (Three of our worst VC stories) #
https://twitter.com/eastdakota/status/2062860530360959273
Matthew Prince(Cloudflare CEO)分享了他经历过的最糟糕的两个 VC 故事:
- 红杉资本的一位合伙人因为不相信女性能够领导一家安全基础设施公司,而拒绝了 Cloudflare 的投资。
- 他被介绍给 Marc Andreessen,会议定在周一。他以为是普通会面,没做准备,但对方以为是正式路演,带来了整个 a16z 团队。会上有人说“你看起来没怎么准备”,他后来把对方的拒绝信裱了起来。
他还引用了 GREG ISENBERG 的一个类似经历:在一次 A 轮融资路演中,一位 GP 当场睡着了 30 多分钟,但无人理会,他只能继续对着一个昏迷的人讲完整个演示。
HN 热度 261 points | 评论 140 comments | 作者:orgonon | 1 day ago #
https://news.ycombinator.com/item?id=48416845
- 小团队加上 Claude Code 订阅和自动化代码,完全可以通过利润再投资来发展,避免外部投资,还能去掉人力协调成本。
- 买几块 GPU 自托管 AI 模型,能避免订阅费、速率限制、隐私和拒绝服务问题,对能用旧 PC 组装的小团队来说几个月就能回本。
- 自托管模型(如 Qwen3.6 27B)在编程调试和研究上效果不错,甚至比某些大型第三方模型更优,而且是免费的。
- 有 VC 让人从佛罗里达飞到德州吃早餐,结果见面就问“有多少付费客户”,听到没有就打发走,连完整的演示都不让做。
- 这种见面完全可以通过邮件沟通,浪费双方时间,VC 就是故意摆权力。
- 另一个经历是遇到假富商“Michael Prozer”,他靠伪造银行记录上过电视真人秀,最终因欺诈被捕。
- VC 要求看到实际付费客户再投资其实是合理的,但问题在于他本应提前说清楚,而不是让人飞过去当面拒绝。
- 愿意为 VC 飞半个国家反而可能传递负面信号:说明这家公司没有其他 VC 感兴趣,显得很绝望。
- 一开始创业者都会犯这种错,后来学会了保护自己的时间,保持多元收入来源。
- 大部分 VC 故事都是恐怖故事,但也有很多安静、专业的 VC,只是他们不张扬。
- 有一个正面例子:创始人用不熟悉初创公司的律师做股票计划,后来公司以近 10 亿美元被收购,收购过程中才发现股票计划有错误……(故事被截断,但说明 VC/法律问题最终能被解决)。
5. 宝可梦绿宝石移植到 WebAssembly(10 万 FPS) (Pokemon Emerald Ported to WebAssembly (100k FPS)) #
这是一个用 WebAssembly 技术移植到浏览器的宝可梦绿宝石(Pokémon Emerald)游戏模拟器。页面包含游戏画面、方向键以及 B、A、SELECT、START 等操作按钮,支持速度切换(默认 1 倍速)。键盘快捷键:方向键控制移动,Z 键对应 A 按钮,X 键对应 B 按钮,Enter 键对应 START,Shift 键对应 SELECT。
HN 热度 259 points | 评论 71 comments | 作者:tripplyons | 13 hours ago #
https://news.ycombinator.com/item?id=48423762
- 喜欢加速功能,但希望指示键盘 z/x 对应 A/B 键,否则默认键位不够直观。
- 在“Fight/Bag/Pokemon/Run”菜单中点击“Pokemon”会导致崩溃。
- 建议允许用户重新映射按键,比如将 Start 从 Enter 改为空格,以及用 WASD 和 H/J 键代替方向键和 A/B 键。
- 确认存档功能正常工作,但交易功能似乎无法使用。
- 出现文本 bug:某些物品名称显示为数字(如“You received a 6”),需要排查。
- 推荐另一个 WASM 移植项目:Xonotic(FPS 游戏),包含技术文章。
- 建议提供两种速度模式(正常/加速),并用一个按键切换。
- 批评当前版权期限过长(超过 20-30 年),认为应缩短以释放文化作品,让大众能自由使用和改编经典游戏。
- 担忧完整反编译项目(如 pokeemerald)可能面临法律风险,尽管需要用户提供 ROM,但内嵌了原始动画资源,存在版权隐患。
- 项目通过将资产存储在专有格式、需要 ROM 作为“钥匙”才能编译来规避版权,但这仍是偷懒之举,应编写提取工具。
6. 现代相机镜头维修的复杂性(2024) (The intracies of modern camera lens repair (2024)) #
https://salvagedcircuitry.com/sigma-45mm.html
这位作者修复了一支 Sigma 45mm f/2.8 镜头。镜头外观完美但电子功能完全失效,镜头卡口也很紧。拆解后检查了柔性排线和控制 PCB,最终发现电路板上的 DC-DC 转换器输入保险丝熔断,导致镜头无法通电。文章详细列出了所需工具、拆解步骤以及电路分析方法,适合有维修经验的读者参考。
HN 热度 241 points | 评论 84 comments | 作者:transistor-man | 24 hours ago #
https://news.ycombinator.com/item?id=48420148
- 保险丝的存在是为了防止火灾,而不是保护元件,半导体比保险丝快得多
- 在精心设计的微电子电路中,保险丝配合齐纳二极管和电阻可以保护电路,但消费级产品通常不这么设计
- 保险丝只能应对输入过压,无法处理负载短路,需要主动限流电路或设计电路能承受短路
- 保险丝是过载保护,真正保护设备需要限流电路或继电器等主动保护
- 保险丝保护元件的观点适用于汽车或工厂级别,但不适用于半导体级别
- 汽车保险丝同样是防止火灾而非保护设备
- 电机保护中保险丝用于过流防止火灾,继电器和接地故障保护用于保护设备
- 现代镜头(如 Tamron)有 USB-C 接口用于固件更新和功能自定义,是巨大进步
- 近十年趋势是高端手动对焦镜头回归,尤其中国厂商(如 Samyang/Rokinon)推出大量优秀手动镜头
- 电影镜头仍多为纯机械手动对焦,自动对焦在视频中不好用,而拍照时自动对焦很实用
7. 超越 fork() + exec() (Moving beyond fork() + exec()) #
https://lwn.net/SubscriberLink/1076018/16f01bbbb8e0d1f0/
Linux 内核长期依赖 fork() + exec() 创建进程,但 fork() 需要复制整个进程状态,开销大且常被 exec() 立即抛弃。Li Chen 提出的“spawn templates”补丁通过缓存可执行文件信息来加速重复调用,benchmark 显示约 2% 提升。
然而社区反馈认为该方案未解决 fork() 本身的高成本。Mateusz Guzik 建议直接创建“干净进程”替代 fork。Christian Brauner 提议基于 pidfd 构建新 API:先用 pidfd_open() 创建空进程,再用 pidfd_config() 配置环境、镜像等,最终在用户空间实现 posix_spawn(),避免隐藏 fork/exec。
Li Chen 同意该方向,未来工作将转向此方案。虽然当前补丁不会被接受,但可能催生更优的进程创建原语。
HN 热度 227 points | 评论 248 comments | 作者:jwilk | 10 hours ago #
https://news.ycombinator.com/item?id=48425528
- fork 是 1970 年代的技巧,已经过时,应该被废弃。
- fork 最初是为了在内存有限的情况下运行大程序,通过交换到磁盘实现。
- QNX 将程序加载放在用户空间库中,这可能是正确的方式。
- 现在内存又贵了,可能回到类似情况。
- fork 是 PDP-7 的错误,原始 UNIX 在 PDP-11 上因无 MMU 而无法并发运行。
- 大多数 OS 在用户空间启动进程,但内核也可以加载静态 ELF 二进制。
- exec 的问题在于进程创建(昂贵的内核操作)和加载(便宜的用户空间)分离。
- Windows 不用 fork,但进程创建慢,性能不是最好的论据。
- fork 的问题在于它迫使 OS 设计决策:写时复制和过度提交,这些决策有重大权衡,不应该是强制性的。
- 写时复制可能是个好主意,fork 能受益于它;对 Linux 没问题。
- 过度提交值得质疑,两者应当基于自身优点讨论,不是 fork 的必然后果。
- fork 不强制过度提交,例如 OpenBSD 不允许过度提交。
- 不过度提交时,fork 在内存不足时会失败,限制程序设计。
- fork+exec 中子进程立即 exec,fork 做无用的引用计数增减,合并系统调用可以避免。
8. Nvidia 提出了一款针对 Windows PC 的强大 CPU 系统 (Nvidia is proposing a beast of a CPU system for Windows PCs) #
https://twitter.com/lemire/status/2062880075117113739
Nvidia 提出了一款针对 Windows PC 的强大 CPU 系统,配备 128GB 统一共享内存和最多 6,144 个最新 CUDA 核心。CPU 部分包含 10 个性能核心(基于 Cortex-X925)和 10 个效率核心,支持 6 个 128 位 SVE2 SIMD 执行单元,性能介于 AMD 和苹果芯片之间。最大的变化是统一内存架构,类似苹果多年前的路线,让 CPU 和 GPU 共享同一个内存池。虽然速度不如专用显存,但成本低且带宽足以本地运行 AI 模型。作者认为本地运行 AI 仍是小众应用,但这类机器很适合玩游戏。接下来值得关注 Intel 和 AMD 的应对措施。作者指出,AMD 近期处理器支持的 AVX-512 指令集在数据处理能力和灵活性上远超 Cortex-X925 的 SVE2,但 Intel 在客户系统上一直不太积极推广 AVX-512。
HN 热度 220 points | 评论 410 comments | 作者:tosh | 11 hours ago #
https://news.ycombinator.com/item?id=48424605
- 统一内存池是系统架构的“游戏规则改变者”,能优化利用率、降低内存成本,尤其适合数据中心外的场景。
- 当前游戏和消费级工作负载并未充分利用 GPU 的 PCIe 带宽或 GDDR 带宽,本地 AI 用例对普通消费者也无明显受益。
- 游戏开发者指出统一内存是当前世代主机与 PC 的主要区别之一,但充分利用 GPU 带宽非常困难,需从高层级规划以避免延迟和缓存失效。
- 引擎选择会影响资源分配,但即使使用 Unreal 引擎也难以完美利用带宽,因为数据需先经过 RAM 再到 GPU。
- 统一内存可能导致不可升级,助长厂商像苹果那样对内存升级收取高价;但 LPCAMM2 等可插拔方案(如 Framework 使用)提供了替代选择。
- 用户偏好可插拔内存选项(如 SODIMM 插槽)和可调视频内存,但轻薄趋势可能阻碍此设计。
- 焊接内存并非必然更快:有实例显示同一处理器下焊接 DDR3 比插槽 DDR4 慢;LPCAMM 在高速下可能存在速度惩罚。
- 共享统一内存池会引入侧信道攻击风险,需依赖内存安全设计(如 Rust)来保障安全。
9. 您客厅里的智能电视是 AI 抓取经济中的一个节点 (The Smart TV in Your LivingRoom Is a Node in the AIScraping Economy) #
该网页是一篇研究性博客文章,揭露了 Bright Data 公司如何利用嵌入在智能电视和手机应用中的 SDK,将用户家庭设备变成 AI 模型训练数据的抓取节点。
文章指出,智能电视是理想的“住宅代理”:24 小时开机、带宽无限、用户很少关注。Bright Data 通过一个公共的未认证配置端点,管理合作方应用中的 SDK。合作方包括 PlayWorks、CloudTV、Longvision 等,其中 PlayWorks 号称覆盖 2.5 亿电视家庭。用户通过类似 Petflix 的免费应用同意“偶尔使用您的 IP 地址下载公共网络数据”,但实际默认每月 WiFi 流量预算高达 200GB。
文章强调,AI 公司因数据中心 IP 被封锁,转而依赖住宅代理网络来抓取训练数据。Bright Data 的 SDK 以 iOS 框架形式存在,启动时未经认证即可获取完整配置。作者通过逆向工程和 30 天流量监测,详细说明了 SDK 的工作原理及对用户隐私和带宽的影响。
HN 热度 195 points | 评论 85 comments | 作者:nikcub | 15 hours ago #
https://news.ycombinator.com/item?id=48422993
- 智能电视的 SDK 会通过 WebSocket 连接 AWS Global Accelerator,讽刺的是抓取方和被抓取网站可能都在 AWS 上。
- 有人立刻将相关域名加入 DNS 屏蔽列表。
- 从不把“智能”设备连 WiFi,只当作 HDMI 显示器使用。
- 担心访客会误操作连上网络,导致电视接收大量广告和固件更新。
- 即使只通过 HDMI 输入,电视仍会通过自动内容识别(ACR)收集观看内容数据。
- 在网络层面通过 MAC 地址黑名单或物理拆除 WiFi 模块来阻止连接。
- 担忧断开网络后,电视收集的数据会在本地积满存储空间,最终导致变砖。
- 建议使用 MAC 地址白名单或子网划分来精细控制联网权限。
- 分享个人网络监控工具显示智能电视会频繁连接 Apple 等媒体服务域名。
10. Zeroserve: 一个可以用 eBPF 编写脚本的零配置网络服务器 (Zeroserve: A zero-config web server you can script with eBPF) #
https://su3.io/posts/introducing-zeroserve
zeroserve 是一个零配置的 HTTPS 服务器,整个配置由 eBPF 程序替代。它把网站打包成一个 tar 文件,直接从 tar 中服务静态文件,支持热重载。所有网络和磁盘 I/O 基于 io_uring,单线程事件循环,性能优异:在单核上处理 HTTPS 请求时,多数场景超越 nginx 和 Caddy。
脚本用 eBPF 编写,在用户空间 JIT 编译后运行,可对每个请求进行改写、认证、限流、反向代理等。脚本链式执行,支持预占位模板替换、JSON 解析、OIDC 登录、AWS SigV4 签名等功能。运行时内存限制 256KB,可抢占,不会阻塞其他请求。
安全方面:仅 TLS 1.3、加密客户端问候(ECH)、SNI 证书选择、JA4 指纹。设计核心是把请求路径的全套逻辑放在一个程序中,从上到下可读,摒弃传统声明式配置与脚本分离的架构。
HN 热度 180 points | 评论 45 comments | 作者:losfair | 9 hours ago #
https://news.ycombinator.com/item?id=48425723
- TechEmpower 基准测试似乎已停止更新,新的热门基准测试是 http-arena。
- LLM 界面设计糟糕,技术厉害但交互体验差。
- LLM 使探索新想法变得便宜快速,但 nginx 本身已足够出色;人们更偏好声明式配置而非代码,内置功能已满足多数需求,无需编写 C 代码。
- 配置格式逐渐复杂化(如 YAML、GitLab),最终沦为特设编程语言;不如直接嵌入 Lua/Python 解释器,Windows INI 文件是数据与代码分离的好例子。
- 用 LLM 可快速将 nginx/caddy 配置翻译为 Zeroserve 代码,或从 Kubernetes Ingress 生成配置;配置只是更高层描述的产物。
- 可通过宏来抽象复杂性并实现配置文件的“配置”。
- 随着 AI 发展,“人类语言 → 机器效果”的映射可能更符合 AI 的直观性。
- 不应过于归功 LLM,项目提交历史显示大量编码智能体参与;但 LLM 确实让逆向工程和按兴趣探索变得更快捷,前提是知道自己在做什么。
- 希望支持嵌入 .rs 文件而非 .c 文件;期待内核加速的 web 服务器,建议使用 forking + SO_REUSEPORT,以及 io_uring 和 kTLS。
- 作者已确认会实现 forking + SO_REUSEPORT。
- 希望整合 XDP 等 BPF 程序类型以向下集成 L7 HTTP 功能。
- 静态文件服务并非当前重点,但仍有用户(如 Ghost 转静态)自行搭建;推荐 redbean.dev 和 cosmopolitan 库。
- 建议用 Cloudflare/GitHub 免费托管静态文件,但拒绝将控制权交给这些大公司。
Hacker News 精彩评论及翻译 #
S&P 500 rejects SpaceX, also blocking entry for Op… #
https://news.ycombinator.com/item?id=48422128
Big relief for me. As a passive investor, I want the indices to follow the same passive strategy they always have, and specifically not make exceptions for specific companies like SpaceX wanted.
Plenty of ways to get exposure to that stock without it going into the indices it is not qualified for.
zhivota
对我来说是个巨大的解脱。作为一个被动投资者,我希望指数继续遵循它们一贯的被动策略,尤其不要像SpaceX所希望的那样为特定公司破例。有很多方式可以投资那支股票,而不必让它进入它并不符合条件的指数。
Google to pay SpaceX $920M a month for compute cap… #
https://news.ycombinator.com/item?id=48424983
This is a masterful piece of financial engineering by Google and SpaceX.
Google purchased 10% of SpaceX over a decade ago. After dilution they probably own around 5%.
SpaceX is valued at a whopping 94x revenue. This deal increases SpaceX’s revenue by $11 billion per year. If SpaceX maintains this revenue multiplier, then this single deal boosts SpaceX’s valuation by 94 x 11 billion = $1 trillion dollars. Google owns 5% of SpaceX, so they make 50 billion dollars. Google spends 10 billion and makes 50 billion, $40 billion profit.
The even better part is that because of this deal, SpaceX is now profitable. The S&P requires companies to demonstrate 12 months of profits before they can enter the S&P 500 index. SpaceX lobbied to have this profitability requirement removed, but S&P said no and refused to rewrite the rules.
Now with this incredible deal, SpaceX is now GAAP profitable under the existing rules, and they get to join the index next year without a rule change.
Truly a brilliant deal for everyone involved.
tristanj
这是谷歌和SpaceX联手完成的一项精妙的金融工程。
十多年前,谷歌收购了SpaceX 10%的股份。经过稀释后,他们可能还持有约5%的股份。
SpaceX的估值高达收入的94倍。这笔交易每年将为SpaceX增加110亿美元的收入。如果SpaceX保持这一收入乘数,那么仅这一笔交易就能将SpaceX的估值提升94×110亿=1万亿美元。谷歌持有SpaceX 5%的股份,因此他们将获利500亿美元。谷歌投入100亿美元,赚取500亿美元,净赚400亿美元。
更妙的是,由于这笔交易,SpaceX现在实现了盈利。标普要求公司必须连续12个月盈利才能纳入标普500指数。SpaceX曾游说取消这一盈利要求,但标普拒绝了,并拒绝修改规则。
如今,凭借这笔不可思议的交易,SpaceX在现有规则下已实现公认会计准则盈利,明年无需修改规则即可加入标普500指数。
对参与各方来说,这确实是一笔绝妙的交易。
Did Claude increase bugs in rsync? #
https://news.ycombinator.com/item?id=48419628
For those commenting, I suggest you read the post linked by the rsync author:
https://medium.com/@tridge60/rsync-and-outrage-d9849599e5a0
(Disclosure: while I haven’t talked with him in years, Tridge was my colleague and mentor for many years. I feel it is worth considering his view before joining a crusade)
RustyRussell
给评论者们的建议:请先读一下rsync作者分享的这篇文章:
https://medium.com/@tridge60/rsync-and-outrage-d9849599e5a0
(声明:虽然多年未与他交谈,但Tridge曾是我的同事和导师多年。我认为在加入声讨之前,值得先考虑他的观点)
Ask HN: What was your “oh shit” moment with GenAI? #
https://news.ycombinator.com/item?id=48418158
I bought an Alesis QS8.1 super cheap in perfect condition (was a top grade digital piano/synth in the 90s).
and then i realized that ALL of the software (which i collected from defunct websites and archived on github) related to it was ancient and after a while of getting tired of using WINE every single time i decided i wanted a cross platform modern equivalent that did everything that several of these different programs did (plus break out some stuff that was now potentially possible with modern computer)
i thought it would be extremely hard because the computer to synth communication is pretty much only via sysex commands (of which the actual wave file encoding protocol was undocumented)
Claude walked me through examining the some of the original software in GHIDRA, and I had a working demo that night…..now im just playing with adding new features to it.
jzemeocala
我以超低价购入了一台成色完美的Alesis QS8.1(90年代顶级数字电钢琴/合成器)。
后来发现所有与之相关的软件(我从已关闭的网站上收集并归档到GitHub)都极其古老,在厌倦了每次都要用WINE运行后,我决定开发一个跨平台的现代替代方案,能整合那些旧程序的所有功能(还能实现一些现代计算机可能实现的新功能)。
起初以为这极其困难,因为电脑与合成器的通信几乎完全依赖SysEx指令(其中波形文件编码协议从未公开)。
Claude引导我用GHIDRA分析了一些原始软件,当晚就做出了可用的演示……现在我在给它添加新功能。
How LLMs work #
https://news.ycombinator.com/item?id=48421462
Back when ChatGPT came out, I was so shocked by how good it was for an “AI” product that I simply had to know how it worked. Over the next month I ended up drawing out a block diagram on a whiteboard I have in my office, with the math involved next to each step in the blackboard. I’d puzzle about each step along the way, and the triumph of completing the drawing was also that of this sense of deep understanding. I kept that drawing up for many months after, and would gaze at it often during meetings and idle moments in wonder.
This is to say: the autoregressive decoder-only transformer llm architecture as pioneered by openai is wildly simple for how revolutionary its results are. I was reading about non-learned classical SLAM systems (uses video + handcrafted math to produce 3d mappings of physical spaces while also locating the camera in those spaces) at the time, and comparatively speaking I’d say the math is about as complicated as ONE of the components in those complex formulations. The only reason frontier LLMs need 6-figure computers to run is because the model designers made the middle bit in those models REALLY BIG, dimensionally speaking. They just took the steam engine, made a few gargantuan versions of it, and are selling them as the ultimate source of power.
This was openai’s entire breakthrough. Making this particular model architecture larger leads to emergent capabilities like being able to pick the best ending to a story/set of instructions or answer questions about broad factual knowledge. I’ve been meanwhile watching these AI companies attempt, successfully, to sell this capability as some sort of robot consciousness hand-crafted by supergeniuses. The fact that they are getting away with it is almost as shocking to me as the discovery itself.
malwrar
当ChatGPT刚出来时,我被它作为一款"AI"产品表现出的优异程度深深震撼,以至于必须搞清楚它的工作原理。接下来的一个月里,我在办公室的白板上画出了框图,黑板上的每一步旁边都标注了相关数学公式。我反复琢磨每一步的推导,完成这幅图时的成就感也源于那种深刻理解的愉悦。之后好几个月我都保留着那幅图,开会或闲暇时常常凝视它,充满惊叹。
也就是说:OpenAI开创的这种仅解码器自回归Transformer大语言模型架构,与其革命性的成果相比,简直简单得令人发指。当时我正在研究非学习的经典SLAM系统(利用视频和手工设计的数学公式生成物理空间的三维地图,同时定位相机在空间中的位置),相比之下,我会说其数学复杂度只相当于那些复杂公式中某个组件的水平。前沿LLM之所以需要六位数的计算机来运行,仅仅是因为模型设计者把模型中间的部分在维度上做得极其庞大。他们只是造出了蒸汽机,然后做了几个巨型版本,就将其作为终极动力源来出售。
这就是OpenAI的全部突破。让这种特定模型架构变得更大,就会涌现出新的能力,比如为故事或指令集选择最佳结局,或者回答广泛的事实性问题。而我一直看着这些AI公司试图并成功地将这种能力包装成某种由超级天才手工打造的机器人意识。他们居然能蒙混过关,这几乎和这个发现本身一样让我震惊。
Meta confirms 1000s of Instagram accounts were hac… #
https://news.ycombinator.com/item?id=48427959
“The tool itself worked properly and functioned as intended; however due to a bug in a separate code path, the system did not properly verify that the email address provided by the individual requesting a password reset matched the email address associated with that user’s Instagram account,” said Meta in its breach notice.
I’m not sure “worked properly” and “as intended” accurately describe this situation.
Cyan488
该工具本身运行正常且达到了预期效果;但由于另一段代码路径存在漏洞,系统未能正确验证申请密码重置的用户所提供的邮箱地址是否与该用户的Instagram账户关联的邮箱一致,“Meta在泄露通知中表示。
我不确定"运行正常"和"达到预期效果"能否准确描述这种情况。
Ask HN: What was your “oh shit” moment with GenAI? #
https://news.ycombinator.com/item?id=48417379
I could go on and on, but Claude recently decompiled the firmware of my camper van, documented all the CAN interfaces, then programmed an ESP32 module to talk to the van’s integrated systems (power, HVAC, lighting, tanks). That sort of embedded systems integration is completely out of my wheelhouse.
I honestly don’t understand AI naysayers. I use Claude every day both professionally as a Solution Architect and personally in a variety of projects I simply could not have ever approached alone.
shreddude
我可以说个不停,但克劳德最近把我房车的固件反编译了,记录了所有CAN接口,然后给一个ESP32模块编程,让它与房车的集成系统(电源、暖通空调、照明、水箱)通信。这种嵌入式系统集成完全超出了我的能力范围。
说实话,我不理解那些反对AI的人。我每天都在用克劳德,工作上作为解决方案架构师,个人生活中也在各种项目里用它——那些项目如果只靠我自己根本不可能接触。
Ask HN: What was your “oh shit” moment with GenAI? #
https://news.ycombinator.com/item?id=48417845
My furnace went out during the 2025 holiday and I couldn’t get an appointment with a repair person for 2 days. It was getting very cold in my house so I went into my attic and made several videos of the furnace attempting to start and gave it to gemini. It diagnosed the issue immediately and had me spin one of the components (a small exhaust fan) while the furnace tried to fire. It came on immediately. I had to do that several times, but it worked until the HVAC service showed up.
andrewthornton
我家的暖气在2025年假期期间坏了,而且两天内都约不到维修师傅。屋里越来越冷,于是我爬进阁楼,拍了几段暖气尝试启动的视频,拿给Gemini看。它立刻诊断出问题,并让我在暖气点火时手动转动其中一个部件(一个小排风扇)。暖气马上就启动了。我重复了几次这个操作,但一直撑到了暖通维修人员上门。
S&P 500 rejects SpaceX, also blocking entry for Op… #
https://news.ycombinator.com/item?id=48425180
Letting new stocks marinate in the market and get 4 quarters of SEC filings along with following all the GAAP accounting practices will definitely help evaluate them before inclusion. The last large boom/bust cycle had a couple of companies, at least, that were doing illegal things. I’m not stating that these three are, just that nobody knows and the process should play out.
I do wonder if any of these three companies are using AI to do their accounting and bookkeeping. What happens when there are AI hallucinations affecting those outcomes?
mattbrewsbytes
让新发行股票在市场上充分沉淀,经历四个季度的SEC申报并遵循所有GAAP会计准则,这无疑有助于在纳入指数前对其进行充分评估。上一次大起大落的周期中,至少有几家公司涉及非法行为。我并非指这三家也是如此,只是没人能确定,而流程理应逐步推进。
我确实好奇,这三家公司中是否有任何一家使用AI进行会计和记账。当AI出现幻觉影响这些结果时,会发生什么?
Google to pay SpaceX $920M a month for compute cap… #
https://news.ycombinator.com/item?id=48426218
I sincerely hope the market is not willing to value this sort of deal at a P/E ratio anywhere near 94.
Off the top of my head, there is a very well established business involving buying expensive things and leasing them to the companies that intend to operate them so they can sell services: aircraft leasing.
AER is the biggest player and they have a P/E ratio of, drumroll please, 6. And I expect that GPUs, despite currently looking like an appreciating asset, will actually depreciate faster than aircraft in the long run.
amluto
我衷心希望市场不会愿意给这类交易接近94倍的市盈率估值。
我随口就能想到一个非常成熟的商业模式:购买昂贵资产,然后租赁给有意运营这些资产以提供服务的公司——飞机租赁就是典型例子。
AER是该领域最大的参与者,它们的市盈率是——请敲鼓吧——6倍。而且我预计,尽管GPU目前看起来像是增值资产,但从长远来看,它们的折旧速度实际上会比飞机更快。
Ask HN: What was your “oh shit” moment with GenAI? #
https://news.ycombinator.com/item?id=48419997
Related story, while applying a firmware update to my Kawai CA49 piano, I bricked it due to flashing the wrong file (The process was broken, and I got desperate and tried something stupid, which bricked the piano). Claude walked me through looking for signs of life, and since OTA from the phone app wasn’t working for me, it downloaded the Kawai Android APK, decompiled the Java, figured out the hardcoded key used for encrypting the firmware update. Extracted the piano firmware update, decrypted it, and then wrote a flashing script to program the piano from my laptop via bluetooth. My piano was back to working within an hour.
jsharf
相关经历:我在给卡瓦依CA49钢琴更新固件时,因为刷错了文件导致变砖(更新过程出了问题,我一时慌乱做了蠢事,结果钢琴彻底无法启动)。Claude一步步帮我检查是否有任何反应迹象,由于手机应用的OTA更新对我无效,它下载了卡瓦依Android APK,反编译了Java代码,找到了用于加密固件更新的硬编码密钥。接着提取钢琴固件更新文件,解密后编写了一个刷机脚本,通过蓝牙从笔记本电脑对钢琴进行编程。不到一小时,我的钢琴就恢复正常了。
Three of our worst VC stories #
https://news.ycombinator.com/item?id=48417744
As with all things, the horror stories just get the most attention. People love to rage. There are plenty of boring (good, even!) VCs out there. They just work more quietly, professionally.
I’ll share a story, but its about a close friend and not me so I won’t name any explicit actors and I’m going to round out the numbers. You either trust me or you don’t, but this is a very direct relationship I have to both the founder and VC.
The story is this: the founder started their company outside of SV, so the lawyers weren’t super familiar with startups and messed up the initial incorporation and stock plan stuff (actually super common: use Stripe Atlas or pay a startup-aware lawyer!). Went under the radar through years. This company ended up being bought for nearly $1B (with a B) after many rounds and a large board.
During the legal work to close the acquisition, they found out this messed up stock plan. Without going into the details, the effect was that instead of taking home $200M, the founder would take home ~$75M. The mistake the lawyer made almost a decade earlier was about to cost him $125M.
Most of the board basically said “too bad so sad, law is law.” But one VC (the one I know, the one I’m talking about) basically strong armed and politicked the whole thing and eventually convinced everyone around the table to give up an equal share of their own holdings to make the founder whole.
Letter to the law: they didn’t have to.
Spirit of being founder friendly: this VC went to bat hard and got everyone to yield to make things “right.”
Also, look, you might argue $75M vs $200M is just “rich vs rich.” Who cares? Sure. That’s not the point.
You don’t hear about stuff like this because honestly its not a big enough deal and feel good stories get way less clicks than pitchfork stories.
mitchellh
和所有事物一样,恐怖故事总是最吸引眼球。人们热衷于愤怒。其实有很多(甚至可以说相当不错的)低调风投,他们只是更安静、更专业地工作。
我分享一个故事,但涉及一位密友而非我本人,所以不会点名具体人物,数字也会模糊处理。信不信由你,但我和这位创始人、这位风投之间都有着非常直接的关系。
故事是这样的:创始人起初在硅谷之外创立公司,律师不熟悉初创企业,搞砸了最初的公司注册和股权计划(其实很常见:要么用Stripe Atlas,要么请懂初创的律师!)。这事多年来都没人发现。这家公司经过多轮融资和庞大董事会后,最终以近10亿美元被收购。
在完成收购的法律程序中,他们发现了这个混乱的股权计划。细节不赘述,结果就是创始人原本能拿到2亿美元,最后只能拿到约7500万美元。律师将近十年前犯的错,让他损失了1.25亿美元。
大部分董事基本都说"太糟了,但法律就是法律”。但有一位风投(就是我认识的那位,我提到的这位)几乎靠强硬手段和政治游说,最终说服了在座所有人各自放弃同等份额的持股,来让创始人获得应得的全部款项。
法律上,他们本不必这么做。
但秉持"对创始人友好"的精神,这位风投奋力周旋,让所有人让步以"拨乱反正"。
另外,你可能会说7500万和2亿不过是"富人与更富"的区别。谁在乎呢?确实。但这不是重点。
你不会听到这类故事,因为说实话它不够轰动,而且正能量故事远不如讨伐檄文点击量高。
Did Claude increase bugs in rsync? #
https://news.ycombinator.com/item?id=48419197
Was just looking at commits and came across a commit and its revert
original commit: https://github.com/RsyncProject/rsync/commit/d046525de39315d625ffaef4fdd6e7cf12148016
- if (!ptr)
- ptr = malloc(num * size);
- else if (ptr == do_calloc)
+ if (!ptr || ptr == do_calloc)
ptr = calloc(num, size);
Written with claude. This is a good example of what slips through LLM attention. It forces all allocations to be calloc as if it is a strict upgrade. For large and recursive allocations, this becomes a significant cost.
reverted in https://github.com/RsyncProject/rsync/commit/7db73ad9a1b8721f14a43219d73127b23b86fe00
if you read the description of revert half carefully, it’s easy to tell that even that was written by an LLM .
I can understand the sentiment of whoever posted the original thread.
GodelNumbering
刚刚看提交记录时,发现了一个提交及其还原操作
原始提交:https://github.com/RsyncProject/rsync/commit/d046525de39315d625ffaef4fdd6e7cf12148016
- if (!ptr)
- ptr = malloc(num * size);
- else if (ptr == do_calloc)
+ if (!ptr || ptr == do_calloc)
ptr = calloc(num, size);
用 Claude 写的。这是 LLM 注意力漏洞的一个典型例子。它强制将所有分配改为 calloc,好像这是一种严格的升级。对于大型和递归分配来说,这会带来显著的性能成本。
已在 https://github.com/RsyncProject/rsync/commit/7db73ad9a1b8721f14a43219d73127b23b86fe00 中还原
如果你稍微仔细阅读还原操作的描述,很容易看出那甚至也是 LLM 写的。
我能理解最初发帖人的感受。
India’s surprise baby bust #
https://news.ycombinator.com/item?id=48413895
This has happened to every single society [*] as it industrializes [0], and offering extensive support and incentives to parents (e.g. as has been tried in Scandinavian countries) does very little to reverse this trend [1, 2].
My hypothesis is that as societies industrialize, they afford their population more and more activities that are simply more fun and rewarding than having children. So many people I know put off having children (or curtailed the number they had) because they were reluctant to give up the activities only available in a childfree/one-and-done life. Ultimately, we are hedonistic creatures, and having kids is antithetical to the myriad hedonic pursuits available in wealthy, industrialized societies.
[*] Israel is the lone exception, due to its Orthodox Jewish population.
[0] https://ourworldindata.org/global-decline-fertility-rate
[2] https://pmc.ncbi.nlm.nih.gov/articles/PMC10049131/
MontyCarloHall
每个社会在工业化过程中都经历过这种情况[*],而为父母提供大量支持与激励措施(例如斯堪的纳维亚国家尝试过的做法)对扭转这一趋势收效甚微[1, 2]。
我的假设是:随着社会工业化,人们可从事的活动越来越多,这些活动比生儿育女更有趣、更有回报。我认识的许多人推迟生育(或减少生育数量),是因为他们不愿放弃只有无子女或只生一个孩子的生活才能拥有的活动。归根结底,我们是享乐动物,而在富裕的工业化社会中,拥有孩子与无数可获得的享乐追求是背道而驰的。
[*] 以色列是唯一的例外,因其正统犹太教人口的存在。
[0] https://ourworldindata.org/global-decline-fertility-rate
[2] https://pmc.ncbi.nlm.nih.gov/articles/PMC10049131/
Google to pay SpaceX $920M a month for compute cap… #
https://news.ycombinator.com/item?id=48426082
Since the S-1 filing, xAI has taken over and is likely the largest share of revenue. I would estimate that ~95%+ of xAI revenue, and 100% of its profit, is from renting their datacenters.
This is a datacenter REIT bolted onto a social media company bolted onto launch business bolted onto a niche ISP. The expected price to sales is ~100x. The best datacenter REITs trade at ~10x and pay a dividend, which SpaceX does not. Meta trades at ~7x sales. Comcast is one of the best-run ISPs, and it pays a 5.5% dividend on a stock trading at < 1x sales.
To say SpaceX is overvalued is to even beginning to convey the magnitude of the situation. It’s going to be very painful when the valuation normalizes.
runako
自S-1文件提交以来,xAI已接管业务,并很可能贡献了最大份额的收入。我估计xAI约95%以上的收入以及100%的利润,都来自其数据中心的租赁。
这本质上是一个数据中心REIT(房地产投资信托基金),上面嫁接了一家社交媒体公司,再嫁接了一家发射业务公司,再嫁接了一家小众互联网服务提供商。其预期市销率约为100倍。而最好的数据中心REITs市销率仅约10倍,并且会派发股息——SpaceX并不派息。Meta的市销率约为7倍。Comcast是运营最出色的互联网服务提供商之一,其股票市销率低于1倍,且派发5.5%的股息。
说SpaceX估值过高,甚至才刚刚开始描述问题的严重性。当估值回归正常时,局面将会非常痛苦。
Meta confirms 1000s of Instagram accounts were hac… #
https://news.ycombinator.com/item?id=48428631
In italian we say “l’operazione è riuscita perfettamente, ma il paziente è morto” -> “the surgery was a complete success, but the patient died”
vb-8448
在意大利语中我们说“l’operazione è riuscita perfettamente, ma il paziente è morto” -> “手术非常成功,但病人死了”
Ask HN: Why is the HN crowd so anti-AI? #
https://news.ycombinator.com/item?id=48421820
I call these AI tools “proprietary non-determenistic database of the free internet”. They belong to american companies which can cut off your access if american government doesn’t like your country’s government. They fed from the free internet that many of us grew up in, store it in humans unreadable form and sell you access to it. If some day claude starts to spit out compiled binaries instead of code nobody will notice, and we’ll essentially get proprietary cloud-hosted compiler that most in the world depends on to build software. With built-in telemetry and backdoors and clause in license that allow full overtake of your business if provider wants it ofc. It’s a great shift from the internet we all know and love towards the new subscription-based access to world’s propriatary knowledge base. It’s a perfect “mind control” tool as well - you don’t need USAID, “free media” and stuff like that in other countries when all people there including politicians ask chatgpt everything from meaning of life to recipies of pancakes. Once you see these political and philosophical dimensions it’s hard to unsee how claudecode running on my PC won’t turn into a weapon some day. But in blissful ignorance it’s fun to use, and companies love it for the promise of replacing people. Amen
oleg_antonyan
我将这些AI工具称为“自由互联网的专有非确定性数据库”。它们属于美国公司,如果美国政府不喜欢你的国家政府,他们可以切断你的访问权限。它们吞噬了我们许多人成长过程中所熟悉的自由互联网,将其以人类不可读的形式存储,然后向你出售访问权限。如果某一天Claude开始输出编译好的二进制文件而不是代码,没人会注意到,而我们将实际上得到一个专有的、云端托管的编译器,世界上大多数人将依赖它来构建软件。当然,它还内置了遥测和后门,以及许可证中的条款,允许提供商在想要的时候完全接管你的业务。这是从我们所熟知和热爱的互联网向基于订阅的全球专有知识库访问的重大转变。它也是一个完美的“思想控制”工具——当其他国家的人们,包括政客,都向ChatGPT询问从生命意义到煎饼食谱的一切问题时,你就不再需要美国国际开发署、“自由媒体”之类的东西了。一旦你看到了这些政治和哲学层面的问题,就很难不去想象某一天运行在我电脑上的Claude代码会不会变成武器。但在无知的幸福中,它用起来很有趣,而公司也喜欢它取代人类的承诺。阿门。
New method turns ocean water into drinking water, … #
https://news.ycombinator.com/item?id=48415560
There is a fundamental minimum amount of energy needed to desalinate: you can’t take less energy to do it,than you could gain back (from osmotic pressure) if you allowed the desalinated water to expand a cylinder containing the residual brine. This is large. This paper is a thermal method, so it doesn’t have an electricity input, but to justify their efficiency claim, they should really compare against what you could do by using the same surface area for solar panels, driving a conventional setup. My (limited) understanding is that conventional reverse osmosis is not far from the theoretical optimum, energy-wise, the main difficulties being operational (the membranes need declogging). And of course RO is more expensive than rain.
This paper is interesting, however, in directly producing crystalline salt, which is lower volume than brine and easier to dispose of, maybe even valuable.
ajb
海水淡化所需能量存在一个基本下限:如果让淡化后的水推动一个装有残余盐水的圆柱体膨胀,你无法用比从渗透压中回收的能量更少的能量来完成这一过程。这个下限值很大。这篇论文采用的是热方法,因此不需要电力输入,但为了证明其效率主张,他们应该与使用相同面积的太阳能板驱动传统装置的方案进行比较。据我(有限的)理解,传统的反渗透法在能量效率上已接近理论最优值,主要困难在于操作层面(膜需要除垢)。当然,反渗透法比雨水更昂贵。
不过,这篇论文的有趣之处在于它直接生产结晶盐,其体积比盐水小,更易处理,甚至可能具有价值。
S&P 500 rejects SpaceX, also blocking entry for Op… #
https://news.ycombinator.com/item?id=48421668
Yep!! Respect to them. I was planning to move to an equal weight index but this gives me a little more time to evaluate options.
zippyman55
是的!!尊重他们。我原本打算转向等权重指数,但这给了我多一点时间来评估选择。